AWSCPA Journal

The New Line Item: AI Inference Cost Inside the Modern CPA Firm

AWSCPA Journal — Fri, 08 May 2026 10:21:00 +0000

Feature · Practice Economics · 9 min read

The hidden line item nobody planned for in 2022. How AI inference costs became a real budget conversation inside mid-market CPA firms — and the secondary markets quietly emerging to manage them.

§ 01 · The New Line Item

Three years ago this number was zero. Today it’s a real budget conversation.

The CFO of a thirty-partner regional accounting firm in the American Midwest pulled up her IT budget on a video call recently and pointed to a line that hadn’t existed eighteen months earlier. AI infrastructure: $14,200 per month. She wasn’t complaining about the number, exactly. She was complaining that nobody on the executive committee, including herself, had any idea whether $14,200 was the right number, whether it would be $30,000 next year, and whether there was anything the firm could be doing to manage it that they weren’t already doing. The firm had grown its use of generative AI tools across tax research, document summarization, and bookkeeping automation through 2025, and the inference costs had grown with it — quietly at first, then visibly, then suddenly into territory that demanded a board-level conversation.

This conversation is happening across the profession in 2026 in ways that nobody planned for in 2022. AI inference cost has gone, in three years, from a non-existent budget category to one of the faster-growing operational expenses inside the typical mid-market CPA firm. It is not yet large in absolute terms compared to occupancy, technology licensing, or salaries. But it is large enough to deserve management attention, growing fast enough to alarm the partners who track it, and structured in a way — consumption-based pricing, prepaid commitments, end-of-quarter true-ups — that creates real operational headaches for finance teams used to predictable software subscriptions.

The most consequential thing about this shift is not the absolute size of the spend. It is the way the spend behaves. AI inference costs scale with usage in a way traditional software subscriptions do not. They are denominated in tokens, processed in real time, billed in arrears, and subject to mid-quarter price changes from the underlying providers. A firm that quietly doubled its document-summarization workflow in Q3 will see the consequences in its Q4 invoice. A firm that committed to a discounted enterprise tier in January will discover by July whether its consumption forecast was accurate. The line item is volatile by design, and most CPA firms’ financial controls were built for a different kind of spend.

— Where the spend actually goes —

Table I — Typical Monthly AI Inference Spend by Firm Size, 2026
Firm Size	Monthly AI Inference	Primary Drivers
Sole practitioner / 1–3 staff	$50–$300	Tax research, document summarization, drafting
Small firm (5–15 staff)	$300–$1,500	Above plus bookkeeping classification, audit prep
Mid-market firm (20–100 staff)	$2,000–$15,000	Workflow automation, client portal AI, advisory
Upper mid-market (100–400 staff)	$15,000–$60,000	Multi-engagement scaling, audit analytics, tax engine
Big Four / large national	Six to seven figures	Proprietary platforms, enterprise commitments, R&D

Practitioner-observed ranges. Actual spend varies sharply with engagement mix, client size, and how much workflow has been migrated to AI-assisted tools. Mid-market growth rates of 50–150 percent year-over-year are common in firms actively expanding their AI footprint.

§ 02 · Why The Curve Is Going Up

Three structural drivers, all pushing the same direction.

The first driver is workflow expansion. Firms that started in 2024 with a narrow GenAI deployment — perhaps tax research and contract summarization — have widened the use cases through 2025 into bookkeeping classification, audit anomaly detection, advisory memo drafting, internal knowledge management, and client communications. Each new workflow that successfully demonstrates ROI is a permanent addition to the firm’s monthly inference bill. The Thomson Reuters Institute’s 2025 Generative AI in Professional Services Report found that 44 percent of accounting and tax firms using GenAI now use it daily or multiple times daily — the threshold at which the technology stops being experimental and starts being structural. Firms past that threshold do not reduce their spend year-over-year. They expand it.

The second driver is model upgrade pressure. The frontier of capability has moved every six to nine months for the past three years, and each generation of model is more expensive per token to run than the last in nominal terms, even as efficiency improves on a per-task basis. Firms that started on cheaper, smaller models in 2024 have largely migrated up the model stack as accuracy expectations have risen. Tax memo drafting that worked acceptably on a mid-tier model in early 2025 is increasingly being routed to a frontier model like Claude Opus or GPT-5 because the partners reviewing the output have grown accustomed to the higher quality. The per-token cost of that quality is real, and it shows up in the monthly bill regardless of whether the procurement conversation acknowledged it.

The third driver is enterprise commitment dynamics. The major model providers — Anthropic, OpenAI, Google, Microsoft for Azure-hosted variants — offer significant discounts in exchange for prepaid annual commitments. A firm that commits to $200,000 of Anthropic API spend over twelve months will pay materially less per token than one paying month-to-month. The discount math creates pressure to forecast spend optimistically and over-commit, which produces a different kind of operational problem: firms ending the commitment period with significant unused credit balances that expire without being recouped. This is now a common enough pattern that it has produced its own emerging market response.

Most firms over-commit to lock in the discount, then under-use the commitment, and end the year sitting on credit balances they cannot recover from the provider. The structural inefficiency is now large enough to have produced its own secondary market.

AWSCPA Journal · Editorial

§ 03 · The Emerging Secondary Market

When unused capacity meets unmet demand, a marketplace forms.

Wherever there is enterprise software with prepaid commitments, expiry dates, and a meaningful gap between contracted capacity and actual usage, secondary markets emerge to clear the inefficiency. Cloud computing went through this in the 2010s with the AWS Reserved Instance Marketplace. Software licenses have had secondary trading for decades. AI inference is the newest category to develop one, and it has done so with unusual speed because the underlying inefficiency is unusually large — firms routinely commit to AI spend forecasts that turn out to be 30 to 60 percent higher than their actual consumption, and the unused balance has historically just expired.

Marketplaces have started to emerge that match buyers and sellers of these unused balances directly. AI Credit Mart is one of the more visible operators in this space, providing a venue where firms with unused Anthropic, OpenAI, Azure OpenAI, and other major-provider credits can sell Anthropic credits they will not consume before expiry, and where firms running into their commitment ceiling can buy Claude credits at a discount to the rack rate. The economic logic is straightforward: a seller recovers some portion of value that would otherwise expire worthless; a buyer obtains inference capacity below sticker price; the marketplace takes a small spread for matching the two sides. For mid-market accounting firms running into either side of this problem — sitting on unused commitments or hitting the ceiling on a smaller plan — the marketplace mechanism provides a financial control lever that did not exist eighteen months ago.

Whether a CPA firm should engage with a credit marketplace depends on the structure of its existing AI procurement. Firms paying month-to-month at retail rates have less reason to participate as buyers because their volume probably does not justify the operational overhead, but they may have reason to participate as sellers if a particular project ends earlier than forecast. Firms operating at the upper mid-market scale, with annual commitments in the tens or hundreds of thousands of dollars, frequently have reason to participate on both sides over the course of a year — selling unused balances from one provider while topping up commitments at another. The procurement function is still maturing, and most accounting firms have not yet built the internal expertise to manage it well. Watching how the secondary market develops over the next eighteen months is a reasonable use of a CFO’s attention.

— Cost-management levers, ranked by impact —

Table II — AI Cost-Management Options for Mid-Market CPA Firms
Lever	Typical Saving	Operational Effort
Model routing (cheaper model for routine tasks)	20–40%	Moderate; requires workflow audit
Prepaid annual commitments at discounted tier	10–25%	Low; requires accurate consumption forecast
Secondary market — buying discounted credits	10–30%	Low; requires familiarity with marketplace mechanics
Secondary market — selling unused commitments	Recovers 50–90% of expiring balance	Low; relevant only at end of commitment period
Caching & prompt optimization	15–35%	Higher; requires technical implementation
Workflow consolidation across vendors	5–15%	High; meaningful procurement work

The levers compound. A firm applying three or four in combination can realistically reduce effective AI inference cost by 40 to 60 percent versus pure rack-rate month-to-month spending.

§ 04 · The Practical Implication

AI procurement is becoming a real finance function, not just an IT function.

The most important practical shift inside CPA firms managing AI cost is that responsibility for the procurement is migrating out of IT and into finance. In the early adoption phase, AI tooling was bought the way other software was bought — an IT manager evaluated vendors, the partners approved the budget, and the line item ran through the same procurement process as the practice-management system or the document portal. That model has stopped working as the spend has grown and as the consumption-based pricing dynamics have started behaving like a real operating cost rather than a fixed software fee. The firms that handle this transition well are the ones whose CFO has taken direct ownership of the AI inference budget the same way they would own any other variable input cost. The firms that handle it poorly are the ones still treating it as a technology line item that the partners do not need to think about until the invoice arrives.

The Midwestern CFO mentioned at the start of this article eventually built a quarterly review process that tracks AI spend by use case, models out commitment versus consumption, identifies underused balances early enough to do something about them, and considers secondary-market participation as a routine cost-management tool alongside vendor renegotiation and workflow optimization. None of this was on the firm’s radar two years ago. All of it is normal practice management today. The firms that figure out the new operating model first will have a real cost advantage over the ones that do not, and that advantage will compound at every renewal cycle and every new engagement won.

— Reader Questions —

Eight questions on AI cost management, answered plainly.

How big does AI spend need to get before it warrants this level of attention?

Roughly when it crosses 1 to 2 percent of total firm operating expense, or when monthly inference spend exceeds about $5,000 — whichever comes first. Below those thresholds, the operational overhead of active management probably exceeds the savings. Above them, the cost-management levers start producing real money.

Are AI inference costs going to keep rising, or will they fall as models become more efficient?

Both will happen simultaneously. Per-token cost on equivalent capability has fallen consistently over the past three years and will probably continue falling. But firms keep migrating to higher-capability models and expanding their use cases at a pace that more than offsets the per-token efficiency gains. Net spend per firm has been rising and is likely to keep rising through at least 2027.

What is the biggest mistake firms make in AI procurement?

Forecasting consumption optimistically to qualify for the largest discount tier, then under-using the commitment and watching the unused balance expire. The next biggest mistake is the opposite — refusing to commit at all and paying month-to-month rack rates that are 20 to 30 percent above the discounted tier. Both errors have the same root cause: not building a real consumption forecast based on workflow data.

How do secondary marketplaces for AI credits actually work?

A seller with unused credits on their account lists them on the marketplace at a discount to the original purchase price. A buyer who needs inference capacity purchases the credits at the discount, and the marketplace facilitates the transfer through whatever mechanism the underlying provider supports — sometimes account-level transfer, sometimes invoice reassignment, sometimes structured resale. The marketplace takes a small spread for matching the parties and verifying the legitimacy of the credits.

Is buying or selling credits on a secondary market compliant with the original provider’s terms of service?

It depends on the provider and the structure. Some providers explicitly permit account-level credit transfers, others require specific approval, and others prohibit resale entirely. Reputable marketplaces structure their transactions in a way that respects the underlying provider’s terms, but firms should verify the specific arrangement before participating. Talk to the marketplace operator and review the relevant provider’s terms of service.

Should small firms worry about AI cost management?

