A customer-facing AI assistant can make a particularly costly mistake without sounding confused: confidently quoting last week's promotional price or explaining a returns policy that the business has already changed.
The answer: once an assistant is live, every change to prices, opening hours or policies needs an owner, an approved source and a test before it becomes customer-facing information. This is a knowledge governance problem, not something a cleverer prompt can reliably solve.
Preparing a knowledge base is only the beginning
Before launch, organize your products, services and common questions. Our company data preparation guide covers that stage. After launch, stock changes, offers expire, prices are revised and opening hours shift. The harder question becomes: who may approve a change, and how do you stop the new answer from conflicting with an older source?
Imagine marketing publishes a one-week offer, sales has a revised price list, and an older PDF is still available to the assistant. Without clear authority and effective dates, those sources may conflict. Do not assume the model will reliably infer which document wins.
Four fields every changing fact needs
An owner: the person or role authorized to approve it. A sales or finance lead may own pricing, while a clinic reception team owns appointment hours.
An approved source: one designated page or record. A screenshot in a group chat should not silently outrank a signed-off policy.
An effective date: when does a price or offer begin and end? If it applies only to one branch or customer group, state the scope.
A fallback: when sources disagree or a condition is missing, the assistant should avoid confident guessing and route the question to a human. “Let me confirm the current price” is more useful than a wrong figure.
A practical price-change workflow
1. Capture the change request
Record what changed, who approved it, when it takes effect and which pages or responses it affects. Avoid scattering instructions across private messages.
2. Update the source of truth and retire the old version
Adding a new file while leaving the old one equally authoritative creates conflict. Archive the obsolete source clearly or revise the controlled record. Keep an internal change history so errors can be traced.
3. Test multiple ways customers ask
For a price change, test “How much?”, “Is the offer still available?”, “What is included?”, “Does that apply to my branch?” and informal Egyptian Arabic with spelling errors. Passing one exact FAQ question is not enough.
4. Test what the assistant must not decide
Ask for an unauthorized discount, an unavailable item or a policy exception without an approved rule. The correct behavior may be an explicit limit and a human handoff, not a plausible-sounding answer.
5. Review early live conversations
Sample replies after the update, note corrections by staff and investigate complaints linked to stale information. If the new source is faulty, use a documented human-reviewed rollback instead of layering improvised instructions on top.
What if marketing and sales disagree?
Do not ask the model to settle an internal policy dispute. Name one decision owner for each information category and an escalation route for conflicts. Marketing may draft an offer, but final commercial terms need the relevant approval. Record that decision so it does not reappear in the next customer conversation.
The NIST AI Risk Management Framework Playbook offers voluntary guidance on governance, measurement and risk management. The relevant lesson is clear accountability and monitoring after deployment. Citing it does not imply that Mr. AI has NIST certification.
How do you measure update quality?
Track the time from an approved change to its appearance in the controlled answer source, the number of staff corrections caused by stale information, appropriate human escalations for missing facts and test results after each revision. A low escalation rate is not success if the assistant is guessing.
According to the current Mr. AI FAQ, teams can add company information, configure rules, test replies and take over conversations when needed. This does not establish enterprise approval workflows, automatic synchronization with every database or built-in document version control. Verify the available tools before promising those capabilities.
Start with one real update this week. Record the previous and approved versions, test ten customer phrasings and name the person who signs off. Then try Mr. AI with your own information or book a demo to discuss the appropriate setup.
Sources: NIST AI RMF Playbook, Mr. AI FAQ.
