Write a Job Description for Your AI Assistant: What It Does, What It Never Does, and Who Checks It

A furniture shop turns on an AI feature in its chat tool with the instruction "answer customers politely". In the first week it promises a customer free delivery to another island, quotes a price from an old catalogue and tells someone the shop is open on a public holiday. The owner turns it off and concludes that AI does not work. A new employee given the same single instruction, no price list and no supervisor would have done about as badly. An AI assistant needs what any new hire needs: a role, sources, limits and someone who checks.

Why a job description

• It forces a narrow scope. "Help with everything" fails. "Draft replies to stock and delivery questions" can succeed.
• It names the sources. The assistant answers from what you give it.
• It sets limits in advance, before a mistake teaches you.
• It gives staff clarity. People know what the assistant does and what remains theirs.
• It makes review possible. You can only judge performance against a defined job.

The one-page template

1. Role in one sentence

"Drafts first replies to incoming customer messages about products, stock, delivery and opening hours, for a staff member to approve."

2. Duties

Specific and countable.

• Draft a reply to each new enquiry within the scope.
• Summarise conversations longer than twenty messages for the staff member taking over.
• Label each conversation by topic.
• Suggest a follow-up date when a customer says "I will think about it".

3. What it never does

The most important section.

• Never states a price that is not in the price list.
• Never offers a discount, free delivery or a deadline.
• Never confirms a refund, return or compensation.
• Never gives medical, legal or financial advice.
• Never asks for card details, passwords or verification codes.
• Never pretends to be a human when asked directly.
• Never replies to complaints or angry messages: it hands over.

4. Sources it may use

• Current price list, with the date.
• Product information and stock, with how often it is updated.
• Delivery areas, costs and times.
• Opening hours and holiday closures.
• Policies: returns, warranty, payment.

If the answer is not in these, the assistant says it will check and hands over. An assistant with no source for a question should not improvise one.

5. Tone

Three lines. How you address customers, how long replies are, which words you use or avoid. Add five real replies from your best staff member as examples.

6. When it hands over to a person

• The customer asks for a person.
• The message is a complaint or shows frustration.
• The question is outside the sources.
• The customer mentions a large order, a custom request or a business purchase.
• Money has already been paid and something is wrong.
• The same question has been asked twice.

And to whom: a named role, within a stated time.

7. Level of autonomy

• Draft only: every reply is approved by a person.
• Auto for a list: it sends without approval only for named question types, such as opening hours and location.
• Outside working hours: an acknowledgment that says when a person will reply, with no promises.

Start with draft only.

8. Who supervises

One named person owns the assistant: updates its sources, reviews its work, decides changes to autonomy.

9. How it is measured

• Share of drafts sent unchanged.
• Wrong facts found in review.
• Time to first reply.
• Conversations handed over, and why.

A worked example

Role: Drafts first replies for the sales inbox of a furniture shop.

Duties: Draft answers about product dimensions, materials, stock and delivery areas. Label conversations. Suggest follow-up dates.

Never: Quote prices not in the list dated this month. Promise delivery dates. Offer discounts. Answer complaints. Discuss custom orders beyond collecting the request.

Sources: Price list, product sheet, delivery table, opening hours, return policy.

Hand over: Custom orders, orders above a set value, any complaint, any payment issue, any question not in the sources. To the sales lead, within working hours.

Autonomy: Draft only for the first month. Then automatic for opening hours, address and catalogue requests.

Supervisor: The sales lead. Review every Friday.

That is one page. It took an hour to write and prevented all three of the mistakes in the opening story.

The weekly review

Thirty minutes.

1. Read twenty drafts, including some that were edited or discarded.
2. Ask why each was changed. Wrong fact, wrong tone, missing information, or outside scope?
3. Fix the cause. Usually the sources are out of date or a rule is missing.
4. Add good examples of replies you approved.
5. Check the handovers. Were customers passed to a person at the right moment?
6. Decide on autonomy. Widen only for a question type that has been right for weeks.

An assistant improves when its sources and instructions improve. It does not learn your business by being left alone.

Being honest with customers

• Do not present the assistant as a person. If a customer asks whether they are speaking to a human, tell the truth.
• Make it easy to reach a person.
• Take responsibility for what it says. If the assistant gave wrong information, the business said it. Correct it and make it right.
• Mind the data. Do not feed it more customer information than the task needs, and follow the privacy rules that apply to you.

Common mistakes

• One vague instruction and no sources.
• Full automation on day one.
• Outdated price lists as the source.
• No list of forbidden actions.
• Letting it answer complaints.
• Nobody assigned to supervise.
• Judging it by impressions, not a weekly sample.
• Hiding that it is an AI.
• Expanding its scope after one good week.

Putting it to work

The job description lives wherever your assistant is configured. In Polynusa, AI drafts replies in the shared inbox for staff to approve, so the draft-only stage is the default way of working, and conversations, notes and follow-up tasks stay with the customer record. See the pricing page.

Summary

Treat an AI assistant as a new hire with a narrow role. Write one page that states its duties, the things it never does, the sources it may use, when it hands over, its level of autonomy, its supervisor and its measures. Start with drafts approved by people, review a sample every week, fix sources and rules based on what you find, widen autonomy slowly, and be honest with customers that an assistant is involved.

FAQ

What tasks suit an AI assistant in a small business?

Tasks that are frequent, follow a pattern and have a clear source of truth: drafting replies to common questions, summarising long conversations, sorting incoming messages by topic, preparing follow-up reminders and pulling details from messages into records. Tasks that need judgement about money, exceptions or upset customers should stay with people.

Should an AI assistant reply to customers on its own?

Start with drafts that a person approves. Move to automatic replies only for narrow question types where the answers have been right consistently over several weeks, the source information is reliable and a mistake would be cheap. Keep anything involving price changes, complaints, refunds or promises under human approval.

How do we know whether the AI assistant is doing a good job?

Review a sample each week. Count how many drafts were sent unchanged, edited or discarded, read the ones that were edited to see why, and note any wrong facts. Track response time and how many conversations were handed to a person. Then update its instructions and sources based on what you find.