AI is coming for the accountancy firm, not just the accountant
In a test published on 1 October, AI did month-end accounting work more accurately and far more cheaply than qualified accountants. For small practices, the risk may be less about jobs and more about losing clients to firms run with AI. An OpenAI-backed group already owns nearly 30 US accounting firms and is building AI into their work.
Good Transformer is Patrick Hussey's AI practice. Want help with this in your business? Get in touch.
On 1 October, the AI data company Mercor published the results of a test in which AI did month-end accounting work more accurately, much faster and far more cheaply than 12 qualified accountants. Eighteen months ago, the accountants would have won.
Most of the reaction will be about jobs. The bigger threat may be to the firms that employ accountants.
OpenAI has a stake in a US group that now owns nearly 30 accounting firms. OpenAI says it is placing its own engineers inside the group's companies. If groups like this start running accounting firms in the UK, they could take clients from small practices.
This piece is about accountancy, but the argument applies to any firm whose work can mostly be checked against rules and records.
A small practice can respond in three ways now. It can hand routine back-office work to AI agents of its own, so it can match the speed and price of rivals that use AI. It can protect the three things a new rival does not have: its notes on each client, its relationships with clients and its own records. And it can run some of its AI on its own computers, so the most sensitive client data stays inside the practice.
What the Mercor test found
Mercor hired 12 accountants in the US, all licensed, with an average of about five and a half years' experience. Mercor calls them junior accountants. Each was given four simplified month-end tasks taken from a benchmark Mercor built with accountants. The tasks were based on real month-end work, such as reconciling a hotel's booking system with its books, comparing a charity's spending with its budget, and rolling lease schedules forward a month.
Each accountant had three hours per task and was paid for the full three hours, with a bonus for every item they got right. On average, the accountants got about 37% of the marked items right. The AI, Anthropic's Claude Opus 5 working alone, got all of them right and finished much faster. Mercor says that, per correct item, the AI cost less than a tenth as much as the accountants (Mercor).
Mercor stresses how recent this is: "just a year and a half ago, the best AI models still fell short of the average accountant's score of around 37%. Today, models ace those same tasks."
In the full paper, Mercor also had the accountants do two of the tasks with an AI assistant. They were slower, and slightly less accurate, than the AI working alone. In one session an accountant overrode the AI's correct answer (Mercor research paper).
For a firm, that result points to handing whole pieces of routine work to AI, with people reviewing the finished work rather than redoing it.
What the Mercor test does not show
The test has clear limits, and Mercor states most of them itself.
- The tasks were shorter than real month-end work, but harder. They were packed with planted, hard-to-spot errors, such as revenue coded to the wrong account. The task authors expected a junior to score only about 30%.
- The accountants had less to go on than usual. They had no colleagues to consult, no history with the client and no way to ask the client anything. Mercor says this made the tasks much harder. One of the accountants said that in real work they can usually ask the client questions as they go.
- AI has not yet fully solved most of Mercor's tasks. On Mercor's full benchmark of 160 tasks, 58% have never been fully solved by any AI model (Mercor leaderboard).
- The test used US rules. The accountants were US-qualified and one task used US lease accounting.
Mercor's business is hiring experts to create training data for AI companies, so it has an interest in the result being noticed. Even so, its own conclusion is measured. It expects accountants' work to shift towards "handling client relationships, asking the right questions of colleagues, or operating in the presence of uncertainty or lack of clear rules".
Much of the work in a typical practice is neither client conversation nor judgement. That work is the most likely to be done by AI soon.
Why AI is likely to reach accountancy before most professions
AI improves fastest on work where an answer can be checked. A system can try a task many times, check each attempt, and learn from the ones that pass.
The AI researcher Jason Wei put it as a rule: "The ease of training AI to solve a task is proportional to how verifiable the task is" (Jason Wei). Andrej Karpathy, a co-founder of OpenAI, made the same point about jobs in November 2025: "The more a task/job is verifiable, the more amenable it is to automation" (Andrej Karpathy).
We find it useful to sort work into three kinds, depending on how fully an answer can be checked.
Bounded work is work where every answer can be checked completely. Mathematics is the clearest case. In September, OpenAI's agents solved a famous maths problem, and a computer checked a formal version of the proof line by line. We covered it in our piece on the Navier-Stokes result.
Software is close behind, because code can be tested. Anthropic says that by May 2026 more than 80% of the code merged into its own codebase was written by Claude, its AI, up from low single figures in early 2025 (Anthropic).
