We audited our AI subscriptions and cancelled most of them

Nobody on the team could tell you, before we started, how many AI tools we were paying for. Not roughly — literally could not produce the number. The finance spreadsheet said eleven line items under a category someone had labelled "software — AI" at some point in the last two years, and when we read the list out loud in a meeting, three people asked what one of the tools even did.

That is the honest starting point, and it is worth sitting with before the audit results, because the instinctive response to a bloated subscription list is to blame the buying decisions. Somebody signed up for something they didn't need. That is rarely what actually happened. Every one of the eleven had a legitimate reason to exist on the day it was bought. The problem was never any single tool's justification. It was that nobody had ever looked at the list as a list.

What the usage log actually showed

We pulled three months of login and session data from every tool with an admin panel that offered it, and asked whoever owned the rest to estimate honestly from memory. The picture that came back was lopsided in a way none of us predicted going in.

Four tools accounted for what we'd guess was somewhere around nine in every ten sessions logged across the whole stack: the coding assistant embedded in the editor, the transcription tool that turns every call into searchable notes, the writing assistant used for first drafts of anything client-facing, and a research tool used almost daily for source-checking. Those four were opened multiple times a week, by multiple people, without anyone having to be reminded they existed.

The other seven had a completely different shape. Not zero use — every one of them had at least a handful of sessions somewhere in the log — but concentrated into a burst of a week or two, almost always traceable to a single project, and then nothing. An image generator that got heavy use during one product launch and hasn't been opened since. An automation tool bought to wire together two systems for a client migration that finished in March. A second coding assistant, evaluated properly for about ten days when the first one was having a rough patch with a particular language, then quietly abandoned once the patch resolved itself. None of these were bad tools. They were tools bought to solve a problem that had an end date, and the subscription didn't know the problem was over.

The pattern behind every wasted subscription

Once we laid the seven low-use tools next to each other, they all told the same story with different props: someone had a task, the task needed a tool, the tool got bought on a card that renews itself, and the task finished while the subscription kept running. Nobody made a bad decision at the point of purchase. The decision that was missing happened later, and it was never a decision at all — it was the absence of one. Cancelling requires an active step. Continuing to pay requires nothing.

This is worth stating plainly because it changes what the fix has to be. If the problem were bad buying decisions, the fix would be a stricter approval process before anyone signs up for anything, and that is the fix most finance teams reach for first. It would have prevented none of our seven. Every one of those purchases was reasonable in the moment it was made. What was missing was a moment, recurring on a schedule, to ask whether it was still reasonable now. That is an attention problem, not a purchasing problem, and it needs a scheduling answer rather than a policy answer.

Three tools, one job, arrived from three directions

The audit's second finding was less about waste than about duplication, and it was the one that made people in the room visibly uncomfortable, because unlike the seven forgotten subscriptions, this one was actively being used — just three times over.

We had a meeting-notes feature bundled into the video-calling platform we already pay for regardless. We had the standalone transcription tool mentioned above, bought before the video platform added its own version and never revisited once it did. And we had a note-taking app with an AI summary feature that a few people had started using because it synced better with their personal setup. All three were doing the same job — turning a conversation into text someone could search later — for three different people who had each solved the same problem independently and never compared notes, which is a slightly funny sentence to write about a note-taking tool.

None of the three people who'd adopted their version was wrong to want it. The problem is structural: overlap like this doesn't announce itself the way an unused subscription does. An unused tool shows up as zero in a usage log. A triplicated tool shows up as three normal, healthy-looking numbers, and normal, healthy-looking numbers don't get questioned in a finance review — they get renewed. Catching this kind of waste needs the same person to be looking at the whole list at once, not three different people each satisfied with their own corner of it.

What broke when we cancelled

We cut eight of the eleven — the seven low-use tools plus two of the three overlapping transcription tools, keeping the one built into the video platform since it was already paid for either way. The honest report on what broke is less dramatic than a cost-cutting article usually wants it to be, and also not nothing, which is the useful, unglamorous truth.

One person lost access to a project archive inside the abandoned automation tool that turned out to still be referenced by a live client dashboard, and rebuilding that reference took roughly a day of someone's time — the exact "exit cost" line we've written about at length in what AI tools actually cost, showing up exactly where that piece predicts it will: not at cancellation, but the day after, in the state your work is left in. Two of the cancellations were completely silent — nobody noticed, nobody asked, the renewal simply didn't happen. And one turned out to be missed within a week, not because the work needed it but because a single person had built a personal habit around a feature nobody else used, which is its own small lesson: a subscription with a usage count of one is either genuinely dispensable or it is one person's entire workflow, and the log alone can't tell you which.

The audit that takes half an hour

We didn't build new software to do this and we're not going to recommend any. The process that worked is a recurring calendar entry, once a quarter, with three steps that take about thirty minutes if the list stays under fifteen tools, which for most small teams it should.

Pull the login or usage history for every tool that has one. For the ones that don't, ask the owner to name, from memory, the last time they used it — an honest "I don't remember" is itself the answer. List every tool that does something another tool on the list also does, and put both names in front of the person who has to choose, rather than letting the choice happen by default to whoever adopted their tool first. Then cancel anything with no real answer to "who used this and when," and set a note for anything you're keeping "just in case" to be reviewed again next quarter rather than left alone indefinitely — a "just in case" subscription that survives three consecutive reviews without being used has stopped being insurance and has become the seven-tool problem all over again, just spread out over more quarters. Free tools sitting unused cost nothing and don't need this treatment at all, which is worth remembering before adding a free trial to the audit — the ones worth taking seriously as a genuinely low-risk way to try something are covered separately in the free tiers worth taking seriously.

What we'd buy again without hesitating

The four tools that carried nine-tenths of the usage log are also the four we'd resubscribe to immediately if the whole stack vanished tomorrow, and it's worth naming why, because it isn't the same reason in each case. The coding assistant and the research tool earned their place through raw frequency — opened daily, by more than one person, without anyone needing to be reminded. The transcription tool earned it through a single job done reliably enough that nobody has looked for a replacement in over a year. The writing assistant earned it through the least glamorous reason of all, which is that removing it for a week as a test produced an immediate, specific complaint from the person who used it most, which is a more honest signal than any usage graph.

The seven we cut, and the two duplicated transcription tools alongside them, weren't bad purchases. They were purchases nobody had ever been asked to defend twice. A surprising number of vendors in this category don't survive their own market for long either — a tool that looked essential eighteen months ago can simply not be there to renew, which is its own argument for reviewing on a schedule rather than trusting a subscription to still make sense because it once did; we keep a running account of which tools in this space quietly disappeared in the AI tool graveyard. The eleven-tool list wasn't a pricing failure. Every one of those prices was published, reasonable, and exactly what it said on the page. It was an attention failure, and the fix for an attention failure was never going to be a better spreadsheet at the moment of purchase. It was asking the same question again, on a schedule, after the moment of purchase had been forgotten.