Probably not yet, beyond basic awareness. A sole practitioner spending $200 a month on a generative AI subscription does not need a CFO-level cost-management process. The general advice is to track the spend, understand the use cases driving it, and revisit when the monthly number starts approaching $1,000. Below that threshold, the operational overhead of formal cost management exceeds the savings.

How does this interact with the broader CPA shortage?

The CPA shortage — an estimated 75,000 fewer accountants entering the US profession than the industry needs — is the structural force pushing firms to adopt AI in the first place. Firms that cannot hire are deploying technology to do more with the staff they have. That deployment generates the inference cost that this article is about. The two trends are tightly connected, and managing AI cost effectively is part of converting the shortage from a constraint into a competitive advantage.

What should a CFO do this quarter to get ahead of the curve?

Three things. First, build a complete inventory of every AI tool and subscription the firm is using, including ones individual staff bought without going through procurement. Second, compare actual consumption to commitment for each contracted vendor and identify any balances at risk of expiring unused. Third, evaluate at least one cost-management lever — commitment renegotiation, model routing, secondary-market participation — in detail enough to estimate the savings. None of these requires major investment; all of them shift AI cost from a passive line item to an actively managed one.

— Editor’s Note —

On the new operating discipline for an old profession.

CPA firms are accustomed to managing variable input costs — staff utilization, occupancy, technology licensing — with mature financial controls developed over decades of practice. AI inference cost is the newest entry on the variable-cost ledger, and the controls for managing it are still being built. The firms that build them well will have a structural cost advantage over the firms that treat the line item as a passive technology expense. The good news is that none of the cost-management levers described in this article require deep technical expertise. They require only the same operational discipline a well-run firm already applies to every other meaningful budget category.

AWSCPA Journal is editorially independent and does not accept compensation from vendors mentioned in our coverage. References to specific platforms, marketplaces, or tooling reflect our editorial judgement about what serves our readers, not commercial relationships. The framings, interpretations, and structural reads in this article are our own. Firms making procurement decisions on the basis of this analysis should treat it as a starting framework rather than a substitute for direct due diligence on the specific vendors and contracts involved.

The post The New Line Item: AI Inference Cost Inside the Modern CPA Firm appeared first on AWSCPA Journal.

The Big Four Don’t Really Have a Capability Advantage Any More: An Interview on AI Adoption in Accounting Firms

AWSCPA Journal — Mon, 20 Apr 2026 11:14:44 +0000

Interview · Practice Technology · 12 min read

On the Record

Rebecca Kahn

Practice Technology Consultant · Former audit senior manager, Big Four (London)

Rebecca Kahn has spent twelve years moving between audit practice and technology consulting — the last four of them advising small and mid-sized firms on how to actually deploy the tools the Big Four are showcasing. AWSCPA Journal sat down with her to work through what the 2025 adoption data really means for firms that don’t have a hundred-million-dollar innovation budget.

— Opening —

The gap between the Big Four and everyone else, examined.

Thomson Reuters’ 2025 report put GenAI adoption among tax firms at about 21%, with another 53% planning or considering it. That’s a significant shift from 2024. What’s actually changed on the ground for firms of the size you work with?

The shift is real, but it’s not uniform. What I’ve observed working with UK and European mid-market firms over the past year is that the conversation has moved from “should we?” to “what and how?” — which is a different problem entirely. A year ago, partners were asking me whether AI was going to displace their juniors. Now they’re asking which workflow they should automate first, whether to buy a point solution or wait for their practice-management vendor to ship something, and how to explain the change to a team that’s already anxious about it. The question has become operational, which is a much healthier place to be than existential.

The Thomson Reuters data showed that 52% of the firms using GenAI are using open-source tools like ChatGPT rather than industry-specific software. That seems surprising.

It shouldn’t be, actually. It’s a sign of how far behind the industry-specific tooling is. When a tax senior needs to draft a client memo or summarise a regulatory change, they reach for the tool that works today — which for most firms is a generic LLM with a business subscription. The specialised tools exist, but they’re expensive, their training data is narrower, and their user experience often lags what consumers have been using at home for two years. That gap will close. But right now, open-source tooling is genuinely better at the long tail of accounting-knowledge-worker tasks than most of what’s being sold as “AI for accountants.”

Is there a risk problem with that? Client data running through consumer AI tools?

A significant one. I’ve done governance reviews for three firms in the last year where I discovered staff were pasting client-identifiable information into consumer ChatGPT accounts to get help drafting something. Not out of malice — just convenience. The firms had no written policy about it. This is one of the areas where smaller firms are actually at more risk than the Big Four, because the Big Four have enterprise licensing agreements with data protection terms that consumer tools don’t offer. The first question I ask any firm deploying GenAI isn’t “what’s your use case” — it’s “what’s your policy on client data entering a model, and how do you enforce it?”

— The numbers behind the conversation —

Table I — GenAI Adoption Trajectory in Tax & Accounting, 2024 → 2025
Status	2024	2025	Direction
Actively using GenAI	Lower base	~21%	Rising
Planning or considering	Roughly half	~53%	Broadly stable
No plans to use	~49%	~25%	Falling fast
Using open-source tools (of adopters)	Not measured	~52%	Dominant in 2025
Using industry-specific tools	Not measured	~17%	Early stage

Source: 2025 Generative AI in Professional Services Report (Thomson Reuters Institute).

— The Big Four Playbook —

Proprietary platforms, enterprise agreements, and a five-year head start.

Deloitte, EY, PwC, KPMG — they’ve all been public about their AI investments. Deloitte’s audit platform has agentic capabilities. EY launched a unified AI platform in 2023 and has announced capabilities supporting 160,000+ global audit engagements. PwC claims 20–50% productivity gains in their internal development. How much of this is real, and how much is marketing?

It’s mostly real, but heavily caveated. The Big Four have spent the last four years investing billions in proprietary audit and tax platforms. The capabilities are genuine. But the productivity numbers they cite tend to be measured in narrow domains where the tooling has been optimised — not across the whole engagement. PwC’s 20–50% gain on development productivity is probably true for specific code-generation and code-review tasks; extrapolating that to “audit engagements run 30% faster” would be a mistake. The tooling is real. The uniform productivity transformation is more measured than the press releases suggest.

What about KPMG’s Trusted AI framework and the governance approach? Does that scale down to smaller firms?

The framework itself does, conceptually. The staffing around it doesn’t. KPMG’s framework works because they have dedicated ethics, risk, and technology teams reviewing model deployments and controls. A fifteen-partner firm can adopt the principles — transparency, human oversight, audit trails, bias monitoring — but they can’t replicate the apparatus. What I recommend to mid-market firms is to adopt a simplified, practical version: a one-page policy, a named AI risk owner, quarterly review of what’s being used, and mandatory training. It’s 5% of what the Big Four do but captures 60% of the value.

The Big Four are spending billions. Realistically, how far behind do their proprietary tools put everyone else?

On audit-specific tooling, probably three to five years. But that gap is narrowing fast, and the commercial equivalents — the CoCounsels and the Caseware AI modules and the Intuit Practice AI tools — are close enough that the directional capabilities are available to anyone willing to buy a subscription. What the Big Four actually retain is not the capability advantage. It’s the integration advantage — the ability to build a tightly-coupled platform where the audit tool talks to the tax tool talks to the advisory tool talks to the client portal. That integration is what’s genuinely hard to replicate, and that’s what keeps the moat around Big Four engagements for the largest clients.

The Big Four don’t really have a capability advantage any more. They have an integration advantage — and for most engagements, that’s a different kind of moat entirely.

Rebecca Kahn

— The Smaller Firm Advantage —

Five workflows where mid-market firms are already competitive.

Let’s flip the question. Where are smaller firms actually holding their own, or even winning?

Five places, consistently. Tax research is the first and strongest — a well-configured AI research tool puts a sole practitioner on surprisingly close footing with a Big Four tax senior for routine research questions. Tax return preparation is the second; the commercial tooling has closed most of the gap. Tax advisory work is the third — AI-assisted scenario modelling lets a small firm offer genuinely strategic conversations they couldn’t previously price. Bookkeeping automation is the fourth, and probably the most transformative for the smallest firms. And document summarisation is the fifth — contracts, invoices, receipts, source documents. It’s unglamorous work, but it’s where most of the labour savings actually sit.

Of those five, which has the clearest ROI?

Bookkeeping automation, without much hesitation. A firm doing monthly bookkeeping for thirty to fifty small-business clients can realistically cut labour on those engagements by 40 to 60 percent using a combination of OCR, ML-based transaction classification, and direct bank feed integration. The savings are immediate, measurable, and survive audit. The other four use cases deliver real value but the ROI is harder to quantify. Bookkeeping is the one where you can show the partners a time-savings number at the end of the quarter and point to it.

Is the ChatGPT-for-tax-research workflow really production-grade?

For first-draft work, yes. For final work, absolutely not — and every tax professional I work with understands this. The workflow that works is using GenAI to produce a first pass, then reviewing it against authoritative sources, then editing it for accuracy and client-specific context. Time savings on a typical research memo are probably 40 to 60 percent. The risk, and it’s a real one, is a junior analyst who skips the review step because the first draft reads convincingly. That’s where the governance policy matters.

The Thomson Reuters data showed 44% of firms using GenAI are using it daily or multiple times a day. That’s extraordinary penetration for a technology most firms had never heard of two years ago.

It is, and it tells you something important. These aren’t firms experimenting. These are firms that have integrated the tool into their daily workflow in a way that would be painful to remove. That’s the threshold at which technology genuinely changes a profession — not when it’s adopted, but when removing it would cause real operational disruption. We crossed that threshold for GenAI in tax and accounting somewhere in late 2024. Most of the profession hasn’t quite realised it yet.

— Where the time savings actually sit —

Table II — Top Five GenAI Use Cases in Tax & Accounting Firms, 2025
Use Case	Typical Labour Saving	Maturity
Tax research	40–60% on research memos	Production-grade
Tax return preparation	30–50% on routine returns	Mature
Tax advisory & scenario modelling	Variable; widens engagement scope	Emerging
Bookkeeping automation	40–60% on recurring engagements	Production-grade
Document summarisation	50–70% on contract/invoice review	Mature

Labour-saving estimates are practitioner-observed ranges from mid-market firm deployments. Actual results depend heavily on data quality, workflow design, and staff training.

— What Actually Stops Firms —

The three barriers that matter, and the three that don’t.

When a firm comes to you saying they’ve tried to deploy AI and it hasn’t worked, what’s usually the real reason?

One of three things, almost always. First: they skipped the data cleanup. They bought a tool, pointed it at a messy chart of accounts and a vendor master full of duplicates, and were disappointed when the outputs were unreliable. The tool wasn’t the problem. Second: they tried to roll it out firmwide at once, without a single well-run pilot to validate the workflow. Third: they didn’t train the team properly, so partners end up with junior staff producing AI-generated outputs they don’t know how to review. The barriers are almost never the technology itself. They’re operational.

And the barriers firms worry about that actually don’t matter much?

