How it happens
Nobody decides to spend USD 3,000 a year on AI. It arrives one reasonable decision at a time.
March: the owner takes a USD 20 chat subscription. June: two staff want their own, so USD 60 a month. August: the accounting software adds an AI tier for USD 15. September: a transcription tool at USD 12, a social scheduler with AI writing at USD 29, an AI feature on the booking system at USD 10. November: a second chat product because somebody preferred it.
That is about USD 146 a month, or USD 1,752 a year, and not one of those decisions was wrong at the time. What is missing is the moment where somebody looks at the whole list.
Per-seat is the mechanism
Almost all of this is priced per person per month, which has a specific property: it grows when you hire and does not shrink when somebody stops using it.
A seat is not released when the person stops opening the app. It is released when somebody notices and cancels. In practice the gap between those two events is measured in quarters.
This is worth understanding as design rather than accident. Per-seat pricing is popular with vendors precisely because revenue tracks headcount rather than usage, which makes it predictable for them and invisible to you.
The quarterly audit, which takes an hour
Four columns, one row per tool: what it costs a month, who is on it, when each of them last used it, and what it replaced.
Getting the usage figure is the only slightly awkward part. Most business plans have an admin screen showing last sign-in. Where there is none, ask; people are honest about tools they have stopped using.
Then three decisions, in this order.
Anything nobody has opened in six weeks goes. Not “we will review it”, cancelled. If it turns out to be needed, resubscribing takes four minutes.
Anything duplicating something else loses one of them. Two chat subscriptions, two transcription tools, an AI writer alongside an AI feature in the tool you already pay for. Pick one and cancel the other, accepting that somebody will mildly prefer the one you dropped.
Anything whose replaced-thing column is blank gets challenged. From the earlier lesson on costs: price it against the hours or the cheaper tool it displaced. A blank column means it displaced nothing, which means it is an addition to the cost base justified by a feeling.
Businesses doing this for the first time typically cut 30 to 50%, and nobody notices the loss.
Watch the annual discount
Paying annually saves 15 to 20% and removes the monthly moment at which you might have cancelled. For a tool you have used daily for a year, take the discount. For anything under three months old, do not: you are buying a twelve-month commitment to a decision you have not yet tested.
Watch what the vendor changes
Two patterns worth diarising a check for.
The AI feature that becomes a tier. A capability included in your plan moves to a higher one at renewal. This is now common, and it usually arrives as an email about exciting improvements.
Credits. Increasingly, products meter AI usage in credits, with an allowance included and top-ups sold. The allowance is generous during launch and less so later. If a tool you depend on works this way, find out what a heavy month costs at top-up rates before you depend on it further.
Where to look for the ones you have forgotten
Two places find almost everything.
The card statement, twelve months of it, read line by line with a highlighter. Small recurring charges are invisible in a monthly total and obvious in a year of them. Look for the ones you cannot immediately name.
The app-permissions screen on your email and storage accounts, which lists everything anybody ever connected. This finds the tools that stopped being paid for and kept their access, which is the worse half of the problem.
The number to keep in your head
Total AI spend as a share of what it saves. From the costs lesson you have an hours figure for at least one task; put a rate against it.
If you are spending USD 1,700 a year and saving six hours a month, that is a defensible trade in most businesses. If you are spending USD 1,700 and cannot name the hours, the audit is overdue, and the honest answer may be that two of the tools were bought during a month when everybody was talking about AI.
Every lesson here is open. Your progress is recorded once you pass the AI at Work exam.
Check your understanding
A four-person firm pays for six AI subscriptions. The audit shows two nobody has opened since April, and the owner’s instinct is to keep them because they were useful when set up and may be needed again. What is the strongest argument against keeping them?
Pick the one you would defend. Your answer stays private.
Stuck on something?Ask Yetunde, your AI tutor