How to Audit Your AI Tool Spend in an Afternoon
Eleven products, $940 a month, two doing the same job and three billing for people who left. None of it was reckless. It accumulated one decision at a time.
A four person agency added up its AI subscriptions last month and found $940 a month across eleven products. Two of them did the same job. Three were still billing for people who had left. One was a personal card nobody had expensed in seven months. Nothing here was reckless, it accumulated one reasonable decision at a time. Knowing how to audit your AI tool spend is less about discipline than about doing a specific, finite piece of work once, and it genuinely fits in an afternoon.
How to audit your AI tool spend, step one: find everything
You cannot audit what you cannot see, and AI spend hides better than most software spend because so much of it starts as an individual trial.
Pull from four places, in this order:
Card and bank statements, last three months. Search descriptors rather than product names. Charges from AI vendors are often billed through payment platforms and appear under names you will not recognise.
Every email inbox, searching for "receipt", "invoice" and "your subscription". This catches personal-card purchases that never reached accounting.
Your identity provider's connected apps list. Google Workspace administrators can view and manage third-party app access, and Microsoft's equivalent sits under enterprise application management. Both surface tools people signed into with a work account and never expensed.
Ask the team, once, in writing. "List every AI tool you use, including free ones." The free ones matter because they are next year's paid ones and because they carry the same data questions.
That fourth step routinely doubles the list. It also surfaces unsanctioned tools already in use, which is a governance finding rather than a spend one, but you will find it here whether you were looking or not.
Step two: one row per tool, four columns that decide everything
Put every tool in a single sheet. Resist adding columns. These four answer the question.
Column | What goes in it | Why it decides |
|---|---|---|
Monthly cost | Normalised to monthly, annual divided by 12 | Ranks the list |
Seats paid vs seats active | "8 paid / 3 used in last 30 days" | Finds the biggest single leak |
Job it does | One short phrase, in your words | Exposes duplicates |
Last real use | A date, from usage data not memory | Separates dormant from valuable |
The "job it does" column is the one people skip and the one that finds the money. Write it in plain language and do not use the vendor's positioning. Two products that both say "AI writing assistant" are duplicates. Two products where you wrote "drafts client emails" and "generates blog outlines" may not be.
Step three: triage into three buckets
Every row lands in exactly one bucket. Decide as you go, do not accumulate a list to think about later.
Seat waste. You pay per seat and are paying for more than you use. This is almost always the largest number in the audit, and it is the easiest to fix because it requires no behaviour change from anyone. Downgrade the seat count to actual active users plus one.
Orphans. Nobody has used it in 60 days. Cancel. The instinct to keep it "in case" is what produced the eleven-product list. If you genuinely need it again, resubscribing takes four minutes, and you will have paid nothing in the interim.
Duplicates. Two or more tools with the same entry in the "job it does" column. Keep the one with better data terms, not the one people prefer, unless the preference is strong enough that dropping it means nobody uses either. Duplicates are the most politically awkward bucket and worth handling by conversation rather than by cancellation email.
The two line items that are almost always double-paid
Two specific overlaps show up in nearly every audit of a small team.
Transcription. Your video conferencing platform almost certainly includes it in the plan you already pay for, while somebody is separately subscribed to a dedicated transcription product. Check what your existing plan covers before renewing the standalone tool.
General assistants. A team of five frequently has a mix of individual consumer subscriptions to two or three different assistants, bought personally at different times. Consolidating to one business plan usually costs less than the individual subscriptions combined and, more importantly, changes the data terms, since business tiers commonly exclude your content from training when consumer tiers do not. If you have never checked which of your tools does what with your inputs, that is worth resolving while you have everything in one sheet, using the approach in how to check if an AI tool trains on your data.
Usage-billed tools need a different check
Anything billed per token or per request does not fit the seat model, and the audit question changes from "how many seats" to "what is the trend".
For those, look at three months of usage rather than one, and ask what caused any spike. A single afternoon of experimentation can produce a bill that looks like a permanent cost increase but is not. Conversely, a slow month-over-month climb with flat customer numbers means something in your product is getting more expensive per user, which is a margin problem worth tracing rather than a spend problem worth cutting. The mechanics of bringing that number down are covered separately in how to reduce AI API costs.
Set a billing alert on every usage-based account before you close the sheet. Not a limit, an alert. Limits break production, alerts wake you up.
Close the loop so you do not repeat this in six months
The audit is worth little if the list regrows. Three lightweight controls prevent most of it:
One payment method for software. A single company card means the statement is the inventory. Personal cards are how tools go missing.
A named owner per tool. Not a committee, one person who is responsible for whether it is still needed. Tools without owners become orphans.
A renewal calendar. Every annual renewal gets a diary entry 30 days ahead. Annual plans are usually cheaper and are also where forgotten spend compounds fastest, because nobody sees a monthly reminder.
What you should not do is impose an approval process on trials. Making experimentation bureaucratic is how you end up back at shadow tooling, just with worse visibility. Cheap trials are fine. Untracked recurring charges are the problem.
What good looks like afterwards
A realistic outcome for a small team is a 25% to 40% reduction with no loss of capability, most of it from seat counts and orphans rather than from giving anything up. If your audit produces a much larger number, the cause is usually one forgotten annual plan.
Once the list is clean, the useful follow-up question is not "how do we spend less" but "what is this producing", which is a different exercise entirely and covered in how to measure AI ROI for a small business. Cutting waste is worth doing first because it needs no judgement calls, and the remaining spend is then small enough to argue about properly.
Frequently asked questions
How often should a small business audit its AI tool spend?
Twice a year is enough for most teams, with a quick pass whenever headcount changes. Departures are the single most common source of orphaned seats, so an audit within a month of anyone leaving catches the majority of the waste on its own.
What is a normal amount for a small business to spend on AI tools?
There is no benchmark worth trusting, because the useful figure is spend per active user against what it replaces. The framing that helps is in how much a small business should spend on AI tools, which is a budgeting question rather than an auditing one.
Should I cancel tools nobody has used in 30 days?
Sixty days is the safer threshold, because some tools are genuinely seasonal or tied to a quarterly process. At 60 days of true non-use, cancel, and note the date in case you need to re-evaluate later.
How do I handle tools people bought on personal cards?
Reimburse what is legitimate, then move it onto the company account rather than leaving it. Personal-card subscriptions are invisible to future audits and, more seriously, the account belongs to the individual rather than the business, which becomes a real problem when they leave.
Do free AI tools belong in the audit?
Yes, though not for cost reasons. They carry the same data handling questions as paid tools, and free tiers are where the next paid subscription comes from, so knowing what is in use is the point rather than what it costs. The wider adoption picture sits in AI for small business.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
