Picture a finance team at month end. A new AI tool drafts 300 journal entries before lunch. Last year the same work took three people four days. The CFO is pleased.
Then the controller asks a simple question. Who checks them?
Every entry looks right. The accounts exist. The debits equal the credits. The narration reads well. But “looks right” and “is right” are different things, and the only way to tell them apart is to look at each entry against its source.
So the team has not saved four days. It has traded four days of typing for an unknown number of days of reviewing. If the review is done properly, much of the saving disappears. If it is not done properly, the errors go into the books.
That scene is an illustration, not a case study. But the trade it describes is real, it is measurable, and it sits underneath almost every AI project now being planned in finance and operations. This article is about that trade.
Producing things. Drafts, entries, emails, reports, reconciliations, code.
The fall in cost has been steep. Epoch AI, a research group that tracks this, estimated in September 2026 that the price of a fixed level of AI performance has dropped by about 47% every quarter since 2023. That is roughly 13 times cheaper each year. On that estimate, an AI task that cost 100 dollars in 2023 costs about five cents today.
The estimate has limits, and its authors say so. It is based on benchmark tests, not real office work, and the figures should not be read as exact. But the direction is not in dispute. Producing a first draft of almost anything is now close to free.
This is the part of the story everyone has heard. It is true. It is also only half of the story.
Checking.
A person still has to confirm that the output is correct before the business acts on it. That step costs what it always cost, because it still runs at the speed of a careful human. And there are four reasons why checking AI work is harder than checking human work, not easier.
When a tired clerk makes an error, it often shows. A field is blank. A total does not add up. The format is off. Experienced reviewers learn to spot these signs quickly.
An AI error has no such signs. The wrong answer arrives in the same clean format as the right one, with the same confident wording. A wrong ledger code looks exactly like a correct ledger code. The reviewer cannot skim. They have to go back to the source.
Vendors usually quote accuracy per field: the share of individual values the AI reads correctly. What a business needs is accuracy per document: the share of documents where every value is correct. You cannot post an invoice that is 90% right.
The gap between those two numbers is large. In a 2026 study of AI reading bank cheques, a small tuned model scored 98.5% at field level. It got every field right on only 88.9% of documents. A large general model scored 90.9% at field level and got the whole document right just 45.4% of the time.
Put plainly: a tool sold as “98% accurate” can still hand you one wrong document in every nine. And you do not know which one.
Real work is rarely one step. An AI agent handling an invoice may read it, match it to an order, pick a tax code, choose an account and draft the entry. Each step can fail, and the failures compound.
The arithmetic is simple. If each step is right 95% of the time, a ten-step task finishes correctly about 60% of the time. A twenty-step task finishes correctly about 36% of the time.
This is why a demo can look excellent and a live process can disappoint. The demo shows one step. The process runs twenty.
The obvious fix would be to let the AI flag its own doubtful answers, so people review only those. Today that does not work well.
AI models are often most confident when they are wrong. A 2026 study of financial documents tested this directly. Using the model’s own confidence, only between 0.1% and 7% of fields could be cleared without human review at an acceptable error level. A separate checking layer, built outside the model, cleared between 39% and 71%.
The lesson is clear. The model can do the reading. It cannot be trusted to grade itself. Something else has to do the grading, and building that something is real work that someone must pay for.
Producing got 13 times cheaper each year. Checking did not get cheaper at all. Every other fact in this article follows from that gap.
This is not only theory. Four pieces of evidence point the same way.
Experts were slower with AI, and did not notice. In 2025 the research group METR ran a controlled trial with 16 experienced software developers working on their own projects. With AI tools allowed, tasks took 19% longer. The developers had expected to be 24% faster. After the trial they still believed they had been 20% faster. The time went into reviewing and correcting what the AI produced. The study is small and covers software, not finance. But it measures the exact trade this article describes, and it shows that people do not feel the cost of checking even while they are paying it.
Unchecked AI work lands on colleagues. Researchers at Stanford and BetterUp Labs surveyed 1,150 full-time employees in the United States in 2025. Forty percent said they had received low-quality AI-generated work from a colleague in the past month. Each case took nearly two hours to sort out. The researchers put the cost at about 186 dollars per employee per month. This is survey data and should be read with care, but the pattern is familiar: when checking is skipped at the source, it is paid for downstream.
