Why I don't say "AI can handle all of it"

When people ask me about putting AI into their business, one question comes up more than any other. Can it just handle all of it?

I understand the appeal. But I don’t say yes, and I want to explain why properly rather than hiding behind caution.

The incidents cluster in one place

Nearly every AI failure I have seen or been called in after has the same shape. It happened in the part that was handed over completely — where output went to a client, or into an invoice, or onto a public page, without a person reading it.

Not because the model was bad. The output was usually plausible. That’s the problem: plausible is the failure mode. Something that reads well and is wrong will pass any check that consists of glancing at it.

If nobody was assigned to look, nobody looked.

A gate is not distrust of the tool

People sometimes hear “keep a human gate” as a lack of confidence in AI. It isn’t. It’s the same reason a second person signs off a payment run, or a pilot runs a checklist they know by heart.

The gate exists because the cost of the two outcomes is not symmetric. Ninety-nine drafts going out slightly faster is worth very little. One wrong figure reaching a client costs you the money, the rework, and some part of the relationship.

When the downside is that lopsided, you don’t optimise for speed at the gate. You put the gate where the damage would land.

Where the gate goes

Not everywhere. A gate on every step is just doing the work yourself with extra ceremony. In practice it belongs at three points:

• Anything that leaves the building. Client-facing documents, quotes, published pages, anything sent under your name.
• Anything involving money. Rates, totals, margin, refunds. See the separate note on why AI should never produce a figure.
• Anything irreversible. Deletions, sends, payments, publishing. Actions you cannot quietly undo.

Everything else — drafting, restructuring, formatting, summarising, checking — can run without you standing over it.

Write the scope down, or it will drift

The part most people skip: the scope a human holds has to be written down.

If it lives in your head, it erodes. A busy week arrives, one approval gets skipped because the last forty were fine, and the boundary moves without anyone deciding to move it. Six weeks later nobody is quite sure what is being checked.

So I put it in a document, in plain language, as a list. These specific things are decided by a person. Everything else the system may do on its own. It reads as bureaucratic and it takes ten minutes, and it is the single thing that keeps a setup working after the novelty wears off.

It also does something useful commercially. When a client asks what happens if the AI gets it wrong, you have an answer that isn’t a promise — it’s a design.

What I say instead

So when someone asks whether AI can handle all of it, my answer is roughly this:

Most of the work, yes. The volume, the first drafts, the repetitive assembly, and — usefully — a lot of the checking. But the decision to release, and the numbers that turn into money, stay with you.

That isn’t a smaller offer. It’s the part that makes the rest safe to use.

---

Guardrail Ops — Hirotoshi Yamaguchi, Gold Coast QLD. Get in touch