Key takeaways
- Use AI for work that is repetitive, written and checkable, such as first drafts of processes, weekly number summaries, first-pass categorisation and repeat customer replies, and never for anything that carries the business's liability unread.
- Keep 3 rules: a person reads everything before it is sent, nothing confidential goes into a tool the business does not control, and the model follows a written process.
- Automate process documentation first: record the person doing the work, have a model draft the process from the recording, and have that person correct it.
- Measure every use, so corrections to drafted replies fall and the weekly reconciliation of categorised transactions gets faster.
- Keep the written process as the source of truth and the tool as its instrument, so a tool update cannot quietly change how the business behaves.
Using AI in a small business means giving a model the business's written processes and letting it draft, sort, and summarise under a person's review, never letting it decide.
Where does AI actually help in a small business?
In work that is repetitive, written, and checkable. A new owner inherits a business that runs on the previous owner's memory, and the fastest way to get that memory onto paper is to talk it out and have a model turn the transcript into a first draft of the process. The same is true of the customer reply that says the same thing 40 times a week, the summary of the week's numbers, the first pass at categorising bank transactions, and the job advertisement that has to be written by Friday.
None of that is judgement. It is drafting, sorting, and summarising, which is what the current models do well and what a small business owner does too much of. The Run guide is where AI sits inside documented, checkable processes, and the operating guide places it after the reading period, once the owner knows what the business's real processes are, because a model can only document a process somebody can describe.
Where must AI not be used?
In anything that carries the business's liability without a person reading it first. Payroll calculations, tax positions, safety procedures, contracts, regulated communications, and any figure that goes to a lender or a buyer. The model does not know what it does not know, it states wrong answers in the same voice as right ones, and it is the owner's name on the document. The Federal Trade Commission has said plainly that claims about what AI does, and outputs that deceive, are the business's responsibility.
The line is not the tool. It is the reader. Every output that leaves the building has a person who read it and would sign it.
What are the three rules?
1. A person reads everything before it is sent. Not skims. Reads. The time saved is in the drafting, and a business that saves that time by not reading is spending it on the customer who received the wrong reply.
2. Nothing confidential goes into a tool the business does not control. Customer lists, financial statements, employee records, and anything covered by a contract stay out of consumer tools whose terms allow the provider to keep the input. Read the data terms of any tool before the first paste, and prefer tools that state the input is not retained or used for training.
3. The process is written down and the model follows it. A model asked to "handle customer emails" will invent a policy. A model given the business's written refund policy, its tone, and 5 examples of good replies will draft replies that match. The written process is the asset; the model is the clerk.
What should a new owner automate first?
The process documentation itself, because everything else depends on it. Walk through each process with the person who does it, record the walk, and have a model produce the first written version. Then the person corrects it. A business that ran on one owner's memory has, in a month, a set of documents a new hire can follow, which is what the Systemizing yourself out course teaches as the route to a business that runs without its owner.
Second, the weekly numbers. A model can turn an export from the accounting system into a plain summary: cash, receipts, payments, and the three things that moved. The owner still reads the export; the summary is what makes the reading fast enough to happen every week.
Third, the writing that is the same every time. Replies to the 10 questions customers always ask, the first draft of a job description, the note to a supplier about a late delivery. Written once as a process, with examples, and drafted by the model from then on.
How do you keep AI from making the business worse?
By measuring what it touches. A model drafting customer replies should be paired with a count of replies that had to be corrected, and the count should fall. A model categorising transactions should be reconciled against the bank statement every week, and the reconciliation should take less time each month. If a use of AI cannot be measured, it is a use the owner cannot tell is working, and a new owner has enough of those already.
The larger risk is the one the previous owner never had: a process that used to live in a person's head now lives in a prompt nobody reviewed, and the business's behaviour changes when the tool updates. Keep the written process as the source of truth and the tool as its instrument, and review the outputs on a schedule rather than when something goes wrong.
What does this look like on a real business?
On the business behind this site, the first 90 days were spent reading, not automating. The processes that were written down after that period were the ones that had been costing the owner hours, and the writing was the work. The tools came later, once there was a process to give them, and the measure of each one was whether the owner's week got shorter without the customers noticing anything. The first ninety days lesson is where that order comes from, and it is the order this page recommends.
Questions people ask next
Which AI tool should a small business use?
The one whose data terms the owner has read. The tool matters less than the rule that nothing confidential goes into a tool the business does not control, and that a person reads every output before it leaves the building.
Can AI do the bookkeeping?
It can take a first pass at categorising transactions. A person reconciles against the bank statement every week, and the reconciliation, not the model, is the bookkeeping.
Should I tell customers a reply was drafted by AI?
The reply is the business's reply, read and sent by a person, so it is the business's responsibility regardless. The Federal Trade Commission holds businesses to their claims about AI and to what their outputs say.
What should I automate first?
Process documentation, because everything else depends on it. Record the walkthrough, let the model draft the document, and have the person who does the work correct it.
Sources
- Federal Trade Commission, keep your AI claims in check Primary source. Read 2026-09-08.
- National Institute of Standards and Technology, AI Risk Management Framework Primary source. Read 2026-09-08.
- Official Business Doctor, the first ninety days after you own it Secondary source. Read 2026-09-08.
Dr. Matt Herrera
Dr. Matt Herrera is the founder of Official Business Doctor. He writes and teaches from direct experience buying a business with cash, financing one with an SBA 7(a) loan, founding one, operating two locations with five retained employees, improving the operation, and later exiting at more than twice his equity.