Cost is the big one. Firms agonise over whether they can afford a $15,000 annual subscription when the actual question is whether they can afford to be eighteen months behind their competitors on workflow. Licensing is also a common source of anxiety that evaporates on examination — most enterprise tools have reasonable data-protection clauses, even if reading them feels painful. And the “will it work with our stack?” question matters less than firms think, because almost every serious vendor has invested heavily in integrations.

What about the CPA shortage? The US is facing approximately 75,000 fewer accountants entering the profession than the industry needs. How does that interact with adoption?

It’s accelerating everything. Firms that dragged their feet on automation through the 2010s are now aggressively catching up because they cannot hire. Every new technology conversation I have with a mid-market firm starts, or ends, with staffing. Automation is no longer an optimisation question; it’s a survival question. That’s also changing the kind of firms that are investing. Five years ago, tech-forward firms were the exception. Today, firms that aren’t investing in automation are the exception, and they’re mostly ones that have decided to shrink rather than adapt.

Is this a specifically American dynamic, or are you seeing it elsewhere?

Everywhere, with local variations. UK practices are dealing with the same staffing pressure. Nordic markets — Sweden, Norway, Denmark — have been ahead of the curve on digital accounting for years because their tax authorities pushed early for digital filing and real-time reporting. That actually gave Nordic firms a head start on data hygiene, which is why Swedish practices like redovisningsbyrå Kungsholmen — working in a market where digital bokföring has been the default for a decade — often have cleaner data foundations for AI deployment than equivalent US or UK firms. The technical readiness is actually better in some of those markets than in the ones where the vendors are headquartered.

That’s counter-intuitive. I’d have guessed the opposite.

Most people do. The assumption is that the US leads because that’s where the platforms are built. But platform availability and practice-level readiness are different things. A Swedish firm that has been running fully digital bookkeeping since 2014 has ten years of clean, structured data. A UK firm that migrated from desktop Sage in 2022 is still cleaning up the legacy. When you deploy an AI transaction-classifier, the Swedish firm gets usable output in the first month. The UK firm spends six months on data remediation first. Geography matters less than people think. Digital maturity matters more.

— What firms report vs what actually blocks them —

Table III — AI Adoption Barriers: Reported vs. Operational Reality
What Firms Say	What’s Usually the Real Issue
“The tools are too expensive”	Opportunity cost of not adopting is usually larger
“Our data isn’t ready”	True — but the cleanup is the project, not a blocker to it
“Our staff will resist”	Mostly a training and framing problem, not a will problem
“We need to wait for better tools”	Current tools are already production-grade for key workflows
“Integration will be painful”	Real concern; largely solved by modern API-first vendors
“We don’t have the expertise”	Valid; external implementation support closes this fast
“We’re too small”	Smaller firms often adopt faster due to lower decision friction

— Looking Ahead —

Agents, end-to-end audit, and what breaks next.

PwC has publicly suggested an end-to-end AI-driven audit solution by 2026. How seriously should the rest of the profession take that?

Seriously, but not literally. “End-to-end AI-driven audit” is a marketing phrase. What PwC and the others will actually ship by 2026 is a tightly integrated suite where AI-assisted tools handle substantially more of the audit workflow than they do today — risk assessment, sampling decisions, document review, anomaly detection, first-draft workpaper preparation. Human judgement will still sign off on every significant conclusion, because professional standards require it. But the ratio of human-hours to audit-quality outputs will shift meaningfully. The signoff won’t go away. The work leading up to it will compress.

Agentic AI — systems that can execute multi-step workflows autonomously — is the newest hype cycle. Is it real for accounting?

In narrow workflows, yes. An agent that can receive an invoice email, extract the data, match it against a PO, route it for approval, and schedule payment — that works today in well-instrumented firms. Deloitte is shipping agentic capabilities in their audit platform; they’re not lying about it. What doesn’t work yet is general-purpose autonomy across diverse, ambiguous tasks. An agent can handle AP end-to-end. It can’t handle “draft our Q3 advisory letter and send it to the client” without human intervention at multiple steps. The envelope expands every six months. But the honest 2026 assessment is that agentic systems are production-ready for constrained workflows and demo-stage for unconstrained ones.

What breaks next? If we’re having this conversation in two years, what will we be saying has shifted?

Three things. First, the pricing model of accounting engagements will have moved decisively away from hourly billing. It’s already breaking; by 2027 it will be broken. Firms still running hourly-billed compliance work will be losing meaningful market share to firms offering fixed-fee, AI-enabled alternatives. Second, the regulatory frameworks will have caught up. The PCAOB and AICPA are actively writing guidance on AI use in audit; in two years, there will be explicit rules about documentation, human oversight, and liability. Third, the consolidation of mid-market firms will be happening noticeably faster. Firms that invested in automation will be acquiring firms that didn’t, at attractive valuations. That’s a pattern I’m already watching begin in the UK market.

Last question. A managing partner reads this interview tomorrow morning. What should they do on Monday?

Three things, in order. First, walk around the office and find out what generative AI tools your staff are actually using today — you will be surprised by the answer, and the answer matters for your data protection position. Second, pick one workflow where you already know you have a bottleneck — bookkeeping, document review, tax research, whichever — and commit to a structured sixty-day pilot on a commercial tool. Don’t try to do more than one at once. Third, write a one-page AI policy that covers client data, disclosure, and mandatory human review. You can refine it later. The worst governance posture is the one you don’t have at all.

— Editor’s Note —

The interview was recorded over two conversations in March and April 2026.

Rebecca Kahn consults independently with accounting and professional-services firms on practice technology strategy. She has no commercial relationship with any of the vendors mentioned in this interview, and the views expressed are her own. Statistics referenced come from the 2025 Generative AI in Professional Services Report published by the Thomson Reuters Institute, with supplementary practitioner estimates where noted.

AWSCPA Journal remains editorially independent of all vendors and consultancies referenced in our coverage. When we cite a specific platform, framework, or practice, it is because we believe the reference serves our readers — not because we were compensated for it. For coverage of how regulators are approaching the same questions, see our earlier brief on the Bank of England and FCA’s 2024 survey of AI in UK financial services.

The post The Big Four Don’t Really Have a Capability Advantage Any More: An Interview on AI Adoption in Accounting Firms appeared first on AWSCPA Journal.

Stanford Study: AI Is Making Accountants Faster and More Detailed, Not Obsolete

AWSCPA Journal — Fri, 17 Apr 2026 11:59:55 +0000

Research Brief · Academic · 5 min read

On the paper

Human + AI in Accounting: Early Evidence from the Field

Jung Ho Choi (Stanford GSB) · Chloe Xie (MIT Sloan)

One of the clearest empirical studies to date on what happens when working accountants actually use generative AI in production. The findings complicate the replacement narrative considerably.

For most of the past decade, accounting has occupied an unflattering position at the top of every list of professions most exposed to automation. The reasoning has usually been simple to the point of being lazy: accounting involves a lot of repetitive, rules-based work, therefore computers will do it, therefore accountants are in trouble. The narrative has been widespread enough to affect enrolment in accounting programmes and to raise genuine anxiety across the profession.

A recent working paper by Jung Ho Choi, an assistant professor of accounting at Stanford Graduate School of Business, and Chloe Xie of MIT Sloan, provides some of the first serious empirical evidence on whether that narrative holds up. It does not, at least in the form it has most often been told. What the data actually shows is more interesting — and more useful — than the simple replacement story.

The study combined survey responses from 277 accountants with task-level operational data from 79 small and mid-sized firms actively using AI-powered accounting tools. The analysis examined how the introduction of AI affected both the volume of work an individual accountant could handle and the quality of the work they produced. What emerges is a picture of genuine augmentation rather than substitution — with important caveats about who benefits, and by how much.

— The Four Headline Findings —

7.5 days

Faster monthly close for accountants using AI tools vs. traditional methods

8.5%

Less time spent on routine back-office processing

+12%

Increase in reporting granularity at firms using generative AI

62%

Of surveyed accountants are concerned about AI-generated errors

FINDING 01

AI-using accountants serve more clients, not fewer.

The strongest single result in the paper is a capacity expansion rather than a labour displacement. Accountants equipped with generative AI tools were able to support more clients per week than their peers using traditional methods, while simultaneously completing monthly statement preparation approximately 7.5 days faster. Routine back-office processing time fell by 8.5%. The time recovered by those efficiency gains did not disappear; it was redirected.

The redirection itself is what matters. Accountants did not use the recovered hours to process more transactions faster. They used them for the parts of the job that had previously been squeezed: business communication, quality assurance, client-facing advisory conversations. In the authors’ framing, AI helped with the “setup” — pulling information, connecting bank transactions, tracking vendor data — and that setup work was precisely the bottleneck preventing accountants from operating at higher client volumes.

FINDING 02

Quality did not fall. It rose.

The most counter-intuitive result in the study is that the throughput gains did not come at the usual cost of quality. If anything, the opposite. Firms using generative AI tools posted a 12% increase in reporting granularity — the level of detail captured in their bookkeeping records. Rather than grouping expenses into broad catch-all buckets such as “payroll,” AI-assisted firms broke payroll down into specific sub-categories: bonuses, benefits, meals, and similar. The resulting financial reports are more informative, easier to audit, and more useful for the underlying business to act on.

This matters more than it first appears. Most adoption studies of new technology in knowledge work show a short-term quantity-for-quality trade-off as practitioners climb the learning curve. The Choi and Xie findings suggest that, in accounting bookkeeping specifically, the technology is well-matched enough to the work that the trade-off either does not appear or is smaller than expected. The authors attribute this to the augmentation model — the AI is handling the tedious categorisation work that accountants were rushing through, freeing the human to apply judgement at a higher resolution.

The technology is not here to replace the human being — it’s here to augment the experts who are already in place.

Chloe Xie · MIT Sloan School of Management

FINDING 03

Seniority determines how much you gain.

One of the study’s more consequential findings for how firms structure AI deployment is that gains are not evenly distributed across experience levels. Senior accountants extracted substantially more value from AI tools than junior staff did — despite being, on many metrics, the group less in need of efficiency help.

The mechanism is straightforward. Senior accountants treated the AI as a collaborator whose outputs required verification. They noticed when the model’s confidence dropped on a particular suggestion and stepped in to apply human judgement where it was most needed. Junior staff, lacking the pattern recognition that comes with years of experience, were more likely to accept AI-generated outputs at face value — including outputs that the system itself had flagged as uncertain. The net result: senior accountants extracted both the efficiency and the quality gains; junior accountants captured efficiency but with a higher residual error rate. The research offers an early empirical look at how AI-generated errors can propagate through human-in-the-loop accounting systems, and where they are most likely to slip through.

FINDING 04

Accountants themselves are cautiously positive.

The survey side of the study captured what working accountants actually think about the tools they are now using. The answer is not uniformly enthusiastic, but it is more constructive than the popular narrative would suggest. Sixty-two percent expressed concern about AI-generated errors. Forty-three percent worried about data security implications. Thirty-seven percent reported anxiety specifically about the effect on job stability.

Against those concerns, though, nearly half of surveyed accountants reported that generative AI tools helped them meet deadlines more reliably and improved the accuracy of their work. Almost two-thirds identified the automation of routine tasks as the single biggest benefit of adopting AI. The picture is of a profession that is neither naively enthusiastic nor reflexively defensive — accountants see real benefits from the technology and also see real risks, and they are engaging with both.

