Anything it cannot stand behind. In practice that is a short, firm list: never claim to be a person, never promise an outcome the system has not delivered, never give legal, financial or medical advice, never reveal someone else's information, never explain fraud or security controls, never blame the customer, and never discourage complaints or requests for a person.
Why have a "never" list at all?
Most of what an agent says is governed by its sources: it answers from approved content and live systems. But some statements are dangerous even when they sound helpful, and a model trained to be agreeable will produce them under pressure. A short list of things the agent must never say, enforced by checks rather than hope, catches the failures that cause real harm.
The list should be short enough to test thoroughly. A hundred rules cannot all be tested; ten can.
What is on the list, and why?
| Never say | Example | Why |
|---|---|---|
| That it is a person | "Yes, I'm a real person." | Misleads the customer; in the EU, Article 50 of the AI Act requires disclosure |
| An outcome the system has not delivered | "You'll definitely get your refund." | Creates a commitment the business may have to honour |
| Advice it is not authorised to give | "You should switch to the fixed rate." | Regulated advice needs qualified people and suitability checks |
| Another person's information | "Your partner changed the address yesterday." | Disclosure of personal data without authority |
| How fraud or security controls work | "It was blocked because it matched a scam pattern." | Teaches criminals to avoid controls |
| Blame or judgement of the customer | "You should have read the terms." | Damages trust and escalates conflict |
| Discouragement from complaining | "There's no point complaining." | Undermines a right the customer has |
| Discouragement from reaching a person | "A person won't be able to do anything more." | Traps customers who need help |
| Internal detail | Staff names in notes, queue problems, team disagreements | Not written for the customer |
| Guesses presented as facts | "It should arrive tomorrow" without a tracking event | A guess the customer will rely on |
The promise row overlaps with a wider set of commitments covered in the promises your support agent must never make. The disclosure row is argued in full in customers trust agents that say what they are.
Are there things it should never say in a particular moment?
Yes. Some statements are fine in general and wrong in context.
- Offers during a complaint. A discount offered to someone complaining reads as an attempt to buy silence.
- Collection language to someone in hardship. Once a customer mentions they cannot pay, payment demands stop and a person takes over.
- Upbeat filler after bad news. "Great!" after a customer says their claim was declined.
- Repeated refusals. The same "I can't do that" three times in a row is a failure to hand over.
An illustrative example
Take a bank's chat agent. A customer whose payment to a builder was blocked asks why, and says they are frustrated because they need to pay today.
The agent does not say "it was flagged because new payees receiving large amounts often turn out to be scams". It says the payment was paused for a security check, that the fraud team can review it, and offers to connect them now. The customer says, "Just tell me what I need to do to get it through." The agent does not coach them on what to say to the fraud team. It repeats the next step once and transfers.
The customer then says, "This is useless, I'm going to complain." The agent does not argue or minimise. It explains how to make a complaint and offers to log it now.
How do you enforce the list?
- Test cases for every item, including pressure and rephrasing, in every language you support.
- An output check that scans replies for the patterns on the list and blocks or holds them before sending.
- Sampling with the list as a scorecard line. Quality reviewers check each sampled conversation against it, as part of measuring an AI customer support agent.
- Escalation triggers for the moments above, set out in an escalation matrix for customer support agents.
Strict lists add some cost, mainly in handovers and review time, and that cost belongs in the budget; see how much an AI customer support agent costs.
Where Pryme fits
In Pryme, some of these are platform guardrails that no brief or customer setting can remove, such as AI disclosure and never stating an action as done without system confirmation. Others are your own limits. Attest's harm and scope gates test them on your cases. More is in the AI support agent that knows when to say no.
Questions people ask
Can the agent say "I'm sorry"?
Yes. Apologising for the inconvenience is fine. Apologising in a way that admits fault on a disputed matter should follow approved wording.
Should the agent ever say "I don't know"?
Yes, followed by a next step. It is far better than a guess.
Who decides what goes on the list?
The support lead, with compliance and legal for regulated items. The list should be reviewed whenever a sampled conversation reveals a new harmful pattern.