An owner I sat with last month had her AI subscription totals on a printed page in front of her. £800 a month across five tools, growing. She knew that number to the pound. When I asked her how many staff hours a week the firm now spent checking AI output, supervising AI runs, or rewriting something AI got wrong, she paused, looked at the ceiling, and said honestly that she had no idea. Probably a lot. She had not been counting. Nobody had.
That gap is the post. Every AI use case carries a labour cost that does not get logged. The visible cost is the subscription bill, sitting tidily in the management accounts. The invisible cost is the staff time spent overseeing the tool while it runs, evaluating what it produced, and recovering when something slipped through. Sized honestly, that hidden labour tax commonly runs between five and fifteen per cent of gross margin in firms running heavier AI use, and it never shows up in the AI column of the P&L because nobody is coding their hours that way.
The aim here is to put the second half of the cost equation back on the table so owners can make decisions with both numbers in view. Where the gain is real, you keep going. Where the oversight is eating the gain, you redesign or you stop. Neither call is possible if you only see the subscription side of the ledger.
What is the hidden labour tax of AI?
It is the staff time spent on three things that AI use generates and traditional accounting does not separate. Oversight is the time staff spend supervising AI tools as they operate, keeping a human in the loop. Evaluation is the deeper checking of AI output against quality, accuracy and risk standards. Recovery is the time spent fixing, rewriting, or remediating AI work that was unusable or wrong.
Workday’s 2026 research puts a number on it. Across a large employee survey, the report finds that nearly forty per cent of AI time savings are consumed by rework, summarised as “for every ten hours of efficiency gained through AI, nearly four hours are lost to fixing its output”. CPA.com’s 2025 accounting data, citing Gartner, runs higher in regulated work, around sixty-nine per cent of the time saved is reabsorbed.
Why does it matter for your business?
The maths only works if both sides of the cost equation are visible. The vendor brochure shows the time saved. The subscription bill shows the licence cost. Neither shows the hours your senior people now spend on oversight, evaluation and recovery, often the highest hourly cost in the firm. Hubstaff finds that while eighty-five per cent of professionals say they use AI, AI-labelled activity is about four per cent of tracked work time.
For a twenty-person firm at 50 pounds loaded per hour, two hours per person per week of AI oversight is roughly 100,000 pounds a year. Against an AI tool stack of 30,000 pounds a year, the labour tax is three times the visible cost and entirely absent from the AI line. In a firm doing 1.5 million pounds of revenue at 35 per cent gross margin, that is around 19 per cent of gross profit going to a cost you have never sized.
Where will you actually meet it?
You meet it in the gap between the productivity numbers people quote in conversation and the bottom-line numbers in the management pack. Staff feel busier, delivering more output, or output of similar quality faster. The vendor case study said forty per cent time saved. Profit per head is unchanged, or worse. That disconnect is the labour tax sitting underneath the surface, absorbing the gain into oversight and expanded workload nobody priced for.
You meet it most acutely in regulated and high-stakes work. A junior uses AI to structure a client memo, a senior spends longer than usual unpicking subtle inaccuracies, the extra time is booked as “partner review” with no flag that the inaccuracies came from AI. A tax manager runs an AI-assisted reconciliation, then an hour goes into investigating anomalies that turn out to be AI edge cases, the hour is booked as “reconciliation review” and the source is invisible.
When does it kill the ROI, and when does it not?
It kills the ROI in three categories of work. High-stakes regulated work where oversight is structurally required, audit, compliance, regulated financial advice. Fact-heavy client-facing work where every claim needs verification before it leaves the firm and the verification time often exceeds the drafting time AI saved. Decision-recommendation work where the cost of acting on a wrong AI output is high and so the evaluation has to be careful and slow.
It does not kill the ROI in three other categories. Low-stakes drafting where minor errors carry little cost. Throughput tasks where AI provides a first pass that humans accept or reject quickly. Internal first-passes that are never seen externally and only need to be roughly right to be useful. The strategic move is to be honest about which category each AI use case sits in before you commit to a year of licences.
How do you size it in a week, without timesheets?
You size it by sampling, not by tracking everything. Pick three to five workflows where AI is in regular use and the work affects revenue or risk. Client report drafting. Management accounts. Marketing campaign production. Audit workpapers. Whichever fits. For each, pick one or two staff members already using AI in that workflow, tell them what you are doing, ask for their help for a week. The goal is honest ratios.
During the week, ask them to annotate their calendars and the artefacts they produce with two simple markers. Where AI was used, and roughly how much time was spent supervising or correcting its output. A one-hour block labelled “Client X report” might be annotated as thirty minutes of AI drafting, twenty-five minutes of intensive review, and five minutes of cleanup. You are after orders of magnitude, not stopwatch readings.
At the end of the week, sit down with each person for thirty minutes and walk the entries. Classify the AI-related time into the three buckets, oversight, evaluation, recovery. Note any incidents where AI mistakes caused downstream problems, client emails, or rework, even if those were not recorded at the time. Then do the arithmetic. Total task time, AI-assisted share, oversight-and-recovery share within that, extrapolated to staff numbers and loaded hourly cost.
The figure will not be precise. It will be honest, which is what is missing. Once the line exists, you can decide where to redesign the workflow, where to invest in better prompts and templates to cut oversight, where to reprice the work because the firm is now delivering it faster, and where to scale AI back because the oversight burden is eating the gain. There is a sister post on the staff time tax of running AI consulting engagements, which sits one level up at the vendor and project layer, where this post stays inside the firm’s own tool use.
If you want to find out what your own AI labour tax actually is, Book a conversation.