— What the Findings Do Not Yet Tell Us —

The next frontier is audit, tax, and valuation.

The study’s sample is concentrated on bookkeeping activities — the foundational recording and organisation of day-to-day financial transactions. That is where the tools are currently mature and where the operational data exists to study. More complex and higher-stakes domains of accounting work — external audit, tax strategy, business valuation, complex advisory — remain largely untouched by the research, for the simple reason that they remain largely untouched by the technology.

That gap will close over the next survey cycle. Audit platforms are already beginning to incorporate AI for document review and anomaly detection. Tax tools are integrating LLM-based research assistants. Valuation work, with its heavy reliance on contextual judgement, is furthest behind. Whether the augmentation finding generalises to those higher-stakes domains — where the cost of an undetected error is much larger — is the empirical question the profession needs answered next.

— Reader Questions —

Eight questions on the study.

What is the main finding of the Choi and Xie study?

That generative AI acts as an augmentation tool rather than a replacement for accountants. AI-using accountants supported more clients, closed the books faster, and produced more granular reports — without a measurable trade-off in quality.

How large was the sample?

Survey responses from 277 accountants combined with task-level operational data from 79 small and mid-sized accounting firms that use AI-powered tools.

How much faster do AI-using accountants close the books?

Approximately 7.5 days faster than accountants using traditional methods, with an 8.5% reduction in time spent on routine back-office processing.

Does report quality suffer when AI is used?

No. Firms using generative AI tools posted a 12% increase in reporting granularity — breaking expenses into more specific categories rather than grouping them into broad buckets. Quality rose rather than fell.

Who benefits most from AI tools in accounting?

Senior accountants. They treat the AI as a collaborator, verify its outputs, and step in where model confidence is low. Junior staff see smaller gains because they are more likely to accept AI outputs at face value, including uncertain ones.

What are accountants most worried about?

Error propagation was the top concern at 62%, followed by data security at 43% and job stability at 37%. These concerns coexist with positive views of the efficiency and accuracy benefits.

Which accounting tasks is AI actually being used for today?

Primarily bookkeeping — transaction classification, reconciliation, and routine record-keeping. More complex domains such as external audit, tax strategy, and valuation remain largely untouched by the current generation of tools.

What should firm leaders take from this?

Three things. Deploy AI to expand capacity rather than to cut headcount. Pair AI tools with senior reviewers, not junior ones, where output quality matters. And treat the current bookkeeping-focused rollout as the first frontier, not the last — the higher-value domains are coming next.

The post Stanford Study: AI Is Making Accountants Faster and More Detailed, Not Obsolete appeared first on AWSCPA Journal.

AI in Public Accounting 2026: What the Technology Actually Does, and What It Doesn’t

AWSCPA Journal — Fri, 17 Apr 2026 11:43:32 +0000

Feature · Technology · 14 min read

The hype cycle is over. What remains is the far more interesting question of what artificial intelligence actually does inside a working CPA firm in 2026 — where it saves time, where it creates risk, and where it still cannot be trusted without a human in the loop.

I. The Situation

From press releases to production workloads.

Two years ago, roughly half of the conversations we had with mid-market CPA firms about artificial intelligence began with some variation of “we’ve been looking at it.” In 2026, the conversation has changed. Firms no longer describe themselves as exploring. They describe themselves as deploying, mid-deployment, or mid-rollback after a deployment that did not deliver. The category has moved from the pilot-programme slide deck into the general ledger itself.

The market numbers support what practitioners are seeing on the ground. Independent market research now places the global AI-in-accounting market on a trajectory toward $10.87 billion in 2026, compounding at approximately 44.6% annually, with small and mid-sized enterprises driving the pull rather than enterprise customers. That last detail is the most important one. A decade ago, advanced automation was a Big Four advantage. It is now available, in usable form, to the twenty-partner firm that signed up for a $400-per-month subscription last quarter.

The operational claims made by vendors are, in some cases, substantial. Well-deployed systems are credibly reducing individual tax-return preparation labour by more than eighty percent. Month-end close cycles are shrinking from weeks to days, with the most advanced adopters pushing toward daily or near-real-time reporting. Audit teams are running anomaly detection across entire transaction populations rather than sampling five to ten percent of activity. None of this is science fiction. All of it is being done, today, inside firms that readers of this publication would recognise.

What follows is a practical, non-promotional assessment of where AI in accounting actually stands in 2026 — where it works, where it doesn’t, what it costs, who is buying it, and what the next two years look like. We draw on vendor data where it is defensible, dismiss it where it isn’t, and attempt to give partners, controllers, and IT directors a clearer picture than the average vendor webinar provides.

II. The Headline Numbers

The state of the market, at a glance.

Table I — AI in Accounting: 2026 at a Glance
Metric	Value	Context
Global AI accounting market, 2026	~$10.87B	Projected market size; SME-driven growth
Market CAGR	~44.6%	SME adoption the dominant driver
UK financial services firms using AI	~75%	Per Bank of England / FCA survey, 2024
Additional UK firms planning deployment	~10%	Within the next three years
Weekly time savings, typical adopter	~5.4 hrs	Gross time savings per knowledge worker, Gartner
Individual tax returns automatable	>80%	Best-in-class deployments
Document analysis time reduction	~50%+	Audit and advisory use cases
Firms reporting skills-gap barrier	~58%	Finance-department self-assessment

III. The Maturity Curve

Leaders, implementers, and beginners.

The most useful framework for thinking about where a firm stands on AI adoption comes from KPMG’s Global AI in Finance Report, which segments the market into three distinct maturity tiers. The tiering is less about technology sophistication than it is about organisational commitment — a leader firm is not necessarily one that bought the most advanced tool, but one that made the operational, cultural, and process changes needed to use that tool across multiple workflows at scale.

The distribution is striking and, for most readers, likely flattering. Only 24% of surveyed finance functions qualify as leaders. 58% — the great middle — are implementers: firms that have deployed AI in one or two specific functions but have not yet achieved integrated, cross-functional adoption. The remaining 18% are beginners, still running pilots or evaluating options. Most firms reading this publication belong, honestly, somewhere in the implementer category.

Table II — The Three Maturity Tiers of AI-in-Finance Adoption
Tier	Share of Firms	Defining Characteristics
Leaders	~24%	AI deployed across multiple workflows; measurable ROI; strategic alignment at partner level
Implementers	~58%	AI deployed in 1–2 functions; value demonstrated but not yet scaled; integration incomplete
Beginners	~18%	Pilot or evaluation stage; no production-grade deployment; skills and governance gaps

The useful question is not which tier your firm sits in today, but what would be required to move up one level. The jump from beginner to implementer is almost always a single, well-defined production deployment in a high-volume, rules-based workflow. The jump from implementer to leader is harder. It requires integration across functions, genuine process redesign, a functioning governance model, and a willingness by partners to change how work is priced and how junior staff are developed.

“

The jump from beginner to implementer is usually one good deployment. The jump from implementer to leader is a change in how the firm is run.

— Editorial Observation

IV. Where It Works

Seven workflows where AI is earning its keep.

A reasonable rule of thumb: AI delivers real, measurable value in any accounting workflow that combines high transaction volume with relatively well-defined rules. It struggles in workflows that require professional judgement under conditions of genuine ambiguity, where training data is sparse, or where the cost of an undetected error is high. The following seven areas have emerged, by 2026, as the strongest production use cases.

i. Bank reconciliation and transaction matching

Mature. Deployed at scale across the industry. Modern cloud accounting platforms now match the majority of routine transactions automatically, with human review reserved for exceptions. For a mid-sized firm, this single capability can remove ten to fifteen hours per week from the close process.

ii. Document classification and OCR

Mature in its core use cases — invoice ingestion, receipt capture, tax form extraction. Accuracy on standard document types now comfortably exceeds 95% for well-trained models on clean inputs. Accuracy degrades with handwritten annotations, non-standard layouts, and multilingual documents, which is where vendor claims of “99% accuracy” should be read sceptically.

iii. Individual tax return preparation

A genuine breakthrough category. Best-in-class deployments now automate more than 80% of the mechanical work of individual tax return preparation, including data gathering, form population, deduction identification, and consistency checks. The residual 20% — judgement calls, complex multi-state situations, and client-specific optimisation — remains firmly with the preparer. The economics of small tax practices have shifted meaningfully as a result.

iv. Audit anomaly detection

Moving rapidly from pilot to production. AI-assisted audit tools can now scan full transaction populations — not samples — and flag anomalies by amount, vendor, timing, and frequency. Duplicate payments, round-dollar transactions, out-of-business-hours postings, and unusual journal-entry patterns surface automatically. The auditor’s role shifts from running tests to interpreting the flags the system produces, which is genuinely higher-value work.

v. Accounts payable automation

A mature category with meaningful ROI. Intelligent invoice scanning, three-way matching, approval-workflow routing, and duplicate detection have collectively moved AP from a high-touch process to a low-touch one. Processing cost per invoice has fallen substantially for firms that adopted well-designed systems; manual error rates have fallen even more.

vi. Cash flow and variance forecasting

Emerging but promising. Predictive models trained on historical payment patterns, seasonality, and client-level behaviour are delivering meaningfully better cash-flow forecasts than traditional spreadsheet-based approaches. Accuracy improvements in the 30–45% range are credible for firms with clean, multi-year transaction data. For firms with messy data, results are considerably more variable.

vii. Research assistance for tax and advisory

Early production, accuracy still variable. Large language models trained on tax codes, accounting standards, and professional guidance now draft initial research memos, summarise regulatory changes, and answer technical questions in conversational form. Every serious deployment pairs the output with mandatory human review because the underlying models still occasionally produce confident-sounding but incorrect citations. Used properly, the time savings are substantial; used carelessly, the liability implications are significant.

Table III — AI Adoption Maturity by Accounting Function, 2026
Function	Maturity	Key Caveat
Bank reconciliation	Production-grade	Exception handling still requires human judgement
OCR / document ingestion	Mature	Accuracy degrades on non-standard layouts
Individual tax preparation	Mature	Review and judgement remain with preparer
Audit anomaly detection	Scaling rapidly	Interpretation of flags is the critical skill
Accounts payable	Mature	Integration with ERP is the main project cost
Cash flow forecasting	Emerging	Requires clean multi-year data
Tax research / advisory memos	Early production	Hallucination risk; mandatory human review
Advisory memo drafting	Early	Liability and professional-standards questions open
Autonomous bookkeeping	Demo stage	Not production-reliable for general use

V. Where It Doesn’t Work

The limits, plainly stated.

Every honest assessment of AI in accounting has to name the things it does not do well. The vendor pitch decks tend to treat these as engineering problems that will be solved next quarter. The operational reality is that several of them are likely to persist through the end of this decade.

It cannot exercise professional judgement. Deciding whether an accounting estimate is reasonable, whether a tax position has substantial authority, or whether an audit exception warrants further investigation requires judgement shaped by years of practice. AI can surface the facts. It cannot form an opinion that meets professional standards.

