There are two stories being told about AI and accounting. The first is the one the firms tell: AI is transforming the profession, improving audit quality, enabling better advisory, and making accountants more productive. The second is what is actually happening. They are not the same story.
The investment numbers sound impressive
PwC announced a billion-dollar investment in generative AI. RSM committed another billion. The Big Four collectively have poured billions into AI tools and platforms. Deloitte launched Zora AI. EY built EY.ai. PwC partnered with OpenAI. Press releases were issued. Chief AI Officers were hired. Innovation labs were opened.
And then you look at what actually changed for the client, and the answer is: the fees went up.
The median net hourly billing rate across accounting firms rose nearly 7% over two years, to $170 per hour. In the same period that firms were publicly celebrating AI-driven efficiency gains, they were charging more per hour, not less. The AI does the work faster. The firm bills the same hours. The partner pockets the difference.
The client paid for professionals. The client received a language model. The partner kept the margin.
The KPMG hypocrisy
In 2025, KPMG International demanded that its own auditor, Grant Thornton, pass along cost savings from AI tools. The fee fell 14%, from $416,000 to $357,000. KPMG argued, correctly, that if AI makes the work faster and cheaper, the client deserves the savings.
The irony was immediate. KPMG demands AI discounts from its vendor while offering none to its own clients. They know AI reduces the cost of the work. They insist on the savings when they are paying. They pocket the savings when they are billing.
PwC's own Chief AI Officer admitted the dynamic publicly. Clients, he said, "would hear us talking about using AI and say, 'We want our fair share of those efficiencies.'" PwC has cut prices for some services. How many, and by how much, they will not say.
The Deloitte disaster
Deloitte's Australian member firm charged the government $290,000 for a report on welfare policy. A university researcher discovered it was full of fabricated references, hallucinated citations, and a fabricated quote attributed to a federal court judgment. Deloitte had used GPT-4o to produce the report and had not adequately checked the output.
An Australian senator called it the kind of work a first-year university student would be expelled for. The UK's accounting regulator subsequently found that the six largest firms were not formally monitoring how AI affected the quality of their audits.
Deloitte's response was to issue a revised version with a footnote acknowledging that AI was used, and to state that "the substance of the independent review is retained." The substance was retained. The citations were fabricated. The client paid $290,000.
What the firms are actually doing with AI
Let me cut through the press releases. There are three things happening simultaneously, and the firms do not want you to see them as connected.
First, they are using AI internally to reduce the hours their juniors spend on reconstruction: data entry, transaction matching, document review. This is real and it works. AI can scan entire transaction populations instead of samples. It can match documents across systems. It can do in hours what a team of juniors does in weeks.
Second, they are not passing the savings to clients. The billing rates go up. The fees stay flat or increase. The margin between what the work costs to produce and what the client pays is widening. Every efficiency gain from AI flows to the partner's compensation, not to the client's invoice.
Third, they are cutting junior headcount. PwC UK has reduced partner numbers and halted tech apprenticeship programmes. KPMG UK achieved record partner pay through aggressive cost-cutting. Big Four firms slashed graduate jobs as AI takes over entry-level tasks. The pyramid is shrinking from the bottom.
Put those three together: AI does the juniors' work, the juniors get cut, the clients pay the same or more, and the partners earn more than ever. That is not transformation. That is extraction.
What AI in accounting should actually look like
There is another model. It starts with a different question: instead of asking "how can AI make the firm more profitable?", it asks "how can AI make accounting accessible to the people who need it?"
The architecture has three layers. The first is AI extraction: a language model reads documents, extracts data, identifies suppliers, amounts, dates. This is a commodity. Every model does it well. The second is a deterministic rules engine: tax law encoded as executable logic. Which VAT rate? Which box? Deductible or not? These are not probabilistic questions. They have correct answers defined by statute. You do not guess. You compute. The third is human judgment: a qualified accountant reviews only the exceptions that the system cannot resolve with confidence.
In this model, the AI does not make the firm more profitable. It makes the firm unnecessary. The reconstruction that justifies the firm's existence is automated. What remains is the human judgment that the training was designed for. And that judgment can be delivered at a fraction of the current cost, because the overhead that the firm requires to exist has been eliminated.
The question is not whether AI can do accounting. It is whether there is any justification for requiring humans to do the work that AI already does correctly.
The three failures to watch for
Not every AI accounting company will succeed. Two have already failed spectacularly, and their failures are instructive.
Botkeeper raised $89.5 million over eleven years. Its technology could clean up years of messy data in minutes and code 80%+ of transactions at 98% accuracy. Then it died in weeks. It had sold to mid-size accounting firms, and when PE consolidation hit those firms, Botkeeper's customers disappeared. The lesson: selling tools to firms is building on rented land.
Bench was the largest online bookkeeping service in North America. It paired AI with human bookkeepers. The AI produced too many errors. The humans had to clean up after it. The cost savings never materialised. Bench shut down in December 2024. The lesson: AI that creates work for the human instead of eliminating it makes the problem worse.
The companies that will succeed are the ones that do not sell tools to accountants and do not depend on them for distribution. They will do the work, directly, for the people who need it. The accountant becomes a reviewer of exceptions, not a processor of everything. The client gets professional-grade compliance at a price they can afford. The firm, as a structure, dissolves.
What this means for you
If you are an accountant: the AI is not your enemy. The firm structure is. AI eliminates the reconstruction work that consumes your time. What remains is advisory, judgment, and client relationships. That is the career you trained for. The question is whether you will do that work inside a firm that extracts your surplus, or independently, at a rate that reflects the value of your judgment.
If you are a freelancer or small business owner: the accounting you are paying for is about to get dramatically cheaper and dramatically better. The current system charges you for a human to type your receipts into a spreadsheet. The new system automates the typing and charges you for the human judgment on the 5% of transactions that genuinely need it. The difference in cost is 70-90%.
If you are an investor: the professional services industry generates over $5 trillion in annual revenue. Most of it pays for reconstruction. AI is automating reconstruction. The firms that reorganise around this reality will capture enormous value. The firms that resist will be undercut by those that don't.
Michael Cutajar, CPA — Founder of Accora.