Semi-bounded work is work where most answers can be checked against records and rules, and judgement decides the rest. Accounts either reconcile or they do not, but deciding how to treat an odd item takes judgement. A VAT return either follows the rules or it does not. Much of accountancy, tax and law is this kind of work.
Unbounded work has no single right answer to check. A novel, a brand campaign or a company strategy is judged by taste and by results that arrive years later. Karpathy lists creative and strategic work among the tasks that "lag by comparison".
AI is likely to take on semi-bounded work next. So after software engineering, accountancy is likely to be one of the first professions where AI does a large share of the work.
OpenAI, which has a stake in a group buying accounting firms, takes the same view. When it took that stake in Thrive Holdings in December 2025, it said "the initial focus is accounting and IT services because these functions run high-volume, rules-driven, workflow-heavy processes" (OpenAI, via the Internet Archive).
How AI could go from doing tasks to running whole firms
What follows is partly our own view of how AI could change accounting firms, so we say which parts are fact.
In 2019 the AI researcher Rich Sutton wrote an essay called The Bitter Lesson. Looking back over 70 years of AI research, he concluded that general methods which make use of more computing power "are ultimately the most effective, and by a large margin" (Rich Sutton).
In plain terms, AI that learns from huge amounts of data, using large amounts of computing power, has repeatedly beaten AI built from experts' hand-written rules. The lesson is bitter because experts keep expecting their own knowledge to win.
So far, the lesson has held for single tasks, such as playing chess and, now, reconciling accounts. We think it may now hold for bigger units of work, first whole jobs and then whole organisations.
Whole jobs are becoming training data. A Mercor job advert from July asked for accountants holding a CPA, ACCA or similar qualification to help by "translating everyday accounting workflows, judgments, and decision-making into structured, high-quality training data" (Mercor). In October 2025, Bloomberg reported that OpenAI was paying more than 100 former investment bankers $150 an hour to build financial models for training its AI (Fortune, reporting Bloomberg).
Agents are being connected to whole organisations. OpenAI's Frontier platform, launched in February, connects AI agents to a company's data, customer records and internal systems so the agents have "shared business context". OpenAI says its agents "build memories, turning past interactions into useful context that improves performance over time".
It adds that as companies use them, "we also learn how the models themselves need to evolve to be more useful for your work" (OpenAI, via the Internet Archive). In effect, each company that runs agents this way builds up a record of how it works. Its agents draw on that record, and OpenAI learns from how they are used.
Agents can organise themselves into swarms. In September, about 10,000 OpenAI agents worked together on the maths problem by sending each other 2.7 million messages (our Navier-Stokes piece has the details). They were given simple tools and left to work out how to coordinate.
Agents have also been set up to run a small business. In Anthropic's office shop experiment, an AI "chief executive" agent supervised an AI shopkeeper, and weeks with a negative profit margin "were largely eliminated". Anthropic is cautious about the result: "the gap between 'capable' and 'completely robust' remains wide" (Anthropic).
Agents can already coordinate with each other and connect to a company's records. They may soon be trained on whole jobs too. We think it plausible, though not proven, that agents like these could start to run a firm's routine operations. Fewer people than today would set the rules and sign off the work.
That would be the bitter lesson applied to a whole firm. The way the firm runs would be learned from data, rather than designed by the people in it.
An accountancy practice looks unusually suited to this. Its processes follow written rules and its filing dates are fixed. Its client work can mostly be checked, and so can the running of the practice itself, from deadlines met to invoices chased. In the terms we used above, both are semi-bounded.
An OpenAI-backed group is already buying US accounting firms and building AI into them
This part is fact. Thrive Holdings owns Current, a group formerly called Crete Professionals Alliance. In June, Current had "more than 2,000 employees at nearly 30 independent accounting firms across the U.S." and more than $500 million in annual revenue. Thrive Holdings and OpenAI "have embedded a team of AI engineers and researchers to build tailored AI products in partnership with accountants".
In a tax pilot across 7,000 returns, Current reported an average 31% saving in preparation time (CPA Practice Advisor).
OpenAI calls the arrangement "a repeatable model that can expand across other industries". Joshua Kushner, the chief executive and founder of Thrive Capital and Thrive Holdings, explains why owning the firms matters. "The businesses we acquire represent the right reward systems for this evolution, bringing together industry expertise and real-world data that can help improve models on specific tasks and capabilities" (OpenAI, via the Internet Archive).
