You'd never onboard a human like this

Picture hiring your first employee and, on day one, handing them your client list, your pricing, your unfinished legal disputes and your inbox password — then telling them to "just help out". You'd never do it. Yet that's exactly how most solo founders onboard AI: full access, no job description, no review.

The result isn't usually a disaster. It's worse — it's quiet mediocrity. The tool does a bit of everything, owns nothing, and you can't tell whether it's helping because you never defined what help looks like.

The decoded idea: AI is staff, not software. Software does the same thing every time; staff need a role, boundaries and a manager. The founders getting real leverage from AI in 2026 aren't the ones with the most tools — they're the ones who wrote the job description.

A job description for AI fits on one page, and it has exactly three sections:

1. Allowed — what it does unsupervised (as first drafts). First drafts of emails, posts and proposals. Research with sources you'll check. Summaries of long documents. Formatting, renaming, reorganising. Meeting notes. This is the intern work: high volume, low stakes, always reviewed before it ships.

2. Never touch — what it doesn't get, full stop. Client names and private data. Contracts and legal wording. Final numbers, prices and anything that goes to HMRC / IRS / KRA or a client as fact. Your point of view — the two lines in any piece only you can write. If leaking it would cost you a client or a night's sleep, it doesn't go in the chat box.

3. Review — how you stay the editor. A weekly 15-minute spot-check: pick three outputs at random, check them against source, log what the tool got wrong, tighten the SOP. The log matters more than the check — after a month you'll know exactly where your AI is strong, where it hallucinates, and what to delegate next.

The practical move. Write the SOP tonight. Twenty minutes, one prompt, edit the brackets:

"I run [business type] for [audience]. Draft a one-page AI usage SOP with three sections: tasks AI may do unsupervised as first drafts, information it must never receive, and a weekly review checklist for me as the editor. Keep it under 300 words and practical."

Then print it — genuinely, physically print it — and keep it where you work. The SOP only works if it's a rule you can see, not a document you archived.

Why this beats "just be careful": rules you have to remember fail when you're busy, and you're always busy. A one-page SOP moves the judgement from every individual paste-decision (hundreds per week, each made when you're rushed) to one decision made once, calmly. That's what systems are for.

The responsible-AI note. Before the SOP can work, vet the tool it governs: read the data policy, switch model-training off, set retention limits. Disclose AI's involvement wherever a client would feel misled to learn of it later. And verify what ships — the model drafts, you decide. The SOP isn't bureaucracy; it's the price of being able to trust your own output at speed.

Where founders get this wrong (so you don't). Three failure modes come up again and again. First, the SOP that's really a wish list — ten sections, printed nowhere, followed never. One page is a feature: if it doesn't fit on one page, you've written policy, not practice. Second, the never-touch list that quietly erodes — it starts firm, then one deadline arrives and a client contract goes into the chat box "just this once". Decide now what your once-exception rule is (ours: never for the never-touch list; that's what makes it a list). Third, skipping the review log because the outputs "seem fine". Seeming fine is exactly what confident wrong answers look like; the log is how you find the 5% that isn't, before a client does.

What this looks like after a month. Week one, the SOP feels like admin. Week two, you stop re-deciding what's safe to paste, and the hesitation tax disappears. Week three, your review log shows a pattern — maybe dates are always right and figures sometimes aren't — so you tighten one line. Week four, delegating something new takes one glance at the page instead of a debate with yourself. That's the compounding: not the twenty minutes it saved, but the hundred small decisions it removed.

One decoded AI move, every week. Follow @aithebrief on Instagram for the daily version, or subscribe free on Substack for the deep one.