‹ back to writing
Nov 20, 2026·6 min read
systemssmall-business

Which AI tasks can a small business stop checking?

Stop checking AI work that stays inside the business and can be undone: internal notes, summaries, first drafts. Keep a human on anything a customer can act on, anything that moves money, and anything you can't take back. In Gusto's 2026 survey of 1,731 small-business owners, about 60% of AI users already let it draft marketing with limited review, yet only 25% of businesses with employees have any AI policy.

A lot of small jobs in my week are done by AI now. First drafts. The nightly log of what ran and what broke. Notes filed into the right folder. A daily post slot.

For the first few weeks I read every line of all of it. The nightly log got the same attention as a post going out under my name.

That was a waste. Most of it I could have skimmed, or not read at all. But two kinds of output still get checked every single time before they go anywhere, and I've never regretted it.

what my AI does · how closely I check
internal note filed · stays inside · undoable → skim
nightly run log · stays inside · undoable → skim
first draft of an essay · stays inside · undoable → edit anyway
post under my name · leaves · public → checked before it goes
message to someone new · leaves · can't unsend → my yes, every time
// The check level was never about how good the AI is. It was about where the output goes.

That is the whole essay. You don't decide what to stop checking by asking how smart the tool is. You decide by asking who sees the output, and whether you can take it back.

Which AI tasks can a small business stop checking?

The ones that stay inside the business and can be undone. Internal notes, meeting summaries, research digests, first drafts a person will edit anyway: spot-check these, don't read every word.

Keep a human on anything a customer can act on, anything that moves money, and anything you can't take back. That line has almost nothing to do with which AI tool you use.

What are small businesses already letting AI do without review?

More than you'd guess, and mostly the right things.

Gusto's State of Small Business 2026, published on 12 November, surveyed 1,731 owners of US businesses with 24 or fewer employees. 67% now use AI for at least some tasks. They asked owners which tasks they'd be comfortable letting AI handle with limited human review.

Owners who already use AI were far more relaxed than owners who don't:

Comfortable with AI doing it on limited review · AI-using owners
  • Drafting marketing content
    0%
  • Writing emails
    0%
  • Scheduling
    0%
  • Bookkeeping
    0%
// Non-adopters: 27%, 20%, 8% and 6%. Source: Gusto State of Small Business 2026.

Gusto found owners "drew a clear line between communication and judgment-heavy work": most comfortable with AI drafting marketing and routine customer emails, far less comfortable letting it do data analysis or answer customer questions directly.

And it's not a niche habit. Bank of America's 2026 Business Owner Report, released on 17 November (an Ipsos survey of 849 small and 250 mid-sized US business owners), found 84% use AI in some capacity. Among users, 52% use it for marketing and 31% for customer service.

So almost a third of AI-using owners already point it at customers. That's exactly where the line matters most.

Why is a customer-facing AI answer different?

Because the customer acts on it, and the answer is yours.

The clearest example is a tribunal case from 2024. In Moffatt v. Air Canada, the airline's website chatbot told a grieving customer they could apply for a bereavement fare after travelling. The airline's own bereavement page said the opposite. The customer booked on the chatbot's word, the airline refused the reduced fare, and the dispute went to British Columbia's Civil Resolution Tribunal.

Air Canada argued it couldn't be held liable for what its chatbot said. The tribunal called that "a remarkable submission" and wrote: "It makes no difference whether the information comes from a static page or a chatbot." The airline was ordered to pay $812.02.

That's a big company and a small sum. For a small business, the money is rarely the problem. The problem is that the customer already acted on a wrong answer before anyone at the business read it.

If a customer can act on it, or you can't take it back, a human reads it first.

What is a review tier?

A review tier is how closely a person checks a given AI task before its output counts. Three are enough for most small teams:

The tier belongs to the task, not the tool. The same model can write your internal notes and your refund replies. Only one of those should be on skim.

How do you decide which tier a task belongs in?

Ask two questions, in order. Then write the answer down.

Sort one AI task in under a minute0 / 5
12345
  1. 1
    Name the taskOne line: what the AI does and where its output goes
  2. 2
    Does it leave the building?A customer, supplier or the public will see it or act on it
  3. 3
    Can you take it back?Money moved, a promise made, a message sent: usually no
  4. 4
    Pick the tierStays inside and undoable: skim. Leaves: check before it leaves. Can't undo: human decides.
  5. 5
    Review in a monthMove a task down a tier only after a month of checks with nothing to fix
// Step 3 is the one people skip. It's the one that decides whether a mistake is a typo or a refund.

Two tasks look alike and land in different tiers all the time. "Summarise yesterday's enquiries for me" is skim. "Reply to yesterday's enquiries" is check before it leaves. Same inbox, same AI, different tier.

Do you need an AI policy for this?

You need one page, and this list is most of it.

In Gusto's survey, only 25% of small businesses with employees have any kind of AI policy, up from 15% a year earlier. 47% have none and no plans to write one.

A long policy is overkill for a ten-person business. But a list of every AI task you run, with its tier next to it, is not a policy document. It's the answer to the question a new hire will ask on day one: "Do I need to check this before it goes out?"

I wrote about the agent-side version of this in The agent needs a yes: anything public, irreversible or expensive waits for a human. The review tier is the same rule, written for the whole team instead of one agent.

So which tasks can you stop checking?

The ones nobody outside the business will ever see, and that you could quietly fix tomorrow.

Everything else keeps a human in the loop. Not because the AI is bad, but because the customer won't know the difference between your AI's answer and yours.

List every task AI does in your business this week. Next to each one, write skim, check, or decide.

Related essays
More essays

Owned, undistributed, and published only here — no newsletter, no funnel, no upsell.

All essays ›
‹ all essays

Let's build
something joyful.

Agents · Engineering · Talks · Podcasts · Interviews · Automations.