Someone is marking the AI's homework. Nobody is counting the hours.

The most expensive AI tool in your business probably isn't on any invoice. It's the person checking its output.

Every AI pitch counts the hours a tool saves. Hardly anyone counts the hours spent reading what it produced, fixing the bits that are wrong, and redoing the bits that were confidently made up. That second number now has a figure on it, and it isn't small.

Pax8, a cloud software distributor, published its Q3 2026 SMB AI Pulse Report on 12 October [1]. It surveyed 1,150 decision-makers at firms with 5 to 499 staff across the US, UK, Germany and Sweden, including 350 in the UK. The findings:

A caveat before anyone quotes this at a board meeting: Pax8 sells AI and cloud services through IT providers, so it has an interest in the answer being "you need help managing this". The UK margin of error is about ±5 points. Compare the Cloud's write-up gives UK-only splits for some figures (44% spending three-plus hours checking, 24% finding unapproved AI) that Pax8's own release doesn't break out, so treat those as reported rather than verified [2].

Even allowing for that, the shape is clear. AI is now ordinary: the ONS says 30% of UK businesses used at least one AI technology in September, up from 9% three years earlier [3]. The checking has grown with it, and in most businesses it sits nowhere in the numbers.

The best evidence isn't in a survey. It's in what accountants say about their clients.

"An AI tool linked to Xero misclassified items and invented commentary." That's from AccountingWEB's editorial piece of 3 August 2026, "Accountants count the cost of clients' AI use", where Dom Noakes of HaysMac described what a client's AI add-on had done to their books [4]. Another practitioner described a client's AI-converted cash book with numerous transactions simply missing.

The 26 comments underneath are blunter. One reader (4 Aug) described AI "balancing" a set of books by inventing offsetting errors, and said AI-assisted clients had doubled the time it took to prepare their VAT returns. Another said AI-assisted data entry can double the fee because of the correction work. A third described an automated QuickBooks query email that wrongly accused a client of misreporting a foreign-currency transaction, and the apology that followed.

The survey behind that article was commissioned by Dext, which sells bookkeeping automation, and run by Censuswide in December 2025: 500 UK accountants and bookkeepers, half of whom had seen clients lose money after acting on incorrect AI advice. Most of those who meet these errors spend up to ten hours a month correcting them [5]. It's vendor-funded and ten months old, so take the direction, not the decimals.

And the question at the heart of all this was asked on UK Business Forums back in June 2023, in a thread titled "Do you encourage your team to use ChatGPT?" (old, but it has aged well). After several members described factually wrong answers, one asked whether having to verify everything elsewhere defeats the productivity gain [6]. Three years on, most firms still can't answer that with a number.

Take an 80-person wholesaler. Its finance and sales teams use AI for four jobs. Here's what each one looks like if someone writes down the time honestly for a month:

Illustrative figures, not survey data.

The vendor dashboards would say 27 hours a week. The truth is 13.5. Half the time "saved" goes on checking and fixing.

At a loaded cost of £35 an hour over 46 working weeks, that's the difference between about £43,500 a year of claimed value and £21,700 of real value. Still worth having. But one line costs more than it saves: the month-end commentary, where the person checking has to rebuild the reasoning from scratch because the AI wrote a plausible explanation for a margin dip that didn't match the ledger.

That's the one to fix or switch off, and nobody would know without the table.

This is a small model in Power BI or Excel, built in about two days, and it treats AI like any other cost centre: what it consumes, what it produces, how often it's wrong.

1. Sources

2. Model

3. Measures

4. Alerts

5. Dashboard

One page is enough: a keep / fix / kill grid with net hours saved on one axis and severe error rate on the other, each use case as a dot, plus a table underneath with owner, review level and cost per net hour.

Sampling, not surveillance. You don't need to check everything. Start by checking 20 items in every 100 for a new use case. When the error rate has stayed low for a month, drop to 5 in 100. When it rises, go back up. Auditors have worked this way for decades, and it turns "we trust it" into "we checked 400 of them and found 3".

Caveats. The pulse hours are self-reported and soft: use them for direction, not decimals. Keep the check log about the output, not the person, or it becomes a performance tool and people stop logging honestly. And if staff names go into the log, your privacy notice needs to cover it.

Not going to build a model? Do these this month:

Measuring hours and ignoring severity. A tool that saves ten hours and once a quarter sends a wrong figure to the bank isn't a ten-hour win. Weight the severe errors.

Checking everything forever. Full review on a mature, low-risk use case is waste, and it's how a good tool gets a bad reputation. Let the sample rate fall as the evidence builds.

Blaming the tool for a process problem. If invoice coding errors cluster on three suppliers with messy master data, fix the data. The AI is just finding the mess faster.

Turning the log into a stick. The first time someone gets told off because the log showed they spent four hours fixing AI output, the log goes quiet. Reward the people who find errors.

AI isn't failing small businesses. It's being measured on the half of the ledger that flatters it. Put the checking time next to the time saved and most firms will find two or three use cases that pay off handsomely, and one that's quietly costing them a day a week.

If you'd like that picture drawn properly, covering your tools, data and controls with a keep / fix / kill list at the end, that's what our AI & Analytics Maturity Audit does. Reply and I'll send the outline.

— Aksel