It fails silently on edge cases. A well-trained model will produce a confident, well-formatted output on an input that falls outside its training distribution. The output will look correct. It will not be correct. For high-stakes work, detecting this failure mode requires an experienced reviewer who knows what to look for.

It is only as good as the underlying data. Approximately 63% of early AI projects in accounting are delayed by data-quality issues. Firms with messy charts of accounts, inconsistent categorisation practices, or fragmented general ledgers find that the AI rollout exposes the underlying hygiene problem rather than solving it.

It does not replace entry-level training. The next generation of senior accountants will have learned the profession in an environment where AI handled much of the routine work. Whether they will have developed the same instinct for numbers that comes from manually preparing thousands of reconciliations is an open question the profession has not seriously grappled with.

VI. The Vendor Landscape

Four categories, four strategies.

The AI-in-accounting vendor landscape has consolidated into four reasonably distinct categories, each with a different approach to integration, pricing, and target customer. A firm evaluating tools in 2026 should understand which category it is actually buying from — the strengths and limitations are structural, not marketing.

i. Legacy enterprise ERPs with bolted-on AI

The SAPs, Oracles, and Microsoft Dynamics of the world. AI added to architectures designed before AI existed. Powerful, extensively integrated, and capable of handling the most complex multi-entity, multi-jurisdictional environments. Expensive to implement, slow to change, and constrained by backward-compatibility obligations that vendors cannot abandon. Best suited to firms over approximately $100 million in revenue.

ii. Cloud accounting platforms for SMEs

Xero, QuickBooks Online, Sage Intacct, FreeAgent, and their regional equivalents. Subscription-priced, rapid to deploy, and now shipping genuinely useful AI features in their standard tiers — reconciliation automation, receipt capture, expense categorisation. Limits emerge at the upper end of the SME range, where consolidation, complex revenue recognition, or advanced compliance needs outgrow the platform.

iii. Specialist automation vendors

Point solutions that sit on top of existing general ledgers — AP automation, AR workflow, intelligent document processing, audit anomaly detection, tax research assistants. Strong results in narrow domains; integration burden grows with the number of tools adopted. Most mid-sized firms end up with four to eight of these, which creates a meaningful integration and vendor-management cost.

iv. AI-native accounting platforms

The newest category. Platforms designed from day one around AI-first workflows rather than retrofitting intelligence onto existing ledger software. Smaller footprint than enterprise ERPs, greater capability than SME platforms, fewer integration gaps than specialist stacks. Trade-off: these are younger products with shorter deployment track records, and buyer diligence on vendor stability, data portability, and implementation support matters proportionally more.

Table IV — Vendor Categories at a Glance
Category	Typical Price Point	Best For
Legacy enterprise ERPs	$500K–$5M+ TCO	Firms >$100M revenue, multi-entity global operations
Cloud SME platforms	$50–$500/month	Sole practitioners through mid-sized firms
Specialist automation	$100–$2K/month per tool	Firms extending capabilities on existing GL
AI-native platforms	$10K–$200K+/year	Mid-market firms rebuilding workflows from scratch

VII. Implementation

Six phases of a disciplined rollout.

Define the problem

Identify a specific workflow that is painful, measurable, and well-understood — slow reconciliation, AP backlog, manual tax data entry. Set a baseline with real numbers. If the team cannot say how long the current process takes, the project is not ready to start.

Choose a pilot

Pick one workflow, one team, one measurable target. A sixty-day pilot with a clear success criterion beats a six-month firm-wide rollout with vague goals. Bank reconciliation and expense categorisation are time-honoured first projects for reason.

Clean the data

Most AI project delays are data-quality problems wearing a different hat. Before deployment, fix the chart-of-accounts drift, resolve vendor-master duplicates, and reconcile the client list between systems. The cleanup is tedious and non-optional.

Train the team

A new tool with an untrained team produces worse results than an old tool with a trained one. Allocate proper training time. Name internal champions. Accept that productivity will dip for two to three months before it rises.

Governance from day one

Role-based access, detailed audit logs, a defined human-in-the-loop review point for high-risk outputs, and a documented policy on AI use in client-facing work. SOC 2 Type II certification and credible data-portability terms should be minimum vendor requirements.

Measure and expand

Track the KPIs set at Phase 01. If the pilot hit its target, expand to an adjacent workflow. If it didn’t, understand why before spending more. Most successful firm-wide rollouts are four to six successful pilots stacked sequentially, not a single Big Bang project.

VIII. The Barriers

What actually stops firms from getting value.

Table V — Common Failure Modes in AI Implementation
Barrier	Frequency	Mitigation
Skills gap in finance team	~58%	Structured training, internal champions, phased rollout
Data quality issues	~63% of early delays	Pre-implementation data cleanup; chart-of-accounts discipline
Legacy system integration	Widespread	API-first vendor selection; accept modernisation cost
Internal resistance to change	Common	Frame as augmentation; show early wins; protect roles
Unclear success metrics	Very common	Define KPIs at project start; baseline before deployment
Vendor lock-in	Underrated	Require data-portability commitments; open APIs
Governance gaps	Material liability risk	Policy, audit logs, human-review checkpoints, SOC 2 Type II

“

Most AI project delays in accounting are data-quality problems wearing a different hat.

— The Implementation Rule

IX. The Consequences

What it means for the profession.

The clearest near-term effect of AI on the profession is not that accountants will be replaced. It is that the economic structure of an accounting firm — specifically, the pyramid that has organised the industry for the past seventy years — will change shape. A traditional firm has a wide base of junior staff doing high-volume, low-judgement work, narrowing upward through seniors, managers, and partners. That pyramid is already flatter at the bottom in firms that have deployed AI aggressively, and will continue to flatten.

This has second-order consequences that the profession has not yet seriously addressed. If a first-year accountant no longer spends a thousand hours doing reconciliations and bookkeeping, how does that person develop the number sense that traditionally defined a competent accountant? If junior hiring contracts, where does the next generation of partners come from? These are strategic questions for managing partners, not technology questions.

New roles are also appearing. “AI accounting analyst,” “AI financial reporting specialist,” and “AI risk and controls specialist” are now genuine job titles rather than consultant invention. The common thread is the ability to validate, interpret, and explain AI-generated work — a skill set that sits between traditional accounting and data science. Firms that figure out how to grow this capability internally will have a durable advantage over those that have to hire it in at market rates.

At the CFO and managing-partner level, AI strategy has become something leaders can no longer delegate. The decisions about which tools to adopt, how to govern them, how to train the workforce around them, and how to reprice engagements as the labour content shrinks — these are business strategy decisions, not IT procurement decisions. The firms that treat them that way will pull steadily ahead of the ones that don’t.

X. What Comes Next

Three things to watch through 2027.

Agentic workflows reaching production reliability. “Agentic AI” — systems that can decompose a multi-step task, pick tools, execute, and deliver without step-by-step supervision — is the category where vendors are currently loudest and results are currently most variable. Demos are impressive. Production reliability on real client work is still uneven. The honest assessment for 2026 is that agentic systems work well inside well-constrained workflows with clean data and break in predictable ways outside them. By 2027, that envelope will likely be meaningfully wider, but it will remain finite.

Regulatory frameworks catching up. The PCAOB, the AICPA, and international standard-setters are actively developing guidance on the use of AI in audit, assurance, and tax practice. Rules that are currently permissive or silent will, over the next eighteen months, become explicit about what constitutes adequate human oversight, how AI use should be documented, and what professional liability attaches to AI-generated output that a reviewer approved. Firms that build their processes now with those frameworks in mind will have significantly less rework to do when the rules land.

The CPA pipeline problem meeting the automation wave. The United States has approximately 75,000 fewer entrants to the profession than demand would support, and the gap is not closing. The interaction between this shortage and the automation wave is the single most important dynamic in the profession’s next five years. Firms that use AI to do more with fewer people will continue to grow. Firms that use it to preserve billable hour models while cutting costs will continue to lose talent to the first category. The competitive reshuffling has already begun.

XI. Reader Questions

Twenty-five questions, answered plainly.

What does “AI in accounting” actually mean in 2026?

It refers to the use of machine learning, natural language processing, and increasingly agentic systems to automate or augment accounting workflows — from bank reconciliation and document ingestion to tax preparation and audit anomaly detection. The term now describes production deployments, not pilots.

Is AI going to replace accountants?

No. It is restructuring the profession rather than eliminating it. Routine, high-volume work is being automated. High-judgement work — audit sign-offs, complex tax planning, advisory engagements, regulatory interpretation — remains firmly with CPAs. The pyramid of the profession is getting narrower at the bottom.

What are the strongest use cases right now?

Bank reconciliation, document OCR and classification, individual tax return preparation, accounts payable automation, and audit anomaly detection. All five combine high transaction volume with well-defined rules — the combination AI handles best.

How much time does AI actually save?

Independent measurements from Gartner place average weekly gross time savings at approximately 5.4 hours per knowledge worker. In specific workflows — AP processing, reconciliation, individual tax prep — the savings can be much larger. Actual net savings depend heavily on implementation quality and data hygiene.

Is the $10.87 billion market number credible?

It is the figure produced by Mordor Intelligence’s 2025 research for the global AI-in-accounting market in 2026. Like all market-sizing exercises it involves assumptions. The order of magnitude is consistent with other independent research and with observable vendor revenue.

Why are SMEs driving the growth rather than enterprise firms?

Because AI has become accessible at subscription prices through cloud platforms. A decade ago, serious automation required enterprise budgets and dedicated IT staff. Now a twenty-partner firm can deploy sophisticated workflows through monthly subscriptions and no-code configuration.

What’s the biggest implementation barrier?

Data quality. Approximately 63% of early AI projects are delayed by underlying data issues — messy charts of accounts, vendor-master duplicates, inconsistent categorisation. The AI deployment exposes the hygiene problem rather than solving it.

How should a small firm start?

Pick one workflow that is painful, measurable, and rules-based — typically bank reconciliation or expense categorisation. Set a specific target. Run a sixty-day pilot. Expand only after the pilot produces measurable results. Most firm-wide rollouts are a sequence of successful small pilots.

How much of an individual tax return can actually be automated?

In best-in-class deployments, more than 80% of the mechanical work — data gathering, form population, consistency checks, deduction identification. The remaining 20% — judgement calls, complex multi-jurisdiction situations, client-specific optimisation — requires a preparer.

Is audit sampling going away?

In well-instrumented engagements, substantially. AI-assisted audit tools can now scan full transaction populations rather than samples. The auditor’s role shifts from test execution to anomaly interpretation. Regulatory frameworks are still catching up to what is technically possible.

What is “agentic AI” and does it work?

Agentic AI refers to systems that can decompose a multi-step task, select tools, execute the work, and produce a deliverable without step-by-step supervision. Demos are impressive. Production reliability is uneven and depends heavily on workflow constraints and data quality. Genuinely reliable agentic accounting workflows exist in narrow domains; broad autonomy is not yet here.

Can AI prepare financial statements end-to-end?

Technically, yes — for well-structured companies with clean data. Practically, a human reviewer is still required before any statement is signed off. The claim of 95%+ accuracy on standard statements is defensible, but the remaining 5% is exactly where professional liability lives.

What about hallucinations in AI-drafted tax research?

