The Real Cost of AI at Work

Licence price is the smallest line on the bill

Finance sees a seat price and calls it the cost of AI at work. Operators see the rest of the bill. Someone writes a brief. Someone waits on the draft. Someone checks names, numbers, and promises. Someone rewrites the generic middle. Someone explains the error that already reached a customer. Those hours are cost. Cheap drafts that need expensive checking are not cheap. They have moved typing into review and hoped nobody would notice. A serious cost picture adds licences, recovered time that actually returned to the diary, and cleanup that would not have existed without the fluent first version.

Two teams can hold the same ChatGPT Business invoice and live in different economies. One team briefs, generates once, and checks against the file. The draft shortens the afternoon. The other team pastes a vague ask, accepts a long page, and spends the evening proving the percentages. Same licence. Opposite unit cost. Copilots inside Office follow the same split. A licence that sits on every seat while three people use it well and forty people generate unread summaries is not a bargain. It is a tax on attention with a vendor name on the invoice.

This essay is a cost sheet you can take to a budget meeting without theatre. You will not get a universal price. You will get the lines that belong on the sheet: seats, admin, training, duplicate tools, review minutes, and error correction. Ignore those and every pilot looks cheap. Include them and some pilots should stop. Stopping a noisy tool is a saving. That sentence is hard in organisations that treat every new seat as progress. Progress is trusted output per hour, not tokens per seat. Write the sheet so a finance partner can challenge the assumptions in one sitting.

Count review time, cleanup, and duplicate tools

Build the sheet in four blocks. Licences and admin: seats, identity, logging, legal review of the terms. Time: briefing, generation, checking, rewrite. Errors: incidents, customer recovery, restated numbers in a later meeting. Duplicates: the same file run through chat and a copilot because nobody wrote a default. The last block is where money hides. People pay twice and still paste into the weaker boundary. A default tool per artefact is a cost control. A menu of seven assistants is a cost multiplier. Training belongs in block one. A feature tour is not training. A decision rule is.

Measure time-to-trusted-output on a real artefact, not tokens. A copilot that turns a two-page memo into six pages of summary has increased reading cost. ChatGPT that invents a figure you then disprove has increased research cost. Put those minutes next to the list price. The comparison that matters is the trusted artefact, signed, in the time you actually had. Anything else is a demo metric. Demo metrics are how expensive tools stay because they look busy. If the signed page did not arrive sooner, the seat did not earn its place, however impressive the draft looked in the chat.

A cost sheet a manager can defend

Pick five artefacts the team already ships. For each, write the current minutes, the AI path, the review minutes, and whether the output was trusted on first check. Run that for two weeks. You now have a unit cost, not a feeling. Add the invoice. Add any incident that would not have happened without a generated sentence. Share the sheet. Kill the path where review exceeded the old typing time. Keep the path where checking was short because the brief was good. That keep-and-kill list is the budget proposal. It is also the adoption plan, which is why finance and operations should see the same numbers.

Price the default tool, not the catalogue. One approved chat for thinking on allowed notes. One copilot where the file already lives. Neither for restricted classes until legal says yes. Write that on one page. Every extra icon is a training cost and a leak path. Revisit quarterly. Seats that produced no trusted artefact lose their licence. That rule sounds harsh. It is how you stop paying for wallpaper. Wallpaper is a feature list that nobody used on a deliverable that left the team. Unused seats still need admin, logging, and an explanation in the next audit.

Role: You are a chief of staff preparing a cost note I will take to a budget meeting.
Task: Turn my figures into a one-page cost sheet for workplace AI.
Context: I will paste seat prices, two weeks of review minutes, incidents, and duplicate tools.
Constraints:
- Do not invent savings. If a number is missing, write missing.
- Separate licence cost from review cost from cleanup cost.
- Do not call a longer draft a productivity gain.
- Flag any path where review exceeded the old typing time.
Output: A table of cost lines, a keep-or-kill note per artefact, and open questions.
Quality checks: List assumptions a finance partner would challenge.

If the quality list is longer than the savings list, you are not ready to renew seats. You are ready to run another two weeks of measurement on the same artefacts, with the same review rule. Renewing on a vendor slide is how the hidden cost becomes a permanent line. A finance partner will not be persuaded by tokens. They will be persuaded by a keep-or-kill list that survived contact with real work and named the minutes that did not return.

Cheap drafts that create expensive checking

The hidden cost is senior review of junior-looking fluency. A generated pack that is almost right takes longer to correct than a short human draft, because the errors are smooth. People read for tone and miss the date. The meeting then runs on a wrong figure. The cleanup is a second meeting. Count that second meeting. If your culture praises volume, the tool will produce volume. Volume is not value. Value is a page someone will sign without a forensic afternoon. Forensic afternoons are how cheap seats become expensive people.

A second leak is dual running. The team keeps the old process and adds AI on top, then counts both as necessary. Dual running is a pilot phase, not a year. Set an end date. Either the AI path replaces a step or it leaves. A third leak is training as entertainment: a webinar, no decision rule, no review checklist. People return to vague prompts and the review tax returns with them. Pay for a short rule, not a long show. The rule should name the default tool, the data class, and the person who still signs. Without those three, the webinar was a cost with sandwiches.

Related reading on StudyGrid

Read next: How to Measure AI Productivity Adoption and Value How to Choose Workplace AI Tools. Those essays sit beside this one. Use them when you need the neighbouring skill, not as a substitute for the check you still have to make.

What to do this week

Time five real artefacts for two weeks: brief, generate, check, send. Put those minutes next to the licence line and next to any cleanup that would not have existed without a generated sentence. Kill one path that costs more than typing. Keep one path that returns time to the diary. Share the sheet with finance and with the people who do the review. That is the real cost of AI at work, written small enough to act on before renewal.