Regulated firms keep a person in the loop. The Bank of England and the Financial Conduct Authority surveyed UK financial firms in October 2026. Eighty percent already had AI agents in use. Only 2% allowed them broad autonomy. More than 90% kept human approval as a control. These firms have the budgets and the talent. They have looked at the problem closely and decided that checking cannot yet be removed.
Many projects are not surviving. Gartner forecast in 2025 that more than 40% of agentic AI projects would be cancelled by the end of 2027, citing rising costs and unclear business value. A business case that counts the saving from producing and leaves out the cost of checking is exactly the kind that fails this way.
A note on evidence
You will have seen confident figures for what AI saves in finance: close cycles cut in half, 85% of invoices processed without a human. Almost all of these come from the companies selling the software. We could not find one independently verified study of AI agent results inside an ERP system. Part 4 of this series looks at that problem in detail.
For businesses in the Gulf, the Levant and South Asia, there is an added reason to care. In several of these markets the first reviewer of your documents is no longer your own team. It is the tax authority’s computer.
In these systems an invoice is checked by a machine, against fixed rules, close to the moment it is issued. That changes the cost of an error. A wrong tax code or a wrong registration number is not found weeks later in an internal review. It is rejected, or recorded, by the authority straight away.
Now add AI that produces invoices and entries at ten times the old volume. If your own checking does not scale with it, the errors still get found. They just get found by the regulator instead of by you.
There is a second, simpler reason. Finance teams in many mid-sized firms in the region are lean. Three or four people often run what a larger company gives to fifteen. A lean team has no spare capacity to absorb a review queue that suddenly grows tenfold.
When the review queue grows faster than the team, something has to give. What usually gives is the quality of the review.
It happens quietly. A reviewer facing 40 items a day reads each one. A reviewer facing 400 starts to skim. Soon they are clicking “approve” on trust. The system still records a human approval on every entry. The audit trail looks perfect. But nobody is checking any more.
This is the most dangerous state a company can be in. It has the cost of a human control and the protection of none. It also believes it is safe.
There is one simple number that shows whether your review is still alive: the rejection rate. If reviewers reject or correct a steady share of what the AI proposes, they are reading. If the rejection rate drifts toward zero while volume climbs, the approval has become a click. Either the AI has become perfect, which is unlikely, or the humans have stopped looking.
None of this is an argument against using AI. It is an argument for planning around the real constraint.
Before approving any AI project, ask two questions. How will the output be checked? What will that cost in people and hours? If the proposal has no answer, the saving it claims is not yet real.
Many checks need no judgment. Does the entry balance? Is the amount within the approved limit? Does the tax number have the right format? Is the period open? Ordinary software does these perfectly, instantly and at no cost per item. Every check you move from a person to a rule shrinks the queue. Keep people for the questions that need judgment.
Checking is slow when the reviewer has to hunt for the source. It is fast when the source is attached. An AI-drafted entry should arrive with the document it came from, and with the exact line highlighted. A reviewer can confirm a highlighted line in seconds. This single design choice does more for review speed than a better AI model.
Because errors multiply across steps, short tasks are far more reliable than long ones. Give the AI one clear step, check the result, then move to the next. Let ordinary software, not the AI, decide the order of the steps.
The goal is not for people to review everything. It is for people to review the right things. Clear the routine items with rules. Send the unusual ones to a person, with the evidence attached. Then track the rejection rate every month, so you know the review is still real.
The bottleneck in business has always been somewhere. For decades it was in producing the work: keying the invoices, preparing the entries, building the report. AI has removed that bottleneck faster than almost anyone expected.
But a business does not run on work produced. It runs on work it can trust. The limit is now how fast you can become sure that something is right. The companies that do well from AI will not be the ones that generate the most. They will be the ones that make trust cheap.
That raises a question about the systems underneath. If checking is now the scarce resource, the most valuable software in the company is the software that checks. For most businesses that is the ERP: the ledger that refuses an entry that does not balance, and the rules that stop a payment above its limit.
Which leads to an uncomfortable thought. For two years the common prediction has been that AI will replace the ERP. The evidence points the other way.
Next in the series · Part 2
If AI does the data entry, what is left of SAP, Sage or Tally? More than most people expect. Part 2 looks at which parts of enterprise software are about to disappear, which parts become more valuable, and why the dull old ledger may turn out to be the most important piece of the AI stack.
Still to come. Part 3: how a supplier’s invoice can give orders to your AI. Part 4: why every AI number you have been shown is unaudited, and what to ask for instead.