Real and material. Large language models occasionally produce confident-sounding citations to cases or statutes that do not exist. Every serious deployment pairs AI-drafted research with mandatory human verification. Firms that have skipped this step have created professional liability exposure.

Should we build in-house AI capability or buy it?

For almost all firms: buy it. Building in-house LLM capability is an expensive, specialised undertaking that makes sense only at the largest scale. Commercial platforms have closed most of the capability gap and compress total cost of ownership significantly.

What percentage of technology budget is AI commanding now?

In progressive firms, between 10% and 25% of technology budget is allocated to AI initiatives, including licensing, implementation, and training. The range is wide because maturity varies. Firms that treat AI as a line item underperform firms that treat it as a strategic category.

What governance framework should we put in place?

At minimum: documented policy on AI use in client-facing work, role-based access controls, audit logs of AI-generated outputs, defined human-review checkpoints for high-risk outputs, SOC 2 Type II vendor certification, and a data-portability agreement. These are baseline expectations, not advanced ones.

What new roles are emerging?

“AI accounting analyst,” “AI financial reporting specialist,” and “AI risk and controls specialist” have become genuine job titles. The common skill set is the ability to validate, interpret, and explain AI-generated work — sitting between traditional accounting and data analytics.

Will AI affect how engagements are priced?

Already is. Fixed-fee and value-based pricing arrangements perform much better under AI-enabled operations than hourly billing does, because the labour content of a given engagement falls sharply. Firms clinging to hourly billing will either reprice or lose work to firms that have switched.

How is the CPA shortage affecting all of this?

It is accelerating adoption. With approximately 75,000 fewer professionals entering the US profession than demand supports, firms face a choice: pay substantially more for scarce talent, turn away work, or automate. The third path is the only scalable one, which is why automation-resistant firms have quietly begun catching up.

What is a reasonable payback period for an AI implementation?

In well-chosen pilots — reconciliation, AP automation, expense categorisation — payback can be under twelve months. For larger platform migrations, two to three years is more realistic once training, data cleanup, and process redesign are included.

What is the biggest mistake firms make?

Buying the tool before fixing the process. Deploying AI onto a broken workflow produces faster broken work. Successful firms redesign the underlying process first, then deploy the tool to support the new design.

Does the Big Four have a decisive advantage here?

In proprietary tooling, yes. Deloitte, EY, KPMG, and PwC invest billions annually in internal platforms smaller firms cannot match. In direction of travel, less so — commercial equivalents typically reach the mid-market within two to five years. The practical implication is that smaller firms follow the trajectory rather than match the tooling.

What about cyber security risks?

Real and rising. CPA firms hold enormous quantities of sensitive financial data, making them attractive targets. Cyber insurance premiums now price firms on their security posture, which has effectively made MFA, endpoint protection, and staff training non-optional. Any AI deployment should be treated as an expansion of the attack surface.

What should we watch over the next twelve months?

Three things: agentic AI reaching production reliability in narrow domains, regulatory frameworks from the PCAOB and AICPA crystallising around AI use in audit and tax, and the continuing interaction between the CPA shortage and automation-driven margin compression.

Is this a bubble or is it real?

The hype cycle around “AI transforms everything” is a bubble and will deflate. The underlying technology change — automation of high-volume, rules-based accounting work — is real, durable, and already reshaping the profession. The firms that distinguish one from the other will navigate the next five years better than the firms that do not.

The post AI in Public Accounting 2026: What the Technology Actually Does, and What It Doesn’t appeared first on AWSCPA Journal.

The AI Leaders Gap: What the Top Quartile of Finance Functions Do Differently

AWSCPA Journal — Tue, 20 May 2025 11:50:12 +0000

Analysis · Finance Leadership · 15 min read

Seventy-one percent of finance functions globally are now using artificial intelligence in some form. But the gap between the firms genuinely winning with it and the ones merely deploying it has widened sharply — and the interesting question is no longer whether to adopt, but what separates the leaders from everyone else.

I. The Gap

Seventy-one percent, and widening.

The most recent round of global survey data on artificial intelligence in finance tells a clearer story than any vendor deck ever will. Seventy-one percent of finance functions across twenty-three industrialised and emerging markets are now using AI in some form. Within the next three years, that figure will approach universality. The technology is no longer a signal of innovation; it is becoming the baseline expectation of how finance teams operate.

What changes beneath that headline number, however, is far more interesting. The firms surveyed were grouped into three maturity tiers: beginners, implementers, and leaders. Eighteen percent of firms are beginners, still piloting and evaluating. Fifty-eight percent are implementers, deploying AI in one or two functions but not yet achieving integrated firm-wide adoption. Twenty-four percent are leaders — firms that have embedded AI across finance and are generating compounding returns from that investment.

The gap between the leaders and everyone else is not, as one might assume, primarily a technology gap. It is an investment gap, a governance gap, a talent gap, and above all a strategic-intent gap. Leaders spend more, deploy across more functions, institutionalise governance earlier, and pull assurance from their auditors into the AI lifecycle itself. They are pulling ahead fast enough that the gap will likely continue to widen through the end of the decade.

This article is a close reading of what, specifically, distinguishes the leaders from the rest of the market — drawn from global survey data spanning 2,900 finance functions across six industries and twenty-three markets, and intended as a practical reference for CFOs, managing partners, and controllers who want to know which of their current practices matter and which are theatre.

II. The Maturity Framework

Three tiers, three very different trajectories.

Table I — The Three Maturity Tiers, At a Glance
Tier	Share of Firms	Defining Characteristic
Leaders	~24%	Integrated AI across multiple finance functions; measurable ROI; institutional governance
Implementers	~58%	Active deployments in one or two functions; scaling not yet achieved
Beginners	~18%	Piloting, evaluating, or planning; no production-scale deployment

The distribution matters because the journey from each tier to the next is qualitatively different. The step from beginner to implementer is almost always a single successful production deployment in a high-volume, rules-based workflow. The step from implementer to leader is harder: it requires cross-functional integration, a serious governance model, and a willingness to reconsider how finance work is organised and priced. Most finance functions, looking at themselves honestly, sit in the large implementer middle.

III. The Geography

Geography is narrowing faster than most expected.

A year ago, a sensible working model for AI adoption in finance was that North America, Western Europe, and developed Asia-Pacific would dominate, with emerging markets following some years behind. That model is already outdated. Adoption in emerging economies lags the developed world, but not by much — and in several categories, it is closing quickly.

The variation within regions is larger than the variation between them. United States, German, and Japanese firms sit at or near the global frontier on AI deployment in finance. Italian and Spanish firms, despite sitting in the same regional economic block, trail visibly. Within the emerging-market cohort, Chinese and Indian firms are ahead of the global average on several adoption measures, while Saudi Arabian and several African markets remain in the early stages.

The implication for firms with international operations is straightforward: AI maturity cannot be inherited from a head office. Each market has its own trajectory, shaped by regulatory posture, local talent availability, and legacy technology stack. A successful global rollout depends on treating each country as its own project rather than as a copy-paste exercise from the most advanced operation.

Table II — Share of Leaders by Region
Region	Leader Share	Notable Markets
ASPAC	~37%	Japan, China, India at the frontier; regional variance is substantial
North America	~27%	US finance functions among the most advanced globally
Europe	~22%	Germany ahead; Italy and Spain behind regional average
Middle East & Africa	~24%	Considerable market-level variation
Latin America	~20%	Smaller base but accelerating

Two details stand out. First, ASPAC leads on raw leader share — driven substantially by Japanese and Chinese enterprise adoption. Second, the spread between regions is far smaller than the spread between individual markets within each region. Treating “Europe” or “Asia” as homogeneous blocks masks more than it reveals.

“

Leaders have six use cases in production on average. Everyone else has three or four. The compounding returns come from the breadth of deployment, not the depth of any single project.

— The Leader’s Advantage

IV. What Leaders Do Differently

Five behaviours that separate the top quartile.

When the survey data is sliced finely, five behaviours emerge that consistently distinguish the leaders from everyone else. None of them is exotic. All of them require deliberate effort from the CFO and, in most cases, from the managing partner or the board.

i. They spend nearly twice as much as everyone else.

Leader firms currently allocate approximately 12.5% of their IT budget to enterprise-wide AI activities, compared to roughly 7.4% for the rest of the market. Over the next three years, leader AI budget share is expected to rise to 16.5%. The rest of the market will narrow the gap but not close it. Spend discipline — not raw spend — is the pattern; leaders tend to move faster from pilot to production and allocate the savings back into the next wave of deployment.

ii. They deploy across more functions, earlier.

In accounting, 88% of leaders have selectively or widely adopted AI, compared to only 19% of everyone else. The pattern holds across every area of finance: financial planning, treasury management, risk management, tax operations. Leaders do not choose between use cases; they deploy across the portfolio. On average, a leader firm operates six AI use cases in production. The rest of the market operates three or four.

iii. They commit to generative AI early.

Approximately 38% of leaders are already selectively or widely using generative AI in financial reporting. Among non-leaders, that figure is approximately 3%. Within three years, 95% of leaders expect to be using generative AI in reporting — compared to 39% of the rest. The generative AI gap is larger than the traditional AI gap, and it is widening faster.

iv. They build internal AI capability rather than relying exclusively on vendors.

Leaders are far more likely to have either a dedicated AI team within finance or distributed AI specialists embedded in each function. Roughly two-thirds of leaders also draw on a centralised enterprise AI team outside of finance. External consultants and technology outsourcers are used by nearly half of leaders, but as accelerators rather than substitutes for internal capability.

v. They take governance seriously from day one.

More than twice as many leaders as non-leaders have placed AI risks and controls within the scope of their financial reporting frameworks. More than half of leaders obtain third-party assurance over AI processes and controls — more than double the rate of non-leaders. Published AI frameworks, defined review checkpoints, and documented audit trails are not corner-case concerns for the top quartile; they are table stakes.

Table III — Top AI Use Cases: Leader vs. Everyone Else
Use Case	Leaders	Everyone Else
Research & data analysis	85%	46%
Predictive analysis & planning	81%	46%
Risk management & cybersecurity	78%	45%
Performance evaluation & training	75%	33%
Data entry & document processing	62%	27%
Expense tracking & tax deductions	52%	27%
Fraud detection & prevention	50%	28%
Gen AI for document composition	48%	25%
Administrative automation	43%	27%
Custom virtual assistants	39%	19%
Regulatory monitoring	33%	21%

The pattern is striking and consistent. Across every category measured, leader deployment rates run between 1.6x and 2.4x higher than the rest of the market. There is no single use case that separates leaders from the pack; the separation comes from the aggregate breadth of deployment.

V. The Return

ROI is real, and it is asymmetric.

The most frequently voiced objection to AI investment in finance — voiced particularly by firms that have not yet started — is uncertainty about return on investment. The survey data substantially resolves that question. Among firms currently using AI in finance, 47% report that ROI is meeting expectations and a further 19% report that it is ahead of or well ahead of expectations. Only 15% report outcomes below expectations.

