Skip to content

Writing an AI use policy for a small service business

One page, eight questions, written for your team rather than a regulator. The template, the three clauses people write without having the controls to back them, and the review that keeps it true.

Sohail Safdar · · 6 min read

An AI use policy for a small business is one page and it is written for your team, not for a regulator.

That framing decides everything about it. The front desk needs to know what the assistant handles, so they do not duplicate it or contradict it in the next message. The owner needs to know who is accountable when it says something wrong, before it does. And a new hire needs to learn both in five minutes rather than by asking three people and getting two answers.

Writing it takes an hour. It makes every later decision faster.

The template, eight clauses

1. Purpose.

"Our AI assistant answers routine customer questions by text and web chat from the facts we approve, so our team can focus on the customers in front of them."

One sentence. What it is for, bounded.

2. What it may answer.

"Services, prices from our published list, hours, locations, policies, and how to book. Only from facts and FAQs approved by the practice manager."

The second sentence is the load-bearing one, and see the warning below about writing it if it is not true.

3. What it must hand to a person.

"Anything clinical, legal or financial. Complaints. Emergencies. Anything about another customer. Any price that depends on seeing the job. Anybody who asks for a person. These go to the duty manager immediately."

This is the clause worth spending the most time on, because it is the only one that prevents something. Write it for your own trade, specifically, and keep adding to it: a real conversation that surprised you is the best evidence you will ever get that this list is incomplete.

4. Supervision.

"For the first two weeks every reply is approved by Jo before it sends. After that Jo reads twenty conversations a week and reviews the unanswered-questions log monthly."

A named person and a number. "Regularly reviewed" is not a supervision clause.

5. Disclosure.

"The assistant introduces itself as our assistant. If asked whether it is a person, it says no and offers one."

The second half is the part that matters and the part most often missing. An assistant that evades the question is an assistant that has misled somebody who asked directly.

6. Data.

"It sees only the customer it is talking to. Customers are never asked for card numbers, Social Security numbers or medical details by text."

If your tool lets you choose how much customer history the assistant can read, say which setting you chose. The narrow one should be the default and the wide one should be a decision somebody made.

7. Mistakes.

"If it says something wrong, the practice manager corrects it with the customer the same day and fixes the underlying fact. Serious mistakes go to the owner and to the vendor."

Two separate actions in there, deliberately. Correcting the customer and fixing the fact are different jobs and doing only the first guarantees a repeat.

8. Owner.

"Jo owns this policy and reviews it every six months."

The three clauses people write without the controls to back them

This is the failure worth naming, because it is the one that makes a policy actively harmful.

"It only answers from approved facts." If your tool will answer anything it is asked, this sentence is false. Either choose a tool that can be constrained to a knowledge base you edit, or write what yours actually does.

"Sensitive questions are escalated before an answer is attempted." Check whether your tool refuses before answering or answers with a disclaimer. Those look similar in a transcript and are different things, and only one of them matches this sentence.

"Nothing sends without review." True in approval mode, false the moment somebody switches one agent to autonomous and does not update the policy. If you run a mix, say which agents are which.

A policy describing controls you do not have is worse than no policy, because it is a written claim about a safeguard, and it was presumably shown to somebody.

What a small business does not need

A page on the ethics of artificial intelligence. A statement of principles. A section about bias in training data that nobody in the building can act on.

Not because those are unimportant in general, but because an internal policy is a document somebody reads at the moment they are deciding something, and every paragraph that cannot be acted on makes the clauses that can be less likely to be read.

If you want a one-line version of the ethics: the assistant may answer, it may not advise, and it may never be the reason somebody did not reach a person.

Keeping it true

Three triggers for an edit, in order of how often they fire:

A conversation that surprised you. Add it to clause 3 the same day, while you still have the words the customer used. This is the one that keeps the policy alive.

A change of tool, channel or mode. A new channel is a new place every clause has to hold. A move to autonomous changes clause 4.

Every six months, by the named owner in clause 8, whether or not anything obvious has changed.

In Frontiva

The AI agents configuration maps onto this template closely enough to be worth listing against it, which also means the policy can describe real controls rather than intentions.

Clause 2 is the knowledge base: structured facts and approved FAQs, and a question matching nothing is logged as a gap rather than answered. Clause 3 is your own escalation keywords plus a sensitive-topic check that runs before the answering path as a deterministic pattern check, which is what makes the "before an answer is attempted" wording true rather than aspirational.

Clause 4 is the action mode, per agent, with approval as the default, plus the readiness suite that keeps unsupervised answering locked until it passes. Clause 5 is a handover message you write yourself rather than one that is generated, and it is sent in approval mode too, because somebody asking for a human should not watch an empty thread while a draft waits in an inbox.

Clause 6 is a memory setting with three choices (only the message being answered, this conversation, or this customer's earlier conversations), defaulting to the middle one, which reads nothing the customer is not already looking at.

Autonomous vs approval mode covers clause 4's decision, and what an AI receptionist should refuse to answer is the longer version of clause 3.

Frequently asked questions

Does a small business need a written AI use policy?

Not for a regulator, in most cases, and yes for your own team. The front desk needs to know what the assistant handles so they do not duplicate it or contradict it. The owner needs to know who is accountable when it says something wrong. And a new hire needs to learn both in five minutes rather than by asking three people. It is one page and it makes every later decision faster.

What should an AI use policy contain?

Eight things: what the AI is for, what it may answer, what it must hand to a person, who supervises it and how, how customers are told they are talking to an assistant, what customer data it can see, what happens when it gets something wrong, and who owns the policy. Each one a sentence or two. A longer document is a document nobody consults at the moment they need it.

What is the most common mistake in one of these policies?

Describing controls you do not have. A policy saying the assistant answers only from approved facts, when your tool will answer anything, is worse than no policy: it is a written statement of a safeguard that does not exist, and it was presumably shown to somebody. Either choose a tool that can do what the policy says, or write what your tool actually does.

Do customers need to see the AI use policy?

No, and one clause from it should be visible: the disclosure that they are talking to an assistant. The rest is internal, in the way a cash-handling procedure is internal. If you are in a regulated profession it is worth having counsel read the hand-off, disclosure and data clauses specifically, because those are the three with external consequences.

How often should it be reviewed?

When you change tools, add a channel, or meet a new kind of question the AI should not answer, and otherwise every six months with a named owner. The trigger that matters most is the third one: a real conversation that surprised you is the best evidence you will get about where the hand-off list is incomplete, and it is worth adding on the day rather than at the review.

Get the next one by email

Email me when Frontiva publishes a new guide or post. At most one a week, nothing else, and an unsubscribe link in every one.

To keep spam out, we record your IP address, approximate location and browser with your message. Privacy Policy

Start free. Be live this week.

Built for dental, med spa, home services, and the businesses that live on the phone.