Field Notes · 06

What a small business should hand to AI, and what it should keep

A working checklist for owners. Five kinds of work a machine can take today, four it should stay away from, and the questions to ask before any customer data leaves the building.

Patrick Jean-Baptiste · September 14, 2026

Most advice about AI for small business falls into two piles. One says it will transform everything by Friday. The other says it is a fad. Owners we talk to are tired of both and want a usable answer to a plain question: what, specifically, should I do with this?

Here is the list we work from. It comes from running the back office of a real practice and from fifteen years of cleaning up after software that promised too much.

Hand these over

Retyping. If a person reads information in one place and keys it into another, a machine can do the first pass. Emailed work orders into a job sheet. Web form submissions into a customer record. Receipts into a ledger. Have a person review the exceptions and spot check the rest.

First drafts of routine messages. Appointment reminders, quote follow ups, review requests, answers to the nine questions everyone asks. The machine drafts in your voice from examples you supply. A person sends.

Sorting the inbox. Reading each incoming message and labeling it: new lead, existing customer, vendor, invoice, junk. This alone gives some owners an hour a day back.

Summaries. A long email thread, a forty page contract, a month of customer comments, reduced to a page with the source passages marked so you can verify them.

The weekly report nobody has time to build. Pulling numbers from two or three systems into one page every Monday morning.

Notice the pattern. Each job is repetitive, the inputs are text, a mistake is cheap to catch, and a person sees the result before it matters.

Keep these

Anything where the relationship is the product. The call after a job goes wrong. The conversation with a family in a hard season. Customers forgive a slow reply from a person sooner than a fast reply from a machine that misses the point.

Final say on money. Let a machine prepare the invoice, flag the late account, draft the estimate. A person approves anything that charges, refunds, or commits you.

Professional judgment. Diagnosis, legal advice, structural decisions, tax positions. A model can help a licensed person work faster. It cannot hold the license or the liability.

Facts you cannot check. These systems sometimes state false things with complete confidence. If nobody on your team could tell whether an answer is wrong, that task is a poor fit.

Before any customer data goes anywhere

Ask these of any tool or any person proposing one, including us.

1. Which company's servers will see this information?
2. Is it used to train their models? Where is that written down?
3. How long is it kept, and can it be deleted on request?
4. If the data is regulated, such as health, financial, or student records, will the vendor sign the agreement your regulator expects?
5. What happens to the workflow if this vendor doubles its price or shuts down?

A vendor who answers all five quickly and in writing is worth talking to. Hesitation on the second question is a reason to walk away.

How to begin

Pick one job from the first list. Choose the one your staff complains about most, since they will help you get it right. Run the machine alongside the person for a few weeks and compare results. Expand only when the comparison bores you.

That pace will feel slow next to the headlines. It is the pace at which things get built that still work in five years, which is the only kind worth paying for.