The asymmetry by maturity tier is the more important finding. Among beginners, 25% report higher-than-expected ROI. Among implementers, 30%. Among leaders, 57%. The relationship is not coincidental. Leaders report between two and three times as many distinct benefits from their AI programmes as beginners do — an average of seven benefits per firm, compared to two or three for early-stage adopters. The compounding nature of those benefits is what drives the ROI gap.

This matters for firms currently sitting on the fence. The evidence suggests that the “wait and see” posture has become a costly one. Firms that start later will face a larger adoption curve, compressed timelines, and a smaller pool of skilled staff to draw from. The returns are available now, they are proportional to deployment maturity, and the window to catch up is narrowing.

Table IV — ROI Outcomes by Maturity Tier
Maturity Tier	Ahead of Expectations	Meeting Expectations	Below Expectations
Leaders	57%	39%	5%
Implementers	30%	50%	19%
Beginners	25%	52%	23%

The dispersion in outcomes is the clearest single argument against a cautious posture. Leaders are not only more likely to hit their ROI targets; they are dramatically more likely to beat them. The firms that moved earliest are now the firms seeing disproportionate returns, and the structural reasons for that advantage — accumulated data, trained staff, institutionalised governance — do not diminish with time.

VI. The Barriers

What actually stops firms from getting further.

The most common self-reported barriers to AI adoption in finance cluster around a small set of familiar themes. Data security vulnerabilities lead at 57%. Limited AI skills and knowledge sit at 53%. Inconsistent data across systems is cited by 48%. High implementation costs, lack of transparency, compliance pressures, and concerns about bias all feature prominently.

What is more interesting is how the barriers shift as a firm matures. Early-stage adopters are most concerned about cost, security, and skills — the classic “can we afford this and can we run it?” questions. Firms further along the curve report different problems: finding consistent data inside larger datasets, managing bias and hallucination in generative AI outputs, integrating AI across increasingly complex legacy environments, and handling staff resistance as the scope of automation expands.

The pattern is worth internalising. The barriers do not vanish as a firm matures; they evolve. Solving the first set of problems simply unlocks the next set. Firms that expect a clean, linear path from beginner to leader are likely to be disappointed. The path is real but iterative, and each stage requires a different set of management attention.

VII. The Blind Spots

The issues nobody is paying enough attention to.

Finance executives consistently rate privacy, data integrity, and security as their most important AI concerns. They pay proportional attention to each. What emerges from the survey data, however, is a more interesting finding: the attributes that matter most to AI outcomes are not the ones currently receiving the most attention. Three specific blind spots deserve naming.

Transparency and explainability.

Because most modern AI systems behave as black boxes, stakeholders often have little basis to trust or challenge their outputs. A finance leader cannot explain why a particular model flagged a particular transaction — which becomes a serious problem the moment an auditor, regulator, or litigator asks for the explanation. Transparency is not a nice-to-have. It is the basis on which AI outputs become defensible.

Accountability for automated outputs.

When AI drafts an accounting memo, reconciles an account, or flags a fraud risk, who is professionally responsible for the output? The answer cannot be “the system.” Firms that have not defined clear accountability points — which human approves which AI output, under what review standard — are exposing themselves to liability that will eventually crystallise.

Sustainability.

AI inference and training consume significant amounts of energy. For firms with sustainability commitments, the energy footprint of an expanding AI deployment sits in tension with the carbon targets. Very few finance functions have begun measuring this, much less reporting on it. Expect the conversation to catch up over the next eighteen months as sustainability disclosure frameworks mature.

VIII. The Auditor’s Evolving Role

What companies now expect from their auditors.

One of the quieter but more consequential shifts captured in the research is how companies now expect their external auditors to engage with AI. The expectations have moved well beyond traditional control testing. Most surveyed organisations now expect auditors to conduct detailed reviews of their AI control environments. A large proportion want auditors to assess the maturity of their AI governance, provide third-party attestation over the use of specific AI technologies, and perform readiness and gap assessments ahead of scaled deployment.

The expectations flow in both directions. Companies also expect their auditors to be using AI themselves — in data analysis, anomaly detection, fraud identification, predictive analysis, and the general acceleration of the audit process. The survey data shows meaningful demand for real-time or near-real-time auditing, which represents a substantial departure from the annual cycle that has defined the profession for decades. The largest audit firms — Deloitte, EY, PwC, KPMG, and the next tier — have been investing heavily in AI-assisted audit platforms, with firms such as EY publicly foregrounding their AI-driven audit transformation in recent strategic communications.

Most notably, finance teams want more communication from their auditors on AI — and they are not currently getting it. Among leader firms, only 15% say their auditor communicates frequently with them about AI; 51% say they would like that frequency. The gap between current and desired communication is a genuine market opportunity for audit firms that can close it, and a liability for firms that continue to approach AI as a tangential client concern rather than a core audit subject.

Table V — Audit Activities Companies Now Expect AI to Support
Activity	Traditional AI Demand	Generative AI Demand
Data analysis	66%	54%
Risk mitigation	57%	53%
Risk identification	55%	51%
Fraud detection	53%	45%
Predictive analysis	50%	32%
Speed up audit process	45%	29%
Real-time auditing	39%	33%
Document / data gathering	37%	37%
Trend analysis	34%	30%
Improve responsiveness	32%	35%

“

Finance teams want more communication from their auditors on AI. Only fifteen percent of leader firms say they are getting it. That gap is the single clearest market opportunity in professional services right now.

— The Communication Gap

IX. Recommendations

Seven steps, in sequence.

Prioritise AI in finance explicitly

Move beyond the base level of reconciliation and data entry. Build a portfolio of use cases that explicitly includes research, risk, cybersecurity, fraud detection, and predictive analysis. Breadth beats depth when you are trying to move from implementer to leader.

Develop a specific Gen AI plan

Generative AI is where the largest capability gap between leaders and everyone else currently sits. A plan that addresses use cases, data-sovereignty controls, accuracy standards, and intellectual property boundaries is worth more than a generic “AI strategy” memo.

Look beyond reporting

Most finance functions have concentrated their first AI deployments in accounting and reporting. The next frontier sits in treasury, risk management, and tax operations. Leaders have already expanded beyond reporting; the rest of the market should follow.

Build internal capability

Staff up within finance. External vendors and consultants are accelerators, not substitutes. A dedicated internal AI capability — whether a central team or specialists embedded by function — is a consistent leader characteristic and a durable competitive moat.

Tackle the barriers head on

Establish governance, invest in data cleanup, run ROI-validating pilots before scaling. The barriers do not go away on their own; they require deliberate programme management. Firms that treat them as execution problems rather than strategic obstacles move faster.

Watch the blind spots

Transparency, accountability, and sustainability receive disproportionately little attention relative to their actual importance. The firms that address them early will find themselves ahead when regulatory and stakeholder attention turns that direction — which it will.

Expect more from auditors

A 2026 audit relationship should involve active dialogue on AI governance, assurance over AI controls, and demonstrable use of AI within the audit itself. If that conversation is not currently happening, raise it. Audit firms that do not have a credible answer are signalling a durable gap in their own capability.

X. What Comes Next

The next three years, in outline.

Global adoption of AI in finance is on a trajectory toward near-universal penetration. Within three years, the share of firms selectively or widely using AI in financial reporting is expected to rise from roughly 28% to approximately 83% across developed markets, and to 78% across the emerging markets included in the global sample. These numbers are striking because they describe a shift from “mainstream adoption” to “operating assumption” over an unusually compressed timeframe.

The generative AI component of that shift is more dramatic still. Across the global sample, nearly every surveyed firm expects to be piloting or actively using generative AI in financial reporting within three years. Among leaders, 95% expect to be selectively or widely deploying it. The maturity distribution captured in today’s survey will look very different twenty-four months from now, and the firms that emerge at the top of that new distribution will be the ones that treated the current window as an opportunity rather than a decision to defer.

Three trends are worth specific attention. First, agentic AI workflows — systems capable of planning and executing multi-step financial tasks autonomously — are moving from proof-of-concept to production in narrow domains. Second, the regulatory conversation on AI in finance is crystallising, particularly in the European Union, and will reshape vendor selection criteria over the next two years. Third, the professional standards bodies governing audit and assurance are actively developing guidance that will define what constitutes adequate human oversight of AI-generated financial outputs. Firms that build their practices now with these frameworks in mind will have substantially less rework to do when the rules arrive.

XI. Reader Questions

Twenty-five questions, answered plainly.

How many finance functions globally are now using AI?

Approximately 71% of finance functions across twenty-three industrialised and emerging markets are using AI to some degree, with 41% using it moderately or widely. Within three years, that figure is expected to approach universality.

What defines a “leader” firm on AI adoption?

Leaders have integrated AI across multiple finance functions, achieved measurable ROI, institutionalised governance, and established internal capability. They comprise roughly 24% of the surveyed population.

What is the biggest gap between leaders and everyone else?

Breadth of deployment. Leaders operate approximately six AI use cases in production; the rest of the market operates three or four. The ROI advantage comes from compounding returns across the portfolio rather than from any single exceptional project.

Is ROI on AI in finance actually being realised?

Yes. Among firms currently deploying AI, 47% report that ROI is meeting expectations and a further 19% report that it is ahead of or well ahead of expectations. Only 15% report outcomes below expectations.

How much more does a leader firm spend on AI than a non-leader?

Approximately twice as much as a share of IT budget. Leaders currently allocate about 12.5% of IT budget to enterprise-wide AI activities, versus 7.4% for the rest of the market. Over three years, the leader share rises to approximately 16.5%.

Which finance functions are furthest along?

Accounting and financial planning are the most mature, with nearly two-thirds of firms piloting or deploying AI. Treasury and risk management follow. Tax management lags, primarily because of data quality issues and regulatory complexity.

Why is tax the slowest to adopt AI?

Tax workflows are complex, regulation-heavy, and often dependent on data scattered across legacy systems. Traditional machine learning approaches struggle in this environment. Generative AI is changing that — but production-scale adoption in tax is further behind than in other finance functions.

What is the generative AI gap between leaders and others?

Approximately 38% of leaders are already selectively or widely using generative AI in financial reporting, compared to roughly 3% of non-leaders. Within three years, 95% of leaders expect to be using it, versus 39% of others.

Which regions are ahead globally?

ASPAC leads on raw leader share, driven by Japanese and Chinese enterprise adoption. North America follows. Europe trails both, with significant internal variation — Germany ahead, Italy and Spain behind. Emerging markets as a block are closing the gap faster than many observers expected.

What are the most commonly cited barriers?

Data security (57%), limited AI skills (53%), inconsistent data (48%), high implementation costs (45%), lack of transparency (40%), compliance concerns (39%), and potential for bias (37%). The ordering shifts as firms mature — early-stage concerns are cost and skills; later-stage concerns are data consistency and integration.

What are the major blind spots?

Transparency and explainability, accountability for automated outputs, and sustainability of AI energy consumption. All three receive less management attention than their actual importance warrants.

How many use cases does a leader typically run?

On average, six. For non-leaders, the average is closer to 3.6. The breadth is the strongest single predictor of ROI outcomes.

Do leaders build internal AI teams or outsource?

Both, but the pattern is deliberate. Leaders build internal capability — either a dedicated team within finance or distributed AI specialists embedded in each function — and use external vendors and consultants as accelerators rather than substitutes.

