Audit Your AI Spend and Usage

You can't fix a stack you can't see

Ask most people who use AI daily how many AI subscriptions they're actually paying for, and you'll get a shrug, a guess, or a number that's wrong by two or three. That's not a character flaw -- it's what happens when tools get added one at a time, each for a reason that made sense in the moment, with nobody ever stepping back to look at the whole picture. A trial that auto-converted to paid. A second assistant a colleague swore by. An image tool bought for one project, never cancelled. Each decision was locally reasonable. The stack, as a whole, was never designed.

This track exists to fix that, and it starts here, because every later lesson -- picking a primary assistant, building the thinker/searcher/maker pattern, running a monthly review -- depends on you actually knowing what's in your stack today. Guessing doesn't work. You need a real audit.

The audit is two columns, not one

The instinct is to audit cost: pull up your bank statement, list every AI-adjacent charge, add them up, wince. That's useful, but it's the wrong primary signal. A $20/month subscription you use five days a week for real work is a bargain. A $12/month subscription you haven't opened in six weeks is dead weight regardless of how small the number looks. Cost tells you what you're spending. It doesn't tell you what you're getting.

The audit that actually changes behavior has two columns working together: what job does this tool do, and how often do you actually use it. Put a tool's job in plain language -- not "AI assistant," but "drafting emails and quick answers." Not "research tool," but "sourced answers I can cite." The job description is what makes the next step possible: spotting when two tools are quietly doing the same job.

Look at the worked example above. Two of the five tools have the exact same "actual job" column -- one general assistant used five to six days a week, and a second general assistant, ostensibly a "backup," used zero to one days a week. That's not redundancy as insurance. That's redundancy as unexamined spend. The audit's job is to make that visible, because it's invisible on a bank statement -- both charges just look like "AI subscription, $20."

How to actually run it: a one-week log, not a guess

The version of this audit that works is embarrassingly simple and takes about ten minutes a day: every time you open an AI tool for a real task, write one line. Tool, task, roughly how long. Do this for seven days, not one -- a single day can't tell a genuinely idle tool from a tool you use every Monday for a weekly report. Underestimating the observation window is the single most common way people flatten a real, if intermittent, usage pattern into a false "I barely use that" verdict.

At the end of the week, you tag every logged session by the JOB it did, not the tool it happened in. This is the step people skip, and it's the one that makes the audit worth running at all -- tagging by job is what surfaces the overlap. If "drafting a client email" shows up next to two different tool names across your week, you've found your first flag.

Reading the flags: keep, overlap, idle

Once the week is tagged and totaled, every line in your stack gets one of three labels. Keep means real weekly use on a job nothing else in your stack covers -- no action needed. Overlap means the job is already covered by another kept tool -- this is your highest-confidence cut, because you're not losing any capability by dropping it, only cost. Idle means low or no use, but with no other tool covering the same role -- this one deserves a real judgment call, not an automatic cut, because "idle" sometimes just means "I forgot this could do X," not "I don't need X."

What this sets up

Everything downstream in this track assumes you did this audit honestly. Lesson 776 asks you to pick a primary assistant -- that choice is easier and more defensible once you already know which tool you reach for most and why. Lesson 777 builds the thinker/searcher/maker pattern directly on top of your "actual job" column. And lesson 782, the monthly review ritual, is this exact same audit, run again on a schedule, so the stack never has the chance to drift back into the sprawl you just cleaned up.

If you skip this lesson's exercise and jump straight to "which assistant should I use," you're optimizing a stack you haven't actually looked at. Do the week-long log first. Everything after this gets easier because of it.