Can I trust it not to make mistakes? Where a human must stay in the loop

Which actions can AI take alone?

On September 29 OpenAI used its developer conference to introduce always-on agents that act without being asked (CNBC). Every major vendor now sells a version of this, so the useful question isn’t whether AI can act. It’s which actions you let it take alone.

Owners are already asking it. Accuracy tops every 2026 survey: 45% in the SBE Council survey in March, 36% in Simply Business, 32% in Bluehost in May and 31% in Bluevine in April. Cost barely registers.

That caution is sound: a language model drafts well, judges poorly, and finishes a sentence rather than stop. A sentence that holds a price, a diagnosis or a promise needs a person.

Three tiers: alone, approved, never

The tiers follow consequence, not difficulty: a booking link and a quote are equally easy for a model, but only the quote can cost a small firm a week’s margin.

The middle tier is the human in the loop, and it’s the one vendors skip. Square’s Managerbot, in open beta, gets it right: it proposes actions and the seller approves them first (QSR Magazine). Trades practitioners draw the same line on estimates: the model drafts the wording, the numbers need local rates and judgment (AI Gatecrashers).

The “never” row is written per business.

• A clinic: no clinical questions, and “Is this normal?” counts as one.
• A brokerage: no feature that isn’t on the MLS sheet (the boat slip that wasn’t there).
• A restaurant: no promise that an allergen is absent.
• Any office: a one-star review gets a human reply. 29% of owners named losing the human touch as their top fear (Bluehost).

Six checkpoints we build in

All of it runs in accounts in your name, on business AI plans that don’t train on your data (more in Who owns your automations when the consultant leaves?).

How do you stay covered when AI is wrong?

Treat what the assistant says as what you said, because tribunals do. In 2024 one ordered Air Canada to pay a customer about C$812 after its chatbot misstated the bereavement fare policy, and rejected the argument that the chatbot was a separate legal actor (IAPP).

In July a business owner sued OpenAI, alleging its chatbot’s guidance on a confidentiality agreement left her with weaker protections (Daily Journal). It’s an allegation, not a ruling, with the same lesson. The tiers are how you stay covered. The assistant answers only from your documents, and anything with a price, a promise or a legal position waits for a person.

Does approving eat the time saved?

Less than checking everything does. A Workday study, which we have only secondhand (The Lonely Entrepreneur), found nearly 40% of the time AI saves lost again to fixing and checking. A Forrester estimate cited by Tendem puts the checking at about 4.3 hours a week per worker. We call it the review tax.

Sorting first keeps it low. The top tier never needs a look, and middle-tier items arrive drafted, so an approval is a one-minute read and a tap. Set your tasks a week, the minutes each takes by hand and the share that needs a person’s okay, and compare the three designs.

What to ask a vendor

Buying an always-on agent instead of having one built? Sort it into tiers on the sales call, and ask about data too: security came right behind accuracy, at 42% in the SBE Council survey.

1. Which actions does it take without approval, and can I shorten that list?
2. Can it send money, change a price or alter a listing alone, and can I turn that off?
3. When it doesn’t know, does it say so and hand off?
4. Where is the log, who can read it, and how long is it kept?
5. Is there a switch I control that stops it, without a support ticket?
6. Is my data used to train the model, and is that in writing?

Write your own three-tier table before you talk to anyone; most owners fill the “never” row in five minutes. You can’t make AI perfect, but you can decide which mistakes it’s allowed to make alone.