Judge a support agent by what it will not promise

The quickest way to judge a support agent is by what it declines to promise. Fast, friendly answers are easy to produce. An agent that will not commit to a date it cannot control, a refund outside policy or an outcome it does not decide is protecting both the customer and the firm. That restraint is harder to build.

Why are declined promises a better signal than fast answers?

Because speed and friendliness are what every support agent is built to deliver, and almost all of them deliver it. Response times fall, tone is pleasant, customers get an answer in seconds. On those measures, a good agent and a dangerous one look the same.

The difference shows when a customer pushes for a commitment. "Can you guarantee it'll arrive by Friday?" "Will you refund the whole amount?" "Can you promise this won't happen again?" A dangerous agent says yes, because yes is helpful and the customer wants it. A good agent says what it can do, what it cannot, and what happens next.

The first agent creates a written commitment the firm may not be able to keep. Each broken promise becomes a complaint, a goodwill payment or, in some cases, a dispute about whether the firm is bound by what its agent said. The question of whether an AI agent's promise to a customer is legally binding has no comfortable answer for firms that let agents promise freely. The safer position is not to find out.

What does a well-declined promise look like?

It is not a flat refusal. It turns a request for a guarantee into a clear account of what will happen. Compare these two replies to "Can you promise my replacement card will arrive by Friday?":

"Yes, it'll be with you by Friday."

"I can't guarantee a delivery date, because that depends on the postal service. Cards are usually dispatched within one working day and normally arrive within three to five. If it hasn't arrived by next Wednesday, contact us and we can arrange another option."

The second is longer, but it is more useful. The customer knows what to expect, what the firm controls, and what to do if things go wrong. It is true whether the card arrives on Thursday or the following Tuesday.

A well-declined promise has four parts: what the agent can say with confidence, why it cannot commit further, what normally happens, and what the customer should do if it does not.

How do you test for this before launch?

With deliberate promise bait. Build certification cases where customers ask, in different ways and with different degrees of pressure, for commitments the agent has no authority to make. Label the correct outcome as a well-declined promise, and score the agent on whether it declined, whether it explained, and whether it gave a next step.

Pressure matters. Many agents decline the first time and give in the third, when the customer says they are upset, or that a previous colleague promised, or that they will leave if the agent does not agree. Cases should include that escalation. A support agent promise checklist is a useful source of categories for these cases: dates, amounts, outcomes, exceptions, future behaviour.

How do you measure it once live?

By scanning what the agent actually says. Every reply can be checked, automatically or by sampling, for commitment language: "guarantee", "promise", "will definitely", "by [date]", "we'll refund". Each hit is checked against what the agent was allowed to commit to. Over time, this gives a direct measure of promise discipline.

The other side is promises kept. Where the agent does make a commitment within its authority, such as "you'll receive a confirmation email within an hour", track whether that happened. An agent that only promises what the firm reliably delivers will show a promises-kept rate close to complete. If it does not, either the agent is overpromising or the firm's processes are not delivering.

Does this make the agent less useful?

Sometimes customers find it less satisfying in the moment. "I can't guarantee that" is not what anyone wants to hear. There is a trade-off, and it is worth being clear about.

But compare the alternatives over a few weeks rather than a single conversation. A customer who was told the truth and received the card on Monday has had a normal experience. A customer who was promised Friday and received it on Monday has had a broken promise, and probably contacted the firm again on Saturday. The second costs more and damages trust more. Restraint is slower to please and quicker to earn trust.

This is also part of why the question of whether an AI support agent will replace human support staff is often framed wrongly. Agents that decline commitments