The four AI tools that actually stuck. A six-month receipt.

When the bill came in for May, I noticed something. Of the eleven AI subscriptions we’d bought across the year, four were still being used. Seven were either cancelled or, worse, being silently paid for by a team that had quietly stopped opening them.

I spent the next month doing something unfashionable: I went and looked at the actual usage logs.

The four that stuck

Whisper transcription for meeting notes. Free, open-weights, runs on a Worker. The team uses it without thinking. It writes the transcript. They read it. End of story. No prompt engineering, no preference settings, no Slack thread about “have you tried this one”.

A small local LLM for code review summaries. Llama 3.1 8B, also free, runs on the edge. It’s not as good as Claude or GPT-4 but it’s instant and it never breaks the build. The team kept using it because it was always there.

One paid Anthropic seat for the lead engineer to do the heavy thinking on architecture decisions. We used to have five seats. Four were untouched after the first six weeks.

A grammar / clarity checker built into the doc tool. Not separately budgeted, but worth flagging because it gets used more than any of the standalone writing assistants we’d bought.

That’s it. Four tools. About $40/month all-in for a team of fourteen.

The pattern

Every one of them shares a property I’ll call “workflow disappearance.” You don’t open the tool. The tool runs inside something you were going to do anyway.

You don’t open Whisper. You finish the meeting and the transcript is in the channel. You don’t open the small LLM. The PR comment summary appears at the top of the review. You don’t open the grammar checker. It underlines a sentence and you keep typing.

The tools that died had the opposite property. A separate app, a separate tab, a separate mental model, and a small daily cost in attention to remember they existed.

The cost-of-attention is real

I think most AI procurement decisions in 2025 underweighted this. The bake-off was about output quality and capability ceiling. Nobody asked, “Does the team have to remember this exists?”

For mid-market enterprises with already-overloaded teams, the answer to that question turned out to be the only one that mattered. The capability gap between “the best chat model” and “a decent inline assistant” was smaller than the attention gap between “another tab” and “no tab.”

What I’d ask before renewing

Three questions, in this order:

Is anyone opening this thing daily? Not “do they say they use it”. Does the usage log say they use it?
If we cancelled tomorrow, would anyone notice within a week? Be honest. If the answer is “probably not within a month,” the tool isn’t sticking.
What workflow does this disappear into? If the answer is “well, you have to remember to do something first”, you’ve already paid for it twice.

The four tools that stuck answered “yes / yes / something we were doing anyway” to all three.

The seven that didn’t, didn’t.

This is part one of a short series on what the post-honeymoon period of generative AI looks like inside a programme office. Part two will cover what we bought to replace the cancelled tools.