What should you never let an AI agent decide?

Why those five specifically?

Because they are the decisions where being wrong is expensive and cannot be undone by trying again. Most agent mistakes are cheap: a bad draft, an odd summary, a wasted follow-up. These five are the ones that create obligations, cost money you cannot recover, or change what your company is.

Taking them one at a time:

Irreversible. Deletions, refunds, cancellations, anything sent to a customer that cannot be unsent. The test is simple: if undoing it requires someone else’s cooperation, it is irreversible for this purpose.

Legally binding. Contract terms, delivery commitments, guarantees, anything a customer could hold you to. Agents produce these accidentally and fluently, usually while being helpful.

Pricing. What you charge and what you discount is the clearest expression of a business’s strategy. It is also the thing an agent will happily concede to close something.

Positioning. What the company says it is. Every published claim is a promise somebody may rely on, and the cost of a wrong one is not a bad paragraph, it is a claim you have to defend.

Spend above a threshold. Advertising in particular. Automated spending is the fastest way for a small mistake to become a large one overnight.

What is safe to fully automate?

Work that is reversible, bounded and checkable. Drafting, ranking, routing, reconciling, summarizing, first-line answers drawn from your own documentation. If the worst outcome is that you redo it, automate it. If the worst outcome is a customer holding you to something, do not.

Notice that this list contains most of the actual hours in a small business. Being strict about the five categories costs you very little throughput, which is why the strictness is cheap and worth keeping.

There is a second category worth automating even though it feels risky: anything where a human would be slower and no better. Scoring four hundred leads, or reconciling a bank feed, is not a judgment call. Doing it by hand is not caution, it is just slower.

How should you set the spend threshold?

At the amount that would make you look twice on a bank statement. Not a round number, and not the largest sum you could tolerate losing — the smallest sum you would want to know about. You can always raise it once you have watched the system for a month.

Founders usually set it too high at first, because a low threshold feels like friction. The friction is the product. A threshold that stops three things a week is working; one that never stops anything is not protecting you from anything either.

Review it after a month of real operation. By then you will know which interruptions were useful and which were noise, and you can tune it with evidence rather than nerves.

What about decisions that are reversible but embarrassing?

Treat published content as a sixth category. A bad article is technically reversible and practically is not, because it was seen, indexed and possibly shared before you noticed. Anything going out under the company name is worth a person reading, even when the risk is only reputational.

The same reasoning applies to outbound at volume. A single bad email is recoverable. Four hundred of them, sent overnight to your entire prospect list, are not — the recipients have already formed a view, and some of them mark it as spam, which affects everything you send afterwards.

The useful test is not whether an action can be undone but whether undoing it restores the position. Deleting a post removes the post. It does not remove the impression, the screenshot, or the search result that has already been cached.

Everything else genuinely reversible — a draft, a ranking, an internal summary, a routing decision — can run without review, and should. Being strict about six categories works precisely because it leaves the other ninety percent of the work alone.

How this works on newc0

These five are the default approval gate, and the spend threshold is set by the founder. Agents draft, queue and wait; they do not send, spend, promise or delete on their own. The items that stop are what fills the morning review, which on an ordinary day is about fifteen minutes.