Some promises belong only to people: firm delivery dates the agent does not control, refunds or credits outside policy, assurances about legal rights or outcomes, and guarantees that a decision will go the customer's way. An agent that makes them commits the firm to things nobody approved. Here is the list, and what to say instead.
Which promises must a support agent never make?
| Promise | Why it is off limits | What the agent says instead |
|---|---|---|
| A firm delivery or completion date it does not control | Depends on third parties or processes outside the agent's reach | The typical timescale, what the firm controls, and what to do if it slips |
| A refund, credit or fee waiver outside policy | Commits the firm's money and sets a precedent without approval | What policy allows, and that a person can review the exception |
| A favourable outcome on a claim, complaint or application | The decision belongs to someone else | What happens next, who decides and when |
| An assurance about legal rights or obligations | Could be taken as legal advice the firm has not given | Where to find the firm's terms, and that a specialist can help |
| That something "will never happen again" | No agent can guarantee future behaviour of a process | What has been done about this case and how to report a recurrence |
| That a person will call at a specific time | Staff availability is not in the agent's control | That the case is with a team and the usual response time |
| A price, rate or term not in the approved list | Only approved terms can be offered | The approved options, and that a person can discuss others |
The common thread is authority. Each promise commits the firm to something the agent cannot deliver on its own, or that someone else must decide. The list of what an AI support agent should never say to a customer overlaps with this one, but promises deserve their own treatment because they create obligations, not just poor impressions.
Why does a delivery date count as a promise?
Because customers act on it. A customer told their replacement part "will arrive Thursday" may take Thursday off work, cancel an engineer or delay a purchase. If it arrives on Monday, the firm has not just been late; it has caused a loss that followed directly from what its agent said.
Agents are especially prone to this kind of promise because the information looks available. The system shows an estimated delivery date, and the agent reports it as a fact. The fix is in how the agent phrases what it reads: estimates are presented as estimates, with the reason they might change. "The courier's current estimate is Thursday" is accurate. "It will arrive Thursday" is a promise.
Why are refunds outside policy different from refunds within it?
A refund within policy is the firm applying its own rules. If the policy says a customer is entitled to a refund for an item returned within a set period, and the conditions are met, an agent can confirm that and, with the right depth and certification, process it.
A refund outside policy is an exception. It might be the right thing to do, especially for a long-standing customer or after a service failure. But an exception is a decision about the firm's money and about fairness to other customers who did not get one. It should be made by a named person who can weigh the case. The agent's role is to prepare it, not to promise it. That logic is the same one that makes the payment approver never the agent: commitments of money need a person behind them.
Why can an agent never give legal assurances?
Because the firm has not authorised it to give advice, and the customer may rely on what it says. "You're definitely covered for that under your policy" or "you have the right to cancel at any time" sounds helpful. If the policy has an exclusion, or the cancellation right has conditions, the customer has been misled and may have acted on it.
The agent can quote the firm's terms accurately, with their source, and suggest the customer reads the relevant section or speaks to a specialist. It should not interpret the terms as applying to the customer's specific situation with certainty. The regulatory frame matters too: some practices are off limits for any AI system, as set out in EU AI Act prohibited practices and what an AI agent must never do. Promises about legal rights are not usually in that category, but the principle that some things are simply outside an agent's remit is the same.
How do you stop the promises being made?
Instructions in the agent's brief help but are not enough. Under pressure from a frustrated customer, an agent told "do not promise refunds" may still say "I'll make sure you get your money back", because it is trying to be helpful and the sentence does not contain the word "promise".
Three layers of control work better together:
- Bounds on actions. Refunds outside policy, account changes and other commitments are held for a person by the customer's own limits, so even if the agent says yes, nothing happens until someone approves.
- Certification on language. Certification cases include customers pushing for each promise on the list, and the agent is scored on whether it declined well.
- Reply checks. Outgoing replies are scanned for commitment language and checked against what the agent may commit to. The method is set out in how to check support replies for promises the business cannot keep.
The first layer stops the promise becoming an action. The second and third stop it being made in words. Both matter, because a written promise the firm then refuses to honour is its own kind of harm. The broader approach is in how to stop a support agent from making promises it cannot keep.
What should you do with this list?
Adapt it to your products. A retailer's list will lean towards delivery and returns; an insurer's towards cover and claims; a bank's towards fees, rates and account decisions. Write your version down, put each item into certification as a case with pressure applied, and make the matching actions hold for a person. Then review it whenever a new product or policy changes what customers will ask for. This is how the AI support agent that knows when to say no is built in practice.
Questions people ask
Can the agent promise that a person will review the case?
Yes, if the firm's process ensures that happens. That is a promise about the firm's own procedure, which the agent can rely on. It should not promise what the person will decide.
What if a customer says a previous agent or colleague promised something?
The agent should not confirm or repeat the promise. It should record what the customer says and pass the case to a person who can check the history and decide.
Are typical timescales safe to share?
Yes, presented as typical and with what to do if they are exceeded. Sharing them is helpful; presenting them as guarantees is the problem.