Frisk Digital
← All articles

AI office assistants

What an AI office assistant should be allowed to do

Set clear limits around answers, information collection and handoff before putting an AI assistant in front of customers.

Published November 8, 2026 · Frisk Digital

Before choosing what an AI office assistant should say, decide what it is allowed to do. In a small business, a casual sentence can sound like a commitment. “We can fit you in Friday” affects the diary. “That should cost about this much” affects a customer's expectations before anyone has seen the work.

Useful boundaries start with the decisions your people need to keep. Write those down first. Then identify the information the assistant can safely provide and the routine details it can collect for review.

Separate information from decisions

An assistant may be able to explain published service information, describe the usual inquiry process and ask relevant intake questions. Those tasks rely on information the business has approved. They still need checking when services, locations or working hours change.

Accepting a job, setting a price, promising attendance or deciding an exception calls for authority. If the owner controls those decisions, the assistant should say that the team needs to confirm them. Avoid soft promises such as “That should be fine” when no one has checked availability or scope.

Write down the things it must not do

Make the limits specific enough to test. “Be sensible” is too vague. A more useful instruction is that the assistant must not confirm appointments, create quotations or send emails unless those actions are explicitly included in an approved setup with appropriate controls.

Other limits depend on the business. An assistant should not advise a customer to attempt dangerous electrical, gas or structural work. It should not invent a service because a customer asks for it. It should not reveal another customer's details or give internal notes to someone who claims to know the owner.

You also need rules for complaints, disputes and requests involving personal information. Decide where these go and who can respond. An intake assistant can record that a person needs help without trying to settle the issue itself.

Collect only what the next person needs

A new inquiry rarely needs identity documents, payment-card information or account passwords. Keep intake questions tied to the task. A description of the work, location and suitable contact details may be enough to start.

Explain why information is being requested and what happens to it. Give the customer an opportunity to correct a summary before it is treated as their brief. Review and storage are separate steps, and the wording should make that distinction clear.

For example, Frisk's live Dani assistant gathers a website brief for review before it is saved. It does not book appointments, quote prices or send emails. Those limits help the customer understand what the conversation accomplishes.

Decide when a person takes over

Useful handoff triggers include uncertainty about the service, a request for a firm price, a complaint, conflicting information or a customer simply asking for a person. Do not make the assistant keep asking questions when it has reached the edge of its role.

The handoff should preserve the useful context. A concise account of the request, confirmed details and unanswered questions gives the next person a place to start. The customer should not have to argue with the assistant to reach a normal contact option.

Test the boundaries with awkward requests

Normal questions show whether the assistant sounds pleasant. Awkward questions show whether its limits work. Try asking for a guaranteed start date, a discount the owner never approved, another customer's contact details or instructions outside the business's services.

Also test innocent ambiguity. A customer may say “Book me in” when they mean “Please call me about this”. The assistant should clarify the request without treating the words as permission to make a booking. Look for language that suggests an action happened when it only collected information.

Keep the setup under review

Nominate someone to own the approved information and review examples of failed or confusing conversations. Change the source material when the business changes. If a new integration is proposed, examine the action, data involved, permissions and failure handling before switching it on.

Frisk's custom AI office assistant work is intended to support staff. Any connection to booking, email or other business systems needs to be scoped separately. A useful assistant makes the first part of a conversation clearer while leaving the owner and team in control of the decisions they are responsible for.

Talk through what needs work.

See our custom AI office assistant service

Start your website brief with Dani →

Call Carl at 304-857-8391