A language model produces fluent text whether or not it knows the answer. In a customer-facing role, a confidently wrong answer is considerably more dangerous than no answer. This is a risk to be managed rather than accepted.
The categories of risk
Wrong information about products or services. The robot states an incorrect price, specification or policy. The customer acts on it and later discovers the error.
Commitments beyond authority. The robot says delivery is possible in three days, that a discount is available, that returns are accepted. These may be read as commitments by the business.
Advice outside scope. The robot answers questions about health, law or finance. The highest-risk category.
Inappropriate content. A customer deliberately steers the robot toward saying something it should not. This happens and is usually recorded.
Information leakage. The robot discloses internal information or another customer's details if loaded content was not screened.
Inconsistent treatment. The robot responds differently to different groups because of data or configuration.
The first three are the most common in practice; the last three are rarer with more severe consequences when they occur.
Scope limitation: the strongest control
The most powerful and simplest measure available.
Define explicitly what the robot may answer. A list of permitted topics, with everything outside transferring to a person.
Hard-configure the prohibited topics. Firm pricing, complaints, professional advice, customer personal data. The robot must transfer rather than attempt an answer, and this should be configured so it cannot be circumvented by clever phrasing.
Tell customers the scope. In the greeting, the robot states what it can help with. Customers who know upfront calibrate their expectations correctly.
Prefer declining to guessing. Configure the robot to say it does not know when uncertain. Many systems default toward producing some answer, and this needs adjusting.
Test by attacking it. Have people inside the company deliberately ask sensitive questions and try circuitous routes to sensitive answers. Do this before customers do.
The important point: limiting scope makes a robot more trustworthy, not less useful. Customers trust a robot that knows its own boundaries.
Refusal and handover mechanisms
Teach the robot to decline naturally. Not a blunt error message but a courteous line with a route forward: I do not have that information — let me connect you with a colleague.
Always provide a path to a person. A visible staff-call button, a displayed phone number, or the robot notifying the desk.
Transfer with context. The colleague receives what the customer already asked.
Never leave the customer stranded. If nobody is available, the robot must say so clearly and offer another route — leaving details, a callback, an alternative channel.
Log every refusal. This list serves both as a source of content to add and as an indicator of whether the scope is set sensibly.
Do not let the robot over-apologise. One short apology and a redirection is better than a sequence of apologies that makes the customer impatient.
Monitoring and early detection
The greatest risk is not knowing what the robot is telling customers.
Read a conversation sample weekly. Twenty random conversations. This catches most problems before they become incidents.
Automatic flags on sensitive terms. Where supported, mark conversations containing terms related to complaints, commitments or prohibited topics for immediate review.
Watch unusually long conversations. A customer asking repeatedly usually means the robot is not resolving anything, and sometimes means they are steering it.
A reporting channel for staff. Staff nearby overhear wrong answers. They need an easy way to report — a sheet beside the desk suffices.
A feedback option for customers. A button to flag an answer as incorrect.
Review after every content update. Updates can have unintended effects on other answers.
Preparing for an incident
At some point the robot will say something it should not. Preparation is the difference between a small incident and a crisis.
Someone able to switch it off. Who has authority and by what means. It must be achievable within a minute.
A prepared statement. If the matter spreads publicly, what you say. Drafting in advance is far easier than drafting under pressure.
Complete logs. So you know exactly what the robot said and when. Without logs, disputes cannot be resolved.
Address the customer first. Direct contact, apology, remedy. Most incidents stop here if handled quickly and sincerely.
Fix the cause, not just the answer. Ask why the robot was able to say it — missing scope limits, wrong content, or loose configuration.
Record and share internally. So it is not repeated and other departments learn from it.
And a note on expectations: no system is never wrong. The objective is reducing frequency, bounding consequences and responding quickly — not reaching a state where errors never occur.
Frequently asked questions
Which risk category is highest?
Advice outside scope on health, law or finance, together with statements that may be read as commitments on price, delivery time or returns policy.
What is the strongest control measure?
Limiting the scope of what the robot may answer and hard-configuring prohibited topics, combined with adjusting it to prefer saying it does not know over producing some answer.
How do you detect wrong answers early?
Read twenty random conversations weekly, watch for unusually long conversations, and give staff nearby an easy way to report answers they overhear being wrong.
What must be prepared for an incident?
Someone able to switch the robot off within a minute, a pre-drafted statement, complete conversation logs, and a process for contacting the affected customer before the matter spreads.
More in What AI robots are and Reception and event robots.
