How to audit your personal AI tools in 60 minutes.

You probably can't name every AI tool touching your work right now. There's the obvious ones you open daily, and then there's the ones renewing quietly at $20 a month since March, doing a job something else already does. The stack audit is how you see the whole thing at once — and it takes an hour.

News coverage keeps circling the same finding: people using a handful of AI tools at once get slower, not faster. The sharpest data point so far is a Boston Consulting Group study published in Harvard Business Review in March 2026, which surveyed 1,488 full-time US workers: self-reported productivity rises with one to three AI tools, then plummets once a fourth joins the mix — what the researchers call “brain fry.” Fourteen percent of workers already report feeling it. Separate MIT Media Lab research points the same direction from another angle: EEG scans found LLM users showed up to 55% weaker brain connectivity than people working unaided.

What the coverage misses is why the load accumulates in the first place — because nobody ever holds the full list in one place. Tool creep happens one reasonable decision at a time. A teammate recommends something, a free trial looks good, a new model drop makes the old assistant feel stale. Each addition makes sense. The accumulation doesn't, because the additions were never in conversation with each other. Nobody chose a stack. A stack happened.

The fix is not discipline. It's an inventory with a decision attached to every line.

1 — Inventory (10 minutes). List every AI tool that touches your work. Everything: the chat assistants, the writing helpers, the image tools, the agents, the AI features buried inside software you already pay for (your editor, your CRM, your email client — those count). Include free tools. If it touches your work, it goes on the list. Most people land between 8 and 20 entries. If you land below 8, check your browser extensions and phone apps.

2 — Assign the job (10 minutes). Next to each tool, write the one job it does. One job, stated as an outcome: first drafts, code review, meeting notes, image work. Here's the honest test: if you can't name the job in five words, the tool doesn't have one. If two tools share a job, flag both — you'll deal with that pair in block six.

3 — Follow the data (10 minutes). For each tool, one question: what leaves my machine, and where does it go? A tool that reads your code and sends it to a third-party model is a different animal from one that runs local, even if the output looks identical. You don't need to read every privacy policy. You need to know, for each line, whether the answer is my files stay mine or my files train someone's model. That answer belongs on the list, because it changes the keep / merge / kill math later.

4 — Last used (5 minutes). Sort the list by when you last actually opened it. This block exists to be uncomfortable. The tool you forgot was on this list is the tool you've been subsidizing.

5 — Renewal dates (10 minutes). Add the billing date and cost for every paid entry. Note which ones auto-renew inside 30 days — those get decided this week, not someday, because the clock is running whether you decide or not.

6 — Keep, merge, or kill (15 minutes).Now the decisions, against one constraint: a personal stack has a limited number of real slots. One anchor you build your work inside, a few tools doing active daily jobs, a short bench of supporting tools — the 1:3:5 shape, though you don't need to adopt the framework to run the audit. The constraint is the point: nine slots, roughly. A tool earns a slot or it doesn't.

• Keep: does a named job, used in the last month, data behavior you accept.
• Merge: two tools, one job — pick the one you'd miss, and actually cancel the other this week.
• Kill: no job, no recent use, or a renewal you forgot about. Cancel today, not "before the next renewal."

The builder runs the audit and finds three assistants, two of which exist because a new model launch made the last one feel stale. The merge is easy; the harder discovery is that the "AI code reviewer" with repo access was never in the data check. That line gets re-decided, not just kept.

The founder finds the stack split between personal habits and company tooling, with the company paying for four seats of a tool two people use. The kill here isn't cognitive, it's financial — and the renewal block is what surfaces it.

The operator finds no obvious fat, but discovers six tools touching client data with six different answers to the data question. Nothing gets killed. The stack gets a rule instead: one answer to that question, or the tool doesn't touch client work.