What You Should Never Let AI Do — and How to Write That Down

We automate businesses for a living, so the next sentence is against interest: the most important decision in your AI operation is the list of things it will never be allowed to do.

For seven weeks this blog has argued the delegation side: find the constraint, hand off the recurring layer, demand work delivered finished. All of it stands. But every owner we talk to carries the same quiet question underneath, and it deserves a direct answer instead of reassurance. What if it answers a client on its own? What if it sends something I’d regret? The question isn’t paranoia. It’s the sound of someone who hasn’t been shown where the line is, because nobody wrote one.

The never list#

Here’s ours, the one we bring to every deployment. Five categories that stay human, whatever the tool can technically do.

Final say on anything that binds you. Signatures, prices, exceptions to your own terms, commitments of money or dates. A system can draft the quote in ninety seconds; a human owns the number. The day a discount goes out that nobody chose, you don’t have an efficiency, you have a liability.

Conversations where the relationship is the content. Bad news, conflict, condolences, a fee increase to a loyal client, anything a person will remember receiving. These can arrive drafted, and the draft genuinely helps. They are written-through, decided and sent by a human, every time.

Judgment calls without a definition of done. Hiring, firing, strategy, whether to bend a rule for a good client. The test hasn’t changed since the delegation post: if you couldn’t hand it to a competent new hire on day one, don’t hand it to a system on day one hundred.

The case the system hasn’t seen. Exceptions route to a human by design, not by luck. A system that guesses on edge cases isn’t autonomous; it’s unsupervised. The confidence to let routine work run comes precisely from knowing the unusual stops and asks.

Anything you couldn’t explain afterward. If nobody can see what went out, when, and why, it doesn’t run. Every action logged, every send traceable. This one protects all the others, because a line you can’t audit is a line you can’t enforce.

Write it down, one page, three verbs#

An unwritten rule exists only in people’s heads. It erodes the moment deadlines close in. What holds is a page anyone can read in two minutes. Ours is organized around verbs; that’s the model we propose for yours.

Reads. What the system can see: the inbox, the calendar, the client files it needs for context. Naming this explicitly also names what it can’t see, which is where your data-privacy worries stop being vague.

Drafts. The work that arrives prepared and waits for a human: sensitive communications, quotes, anything on the never list’s edge. The rule at Académie Lavalloise, the 250-student school that runs on Avelle, is one sentence long: Avelle drafts, a human sends. That single sentence, written and known by everyone, is why trust holds there. Nobody wonders. It’s on the page.

Delivers. The recurring layer that arrives finished and logged: confirmations, reminders, reports, the routine answers you’ve given a hundred times. This list starts small and grows. Lines move down the page as validation earns its way to autonomy, and that movement is a decision you make in review, never a drift that happens to you.

Then the fourth section, the shortest and the one that does the most work: Never. The five categories above, in your own words, with your own examples.

Why the page matters more than the tool#

Written limits are what make delegation safe to say yes to, and they compound in three directions at once. Your team trusts the system because they know what it will never do to a client relationship they’ve spent years building. You sleep because the 11 p.m. what-if has a written answer. And any provider you evaluate can be handed the page and asked: can you operate inside this? Nobody actually wants AI; you want finished work inside limits you chose. A vendor who pushes back on your never list has already told you everything.

Notice what this page is not. It’s not a brake on automation; it’s the reason the automation gets to run at full speed everywhere else. The school doesn’t send fewer communications because a human signs the sensitive ones. It sends more, faster, because nobody’s afraid of the machine anymore.

Write yours this week#

Twenty minutes, four headings: reads, drafts, delivers, never. Fill the never section first, because it’s the one your gut already knows. If you want a second pair of eyes, a thirty-minute conversation is enough for us to tell you where your lines sit compared to what we run every day, and which ones we’d move.

You don’t need more confidence in AI. You need your limits in writing, and then the confidence takes care of itself.

---

Matthew Lacopo leads client engagement and content at The Agensee, a managed AI operations firm based in Montréal serving SMBs across North America, which he cofounded with Anthony Lacopo.