A technical challenge? Explore it and build a prototype.Explore the Lab

Business knowledge · 3 min read

The information your AI needs to answer customer questions

You want an AI assistant to answer customer questions. Giving it your procedures is a start, but some answers also depend on the customer’s agreement and what is available today. A course-date change shows why those details need to come together before the assistant makes a promise.

Written by Ludovic ChennebergPublished Updated

At a training centre, a transfer request is reviewed alongside the contract and session bookings.
AI-generated illustrative scene based on the article’s example: a written policy must be considered alongside the contract, current bookings and decision authority.

In this fictional example, a training provider receives a request: “I am booked for Tuesday. Can I attend on Thursday instead?” Its registration rules allow date changes when a place is available. The assistant finds the rule and prepares to say yes.

The coordinator knows that Thursday’s places are being held for a corporate customer. She needs to check whether one can be released. She also consults the participant’s agreement to establish which fees apply. The general rule alone cannot confirm this person’s transfer.

Watch how an employee finds the answer

To prepare the assistant, ask the coordinator to handle a request while you observe. Record the documents and applications she consults, what she looks for and which decisions require another person’s approval.

Then examine a request with different conditions. For example, the provider may have moved a course itself and waived transfer fees for affected participants. A note saying only “free transfer” loses the reason for the exception. Preserve enough context to prevent someone from applying it to a different situation.

This review may also expose an internal disagreement. Sales may treat a verbal commitment as a reserved place, while administration waits for a confirmed registration. The responsible manager needs to settle which rule applies before asking AI to follow it.

Connect each question to the right source

Information changes at different speeds. A registration rule may remain valid for months. A place can be booked while a customer is waiting for an answer.

What the employee needs to knowWhere the provider should check
Is a date change permitted?The current registration rules.
Can this participant have a place?The booking system, including places held for other customers.
Which fees apply to this customer?Their agreement and the applicable fee rules.
Who can approve an exception?The approval rule and the recorded decision.

A participant list exported last week may help an instructor prepare a course. It cannot establish whether a place is available today. The application needs current availability when answering and a booking check when confirming the place.

Document retrieval for AI finds relevant information and supplies it to the model before it answers. Updated documents can be used without retraining the model. The organization still needs to choose the right sources and assign responsibility for maintaining them.

Distinguish a proposed answer from a confirmed booking

The assistant can draft an answer explaining that a date change is allowed but that Thursday’s place still needs checking. The employee can then see what remains to be done. The transfer can be confirmed to the customer once the place has actually been reserved and any required approvals have been obtained.

Access should match that work. Reading this participant’s agreement should not expose every other customer’s negotiated prices. OWASP’s recommendations call for connected systems to enforce permissions rather than leaving the model to decide what it may read or change.

When information is missing, specify what the assistant should ask and who should receive the request. “I need to have the place confirmed” is more useful to a customer than an unverified booking promise.

Test the explanations with another employee

Before testing AI, give a few requests to a colleague who is less familiar with the work. Can they use the prepared material to find the agreement, check availability and identify who approves an exception? Their questions will expose information that remains hard to find or understand.

Then use the same requests to test the assistant. Compare its answer with the available facts and the expected decision. Include a case where a place is no longer available and one where approval is still pending.

When a rule changes, update the source and repeat the affected tests. The next customer should receive an answer based on the current conditions of their registration, with a clear distinction between what is possible and what has been confirmed.

Was this article useful to you?

Your feedback goes privately to TechnoConception. No account or email is required.

Responses expire after 90 days. They are not published or used as sales enquiries. A daily pseudonymous network identifier expires after 24 hours to limit abuse. Expired records are deleted by an hourly cleanup. The feedback database does not store your IP address; hosting security logs may still process it.

This browser stores only a receipt and your rating until the response expires, to reduce accidental duplicates. It does not store your comment or editing key. You can update a response during this page visit; after leaving, it becomes read-only in this browser.

Privacy policy

All insightsRSS feed

Related guides

An analyst clarifies a new-software request by distinguishing an incident, a licence need and a purchase.

AI expertise

How to assess a recommendation for an AI tool

An internal IT support example explains how a business owner or founder can assess an AI proposal and understand the technical choices behind it.

3 min read
A corrected AI draft sits beside a work calendar where a remaining task is assigned to available time.

Productivity

AI felt faster. The stopwatch disagreed.

Read METR’s productivity findings in context, then use two hypothetical calculations to distinguish complete effort, freed capacity and actual spending reductions.

5 min read