Choose the work first.

The research says the help is real. In an MIT study, people using AI finished writing tasks 40% faster, and the work itself scored 18% better. So the question facing a small operator was never whether AI helps. It's where — because your week has thirty things in it, AI could touch twenty of them, and picking the wrong one first is how this technology turns into a subscription you resent.

Playbook No. 03 gave you the recipe: one page that tells an AI how to do a piece of work your way. This playbook is the step before the recipe. Choosing what to cook. Here is a three-question test. A guide to making money with AI has been making the rounds this month, and stripped of the hype, its best idea is this filter. Run any task through it before you hand that task to AI.

Time to run it: about 30 minutes with your real task list — our editorial estimate, not a measured result. Cost: $0. Cadence: run it before every new AI adoption, and again when the work changes.

Does somebody already pay for this result?

Not "would it be nice." Money, salary, a budget line — yours or a client's. If the finished work is something people pay for today, AI help has somewhere real to land. If nobody has ever paid for it, you're automating a hobby.

Does AI make you faster — instead of doing it without you?

The valuable version of this technology drafts, sorts, summarizes, and prepares, while your judgment stays in charge. If the plan removes you from work where your name and your judgment are the product, the test fails — no matter how impressive the demo looks.

Does it need knowledge of your field?

Your donors, your clients, your neighborhood, your rules. The tasks where AI helps most are the ones it can't do well without what you know. Generic work gets generic results; work soaked in your knowledge is where the help compounds.

Three yeses: that's your first task. Anything less: wait.

Watch it work on a shortlist. Picture the director of a small neighborhood nonprofit — three staff, a waiting list, a grant report due every quarter. (An illustrative composite, drawn the way these decisions actually arrive — not one real organization's story.) Her list of candidates:

Drafting the quarterly grant report from her own program notes. People pay for this result — grants literally depend on it. AI assembles the draft from her records; she rewrites the parts only she can judge. And it runs on knowledge no outsider has: her programs, her funders, her numbers. Three yeses. This is the one.

Drafting donor thank-yous. Fundraising is paid work, the drafting speeds her up, and the knowledge is hers — which donor gave to what, and why. Three yeses again, a close second — and if you wrote the one-page playbook from Playbook No. 03, you already own this recipe.

Answering client questions on the website, unsupervised. Fails question two. The answers are the trust her organization runs on; a wrong one costs more than the hours saved. AI can draft answers for a person to approve — that's a different, smaller task. On its own, it's a no.

Sorting the bookkeeping into categories. Paid work, and AI can sort faster than she can. But the categories are generic — the same for every organization — so it fails question three, barely. Fine to hand over later. Not first.

Pick two skills that stack, not five

One more discipline from that same guide, and it's the one most people skip: pick two skills that stack, not five. For an operator, the pair is obvious. Learn to design the workflow — decide what the work is, in what order, with what checks — and learn to set the AI up once with your standing instructions, so you're not re-explaining yourself every session. The two feed each other: the workflow tells the AI what to do, the setup tells it who it's doing it for.

Every correction becomes a standing instruction

And the habit that makes both skills grow is one you already know from Playbook No. 03: when the AI gets it wrong, fix the recipe — and now you know the general rule behind it. Every correction becomes a standing instruction. Correct it once, in writing, and you never correct it the same way twice.

The house rules haven't changed. A person approves before anything leaves the building. The recipe lives in a document you own. This playbook just moves the decision upstream: the most expensive AI mistake a small operator can make isn't picking the wrong tool. It's handing good tools the wrong work.

The one action: this week, before you buy, build, or subscribe to anything, take the five tasks you're most tempted to hand to AI and run each one through the three questions, in writing. Start with the one that earns three yeses — and write its one-page recipe next.

The steelman for the skeptics: "Isn't this just slowing me down? Everything is automatable eventually — why not hand it all over now?" Because sequence is the whole economics. Start with work that fails the test and you spend your scarcest hours supervising a tool on tasks nobody pays for, in areas where your judgment was the product — and you conclude, wrongly, that the technology doesn't work. Start with a three-yeses task and the first win funds the patience for everything after. The test isn't a brake. It's the order of operations.

Time saved: the weeks people lose automating the wrong task first. Value: your scarcest hours aimed where the help is real.

Published Friday, November 6, 2026. The Issue #9 newsletter edition wraps this playbook and lands in inboxes separately — browse the newsletter archive. This page is the playbook itself, frozen as published: playbooks never change after they go up. If a better way appears, it gets a new number, not a rewrite.

Free forever. One playbook every Friday.

Back to the shelf · ← Playbook No. 08 — Never miss a caller