Most finance teams now use AI somewhere in their work. Far fewer have worked out what it costs to check what the AI produces. A survey published this month puts a number on that cost, and it is larger than most people would guess.
What the survey found
The 2026 CFO Sentiments Survey, run by Global Surveyz Research for the finance software company Datarails and published on 6 October 2026, asked 270 CFOs and finance leaders how AI is changing their departments. Its headline finding: finance teams spend 26% of their working week verifying or correcting output from AI tools. 96% of respondents spend at least 10% of their work time on these checks, and 8% spend more than half.
The reasons for the checking are just as telling. In the full report, 75% cite the "inability to audit or explain how AI reached its output" as something that stops them trusting it, 71% cite accuracy and hallucination, and 51% say they could not explain or trace how AI reached a number when asked. The most common frustration, named by 65% in the survey summary, is AI giving "confident answers based on the wrong data".
Trust varies a great deal by task. As CFO.com reported, 23% of respondents trust AI without human review for revenue forecasting, 13% for variance analysis, 5% for board-ready financial reports and 4% for month-end close tasks. The report puts drafting emails, by contrast, at 89%.
Two caveats before going further. Respondents were US organisations with more than 1,000 employees and at least $100 million in revenue, so this is not a survey of Australian Shopify brands. And it was commissioned by a vendor that sells finance software. Even so, the pattern is worth taking seriously, because it matches what anyone who has reviewed someone else's reconciliation already knows.
Why checking costs so much
Checking is expensive when the output arrives without its working. If an AI tool tells you a payout is "reconciled" and gives you a journal, the only way to be sure is to rebuild the reconciliation yourself. At that point the tool has saved you very little, and you have paid for the work twice.
The survey hints at the same thing from another angle. In the report, 76% of teams whose data sits in disconnected systems spend more than a quarter of their time verifying, against 31% of teams with mostly centralised data. When the figures behind an answer are scattered, every check becomes a hunt.
So the useful question is not "how accurate is the AI?" but "how quickly can a competent person confirm that this particular output is right?" A tool that is right 99% of the time but has to be fully redone to find the 1% costs more review time than one that is right 95% of the time and points at exactly where it is unsure.
A bar your auditor already uses
Auditors have long faced this problem with reports produced by a client's own systems. Australian auditing standard ASA 500 Audit Evidence, paragraph 9, says that when using information produced by the entity, the auditor must evaluate whether it is sufficiently reliable, including, as necessary, by obtaining evidence about its accuracy and completeness and evaluating whether it is sufficiently precise and detailed for their purposes.
Those three tests make a practical standard for any AI output you are asked to approve:
- Complete. Does it account for every transaction in the period, with nothing silently dropped?
- Accurate. Do the amounts agree to an independent source, such as the bank deposit or the processor's own report?
- Precise and detailed enough. Can you see each component and trace it, or only a summary you have to take on faith?
An output that passes all three can be reviewed in minutes. One that fails any of them has to be rebuilt.
What a checkable reconciliation looks like
Take a single Shopify Payments payout that lands in the bank as $12,498.35. A tool that only reports "matched" leaves you to work out why. A checkable version shows the decomposition:
| Component | Amount |
|---|---|
| Gross sales | $14,236.90 |
| Refunds | −$1,102.50 |
| Processing fees | −$251.05 |
| Chargeback | −$385.00 |
| Net payout | $12,498.35 |
- Gross sales
- Amount
- $14,236.90
- Refunds
- Amount
- −$1,102.50
- Processing fees
- Amount
- −$251.05
- Chargeback
- Amount
- −$385.00
- Net payout
- Amount
- $12,498.35
The reviewer's checks are now short and mechanical:
- Does it tie? $14,236.90 less $1,102.50, $251.05 and $385.00 is $12,498.35, which agrees to the bank line to the cent.
- Is it complete? The count of transactions in the decomposition matches the count in the processor's payout report.
- Can each line be traced? Every amount links back to the processor transactions that make it up, so a sample can be checked in seconds.
- Is the treatment explained? The chargeback is booked as a dispute, not netted into refunds, and the entry says which rule put it there.
- Are exceptions visible? Anything the tool could not resolve is listed with a reason, rather than parked in a suspense account for someone to find later.
None of these checks requires trusting the AI. They require the AI's output to carry its evidence. That is the difference between a quarter of the week spent reviewing and a quarter of the week spent redoing.
It also explains why the trust figure for month-end close is so low. Drafting an email is easy to judge by reading it. A journal entry is not: you cannot tell by looking whether $251.05 of fees is right. Without the arithmetic and the source transactions beside it, a close entry from an AI tool is a claim, and finance teams are right to treat it as one.
What this means for your close
If AI is already part of your month-end, it is worth timing the review for one close. Note which outputs you accept after a quick look and which ones you end up rebuilding. The ones you rebuild usually have the same problem: no tie-out, no trace back to source, no list of what was left unresolved.
When you evaluate any tool for the close, ask the three ASA 500 questions of its output before you ask about its features. Is it complete, is it accurate against an independent source, and is it detailed enough to check? A tool that answers those on the page will save you time. One that does not will move the work from preparing to checking.
That principle is the basis of how we built Vert: calculations done in code, every payout proven back to the deposit to the cent, exceptions listed with their reasons, and nothing posted until you approve it. We go further into why in Why Finance Needs Deterministic Agents. If you would like to see that on a month of your own payouts, you can join the pilot.