Put simply, the firms are not only customers for AI. Their day-to-day work is data that makes the AI better at accounting.
We have found no UK group doing this yet. Current still matters to UK practices.
It is one group, of about 30 firms, and its work is still done mostly by people. It is building software and working methods that let those people do the same work in less time. Even if Current never buys a UK practice, the software and methods it develops can be sold or copied here.
What this could mean for the profession
ICAEW reports that the Big Four "have reduced recruitment for junior roles" in the last two years. It adds that offshoring and cost cutting make it unclear how much of that is down to AI (ICAEW).
In ICAEW's 2026 survey of mid-tier firms, 68% agreed AI "will reduce the need for early career accountants". The same survey found 83% agreed "demand for accountants will continue, but for a redefined profession" (ICAEW).
Clients are changing where they get advice. Ravical, which sells AI software to accountants, commissioned research among 500 UK small and medium businesses. Of those, 70% said they turn to AI first for financial and tax advice, "without checking with their accountant" (Ravical). A client who already asks AI first has less reason to stay with a practice that is slower and dearer than a rival using AI.
We think firms run mostly by AI could be competing for small practices' clients within a few years.
That could mean fewer, bigger firms in the profession. It could mean job losses, starting with trainees, whose routine work AI can now do. We wrote about that in our piece on AI taking the work that trains juniors. And it could mean firms that are partly automated competing on price with practices that are not.
Larger firms may automate quickly. Small practices could also lose clients to groups that own firms and work closely with the companies that make AI models, as Current does with OpenAI.
AI is not just coming for accountants. It is coming for the accountancy firm.
A small practice can prepare for this now, at little cost.
What a small practice can do now
Three things matter most.
Put agents to work in your own practice. AI agents can now sign in to your systems, work through a job and ask a person before doing anything sensitive. A practice that uses them on its own routine work can keep its turnaround times and fees competitive.
Our free guide, Your first business agent, compares the agents available today and explains how to choose one. Our pieces on picking the first job to hand over and writing an agent's rules on one page go further.
Protect what a new rival does not have. A rival using AI can match your speed and your prices. It does not have your notes on each client, your relationships with them or your records of years of work for them. These three things are what set your practice apart, so they need looking after on purpose.
Your client notes. In Mercor's test, the accountants had no history with the client and could not ask questions. Your practice has that history. Most of it sits in people's heads, and it leaves when they do.
The bitter lesson says AI does better learning from data than following rules experts write for it. Facts about your clients are data, not rules, and only your practice has them. Once they are written down, your own agents can use them on every job for the client concerned.
We explain how to set this up in our piece on teaching AI your business. Notes about individual people are personal data, so keep them to what your engagement letter and data protection law allow. Our guide to GDPR and AI covers the basics.
Your relationships. The accountants in Mercor's test had never met the client. Your clients know who looks after them, and a new rival has to build that from nothing. In the same Ravical research, 94% of the businesses surveyed said they had lost knowledge or continuity when a key contact at their accountancy firm left or changed.
So make each relationship belong to the practice as well as to one person. Name the partner who looks after each client, and make sure a second person knows the client too. Spend some of the time agents save on meeting clients.
Your records. Your practice holds years of records for each client: their accounts, the adjustments you made, the questions they asked and how you answered. Kushner's point above, that a firm's real-world data can help improve AI, applies to your practice too. Organised well, your records can make your own agents better at your clients' work.
Keep those records in systems you control and can export from, and read what each software provider's terms allow it to do with them. If you specialise, anonymised figures across similar clients let you tell each client how they compare. A new rival cannot do that until it has clients of its own.
Consider running some AI on your own computers. Some AI models can be downloaded and run on a practice's own computers or server. They are often called open-source or open-weight models. Client data sent to a model run this way stays on the practice's own equipment.
This matters more now that OpenAI, one of the largest AI companies, also has a stake in a group of accounting firms. A paid business account from a major provider normally comes with terms saying your data is not used to train its models, and that is a reasonable safeguard. Running a model on your own computers means you do not have to rely on those terms. Your clients' data cannot help train AI that a rival firm will use.
There is a trade-off. Models small enough to run on a practice's own computers are usually less capable than the largest models offered online, and someone has to set them up and keep them running. They suit well-defined jobs that repeat and involve sensitive data, such as first-pass bank coding. Many practices will use both: an online service for general work, and a model on their own computers for the most sensitive client data.
A five-step plan for a small practice
Copy this, fill it in over the next 90 days, and review it with your partners.