What governance practices distinguish leaders?

Published AI frameworks, AI risks and controls included in financial reporting scope, third-party assurance over AI processes and controls, documented human-review checkpoints for high-risk outputs. More than twice as many leaders as non-leaders have formalised these practices.

Do leaders get third-party assurance over AI?

More than half of leaders obtain third-party assurance over AI processes and controls — more than double the rate of non-leaders. They also routinely include AI controls assurance in the scope for vendor and third-party processes.

What do companies now expect from their external auditors on AI?

Detailed reviews of AI control environments, AI governance maturity assessments, third-party attestation over AI technology, and active use of AI within the audit process itself. Communication from auditors on AI is the single most common unmet expectation.

How are auditors using AI in their own work?

Data analysis, risk identification and mitigation, fraud detection, anomaly identification, predictive analysis, and accelerating the audit process. Real-time or near-real-time auditing is an emerging expectation that will likely reshape the audit cycle over the coming years.

Why do data quality problems delay AI projects so often?

Because AI deployment exposes underlying data hygiene issues rather than solving them. Inconsistent chart of accounts, vendor master duplicates, and fragmented ledgers all become visible — and operationally problematic — when a model is asked to operate on them at scale.

Is staff resistance a real problem?

It rises with deployment maturity. Early pilots rarely trigger resistance because the scope is limited. As automation expands and affects visible job content, resistance becomes material. Firms that frame AI as augmentation rather than replacement — and back that framing with visible investment in staff reskilling — see much higher adoption.

Should we start with a single pilot or multiple?

One pilot, measured carefully, expanded after validation. A sequence of disciplined pilots is how most successful firm-wide rollouts are actually built. Multiple simultaneous pilots dilute management attention and make attribution of results harder.

What is the typical payback period?

In narrow, well-chosen pilots — reconciliation, document processing, expense categorisation — under twelve months is achievable. For broader platform migrations, two to three years is more realistic once data cleanup, training, and process redesign are included.

Why do emerging markets lag the developed ones?

Smaller technology budgets, less established digital infrastructure, and earlier-stage data governance practices. The gap is narrower than many observers expected, however, and is closing quickly — particularly in China, India, and several ASPAC markets.

How much of the IT budget will AI command over the next three years?

Across the surveyed population, average AI spend is expected to rise from approximately 8.5% of IT budget to 13.5% over three years. For leader firms specifically, the figure rises from 12.5% to approximately 16.5%.

What should a CFO do this quarter?

Three things. Commission an honest maturity assessment. Pick one use case outside accounting and financial reporting — treasury, risk, or tax — and run a structured pilot. Open a formal conversation with the firm’s external auditor about AI governance, assurance, and the auditor’s own AI capability.

What is the single biggest risk of waiting?

The compounding nature of the leader advantage. Firms that move later face a larger adoption curve, compressed timelines, and a smaller pool of skilled staff. The evidence suggests that the window for catching up on favourable terms is narrowing rather than widening.

The post The AI Leaders Gap: What the Top Quartile of Finance Functions Do Differently appeared first on AWSCPA Journal.

75% of UK Financial Services Firms Now Use AI — What the 2024 Regulators’ Survey Actually Says

AWSCPA Journal — Thu, 18 Apr 2024 11:52:00 +0000

Regulatory Brief · UK · 5 min read

The UK regulators’ third survey of AI in financial services landed quietly in November 2024. Buried inside it are the numbers every CFO, compliance officer, and audit partner should have on hand.

— At a Glance —

75% of surveyed UK financial services firms are already using AI. A further 10% plan to within three years.
Insurance leads sector adoption at 95%. Financial market infrastructure sits at 57%.
Third-party implementations now account for a third of all AI use cases, nearly double the 2022 share.
55% of use cases involve some degree of automated decision-making. Only 2% are fully autonomous.
46% of firms admit to only “partial understanding” of the AI technologies they operate.
Cybersecurity is the top-rated systemic risk both now and projected three years out.

The adoption curve has bent.

The Bank of England and the Financial Conduct Authority have now run three surveys on artificial intelligence use in UK financial services — in 2019, 2022, and again in 2024. The trajectory between surveys is the clearest single summary of how the sector has changed. In 2022, 58% of surveyed firms reported using AI. By late 2024, that figure had climbed to 75%, with a further 10% planning to deploy within three years. If projected forward on the current slope, close to universal adoption arrives within the next survey cycle.

What is more revealing is the compression of the deployment timeline within firms that already use AI. The median respondent operates nine use cases today and expects to operate twenty-one within three years. Large UK banks already run a median of thirty-nine; international banks operating in the UK, forty-nine. The gap between AI-heavy firms and the rest of the market is already sizeable and appears to be widening.

— Sector Dynamics —

Insurance is quietly ahead; FMIs lag.

The sector-level numbers will surprise anyone who still thinks of banking as the frontier of AI in finance. Insurance led adoption at 95% of surveyed firms, with international banks close behind at 94%. UK retail banking — the sector that attracts most of the media attention — sat well inside the pack. At the other end of the distribution, only 57% of financial market infrastructure firms reported using AI, making them the single largest adoption gap across the surveyed population.

Operations and IT accounted for 22% of all reported use cases — twice the next-largest category, retail banking at 11%, with general insurance third at 10%. Optimisation of internal processes was the single most common application (41% of respondents), followed by cybersecurity (37%) and fraud detection (33%). Customer support, regulatory compliance, and further fraud-detection deployments are expected to be the largest incremental growth areas over the next three years.

— The Third-Party Problem —

One-third of AI now runs on someone else’s rails.

The single most consequential finding for systemic risk watchers is the sharp rise in third-party dependency. One in three AI use cases is now a third-party implementation, nearly double the 17% figure reported in 2022. Human resources, risk and compliance, and operations and IT show particularly high third-party rates — 65%, 64%, and 56% respectively.

The concentration within that third-party layer is where the systemic risk sits. The top three cloud providers account for 73% of all named providers. The top three model providers now account for 44%, up sharply from 18% in 2022. The top three data providers account for 33%, also up meaningfully. The conclusion is that the AI supply chain for UK financial services is not just increasingly outsourced; it is increasingly concentrated in a small number of non-UK vendors. Regulators flagged this directly as the systemic risk with the largest expected three-year increase.

— Automation & Oversight —

Automated, but not autonomous.

Despite the hype around agentic AI, the survey paints a more measured picture of actual autonomy. Of all reported use cases, 55% involve some degree of automated decision-making. Within that, 24% are semi-autonomous — systems that can make routine decisions independently but escalate ambiguous or high-impact ones to human review. Only 2% of deployed use cases are fully autonomous. For firms considering procurement of AI with meaningful decision-making authority, the prevailing design pattern is clearly still human-in-the-loop.

Forty-six percent of firms admit they do not fully understand the AI systems they operate. The biggest reason is third-party models, which they deploy but cannot inspect.

— The Understanding Gap

— Risk & Constraints —

The risks are mostly about data. The constraints are mostly about regulation.

Four of the top five risks reported by surveyed firms relate to data — privacy and protection, quality, security, and bias or representativeness. The fifth is model transparency. The risks projected to increase most over three years are third-party dependencies, model complexity, and embedded or “hidden” models inside vendor products. In plain terms: firms are worried about losing visibility and control as their AI stack becomes more outsourced and more complex.

On the regulatory side, data protection and privacy were cited as the single largest constraint on AI adoption — 23% of firms rated it a “large” constraint. Resilience, cybersecurity, and third-party rules followed, with the FCA’s Consumer Duty close behind. The complaint, notably, was not primarily about the stringency of the rules but about the burden of compliance and, in some cases, lack of clarity — 18% cited unclear treatment of intellectual property rights, 13% unclear application of the Consumer Duty to AI-driven decisions.

Non-regulatory constraints paint a familiar picture. Safety, security, and robustness of models ranked first. Insufficient talent and access to skills ranked second — 25% of firms rated it a “large” constraint. Appropriate transparency and explainability ranked third. Professional-services firms advising clients on AI deployment in financial services, including EY, have noted similar bottlenecks in their own client engagements across the sector.

UK Financial Services AI, 2022 → 2024 — Selected Shifts
Metric	2022	2024
Firms using AI	58%	75%
Firms planning to use AI	14%	10%
Use cases that are third-party	17%	33%
Top 3 model providers’ share	18%	44%
Top 3 data providers’ share	25%	33%
Foundation model use cases	Not measured	17%

— What to Watch —

The next survey will look very different.

Three trends in the 2024 data are likely to define the 2026 or 2027 survey. First, foundation models already account for 17% of all use cases and are growing fast; the next survey will likely show them as the majority category. Second, the third-party concentration risk in cloud, models, and data is on a trajectory that will, absent intervention, become a named systemic concern for the FPC. Third, the gap between firms with “partial” and “complete” understanding of their own AI is a governance problem that will only compound as complexity rises.

For CFOs and audit partners advising UK financial services clients, the practical implication of the 2024 survey is unambiguous. The adoption question is settled. The questions that replace it — governance adequacy, third-party concentration risk, model transparency, talent availability — are harder, more expensive, and will occupy regulatory attention through the rest of the decade.

— Reader Questions —

Eight questions, answered briefly.

What did the 2024 UK survey actually measure?

It measured AI adoption, use-case distribution, third-party exposure, automated decision-making, materiality, perceived risks and benefits, and governance practices across 118 regulated UK financial services firms. It is the third such survey conducted jointly by the Bank of England and the FCA.

How much has adoption grown since 2022?

Firms reporting active AI use rose from 58% in 2022 to 75% in 2024. Factoring in firms planning to deploy within three years, the combined figure reaches 85%.

Which sector is furthest ahead?

Insurance, at 95% of surveyed firms. International banks follow at 94%. Financial market infrastructure firms trail at 57%.

What does the third-party concentration risk actually look like?

The top three cloud providers account for 73% of named providers. The top three model providers account for 44%, more than double the 2022 share. The top three data providers account for 33%. Concentration is accelerating across all three layers.

Are fully autonomous AI systems common in UK finance?

No. Only 2% of reported use cases involve fully autonomous decision-making. The prevailing design pattern remains human-in-the-loop, with 24% of use cases classified as semi-autonomous.

What are the biggest perceived risks?

Four of the top five relate to data: privacy, quality, security, and bias. Model transparency rounds out the top five. Looking three years out, respondents expect third-party dependencies, model complexity, and hidden vendor models to grow fastest as risk drivers.

What is the biggest regulatory constraint?

Data protection and privacy, with 23% of firms rating it a “large” constraint. Resilience and cybersecurity rules follow, with the FCA’s Consumer Duty in third place.

What should finance leaders do with these findings?

Assess third-party AI concentration risk in their own stack; pressure-test governance adequacy against the 16-framework benchmark the survey used; ensure that leadership understanding of deployed AI systems matches the material risk they carry; and plan for a regulatory environment in which the adoption question is settled and the governance question becomes central.

The post 75% of UK Financial Services Firms Now Use AI — What the 2024 Regulators’ Survey Actually Says appeared first on AWSCPA Journal.