A 90-DAY PLAN FOR A SMALL PRACTICE
Step 1. Sort your services by how easily the answer can be checked (week 1)
Put every service you sell in one of three columns.
Checkable: the answer can be checked against records and rules
(bank reconciliations, VAT returns, payroll, month-end journals)
Partly checkable: rules apply, but judgement decides the answer
(tax planning, management-accounts commentary, year-end adjustments)
Judgement: the client pays for advice and trust
(selling a business, raising finance, a difficult tax enquiry)
In the terms above, the first two columns are semi-bounded work
and the third is unbounded. None of a practice's services is fully bounded.
Expect prices for the checkable services to fall first.
Step 2. Put one agent to work in your own back office (weeks 2 to 6)
Pick one recurring job: chasing clients for missing records,
first-pass bank coding, or the month-end checklist.
Write down what the agent may see, what it may do alone,
and what it must bring to a person.
Name one person who checks its work every week.
For jobs on sensitive client data, consider a model
that runs on your own computers.
Step 3. Write down what you know about each client (weeks 2 to 8)
One page per client, kept with their file:
- the partner who looks after them, and a second person who knows them
- who decides, and how they like to hear from you
- the history behind their numbers (the one-off sale, the dispute, the seasonal dip)
- their plans and worries for the next two years
- what went wrong before, and what they valued
If you specialise (dental practices, farms, charities), also keep anonymised
figures across those clients, so you can tell each one how they compare.
Record only what your engagement letter and data protection law allow.
Check you can export all your client records from every system you use.
Step 4. Change what your juniors do (from week 6)
Move them from preparing the numbers to checking the agent's work
and talking to clients. Each month, have each junior join a partner
in at least one client meeting.
Step 5. Re-price the checkable work, and sell more advice (by week 12)
Decide which checkable services become fixed-price, faster or bundled.
Write one paragraph for clients on what your fee now pays for.
Add one advisory service that uses what you wrote down in step 3.
Step 5 needs the most thought. Our piece on what to charge when AI cuts the hours works through it.
What this means beyond accountancy
Until now, most fears about AI and work have been about individual jobs: will AI do your job, or your trainee's? That question still matters. But if agents can learn whole jobs and coordinate across a whole firm, the bigger risk for a small business owner is losing the business, not one role in it.
Accountancy is likely to be among the first professions to face this, because so much of its work can be checked and because an OpenAI-backed group is already buying firms. Insurance claims and much legal work rest on rules and records in the same way.
The response is the same in every case. Use AI agents in your own business before a rival built on them arrives. Protect what only you have: your client notes, your relationships and your records. Start with step 1 this week. It takes about an hour.
Common questions
Does the Mercor test mean AI can replace accountants?
The test alone does not show that. It used four simplified tasks, packed with planted errors, done by accountants with no colleagues or client history. On Mercor's full benchmark of 160 tasks, most are still not fully solved by any AI model.
What the test does show is that routine, checkable accounting work can now be done by AI quickly and cheaply. Eighteen months ago, the best AI models scored below the accountants on the same tasks.
Is anyone buying UK accounting practices to run them with AI?
We have found no UK group doing this yet. The main example we know of is in the US: Current, owned by Thrive Holdings, in which OpenAI has a stake. Large firms with UK practices already use AI agents in their own work. EY has rolled out agents across its global audit business for defined tasks, and each auditor remains responsible for reviewing the agents' work (ICAEW).
Is it safe to put client information into an AI agent?
It can be, on a paid business account whose terms say your data is not used to train the provider's models. For the most sensitive work, a model running on your own computers keeps client data inside the practice. Stay within your engagement letter and data protection law. Notes about individual people are personal data. Our guide to GDPR and AI explains what the law asks of your firm, and our piece on AI for accountants covers client data in practice.
Which AI agent should a small practice start with?
That depends on the software you already use and who will set it up. Our free guide, Your first business agent, compares the agents available today and gives four questions to narrow the choice.
Which other professions are most exposed to AI?
Any profession whose work can mostly be checked against rules and records is exposed in the same way. That includes insurance claims handling and much legal work. Work judged by taste or by long-term results, such as strategy or creative work, is likely to change more slowly.
If you want help working out which parts of your practice to hand to AI first, and how to keep your clients while you do, book a conversation.
This article is general information, not legal advice.
Work with Patrick
Want to use AI like this in your own work?
AI Coaching for Leaders is one-to-one help built around your real tasks. Email, call or book a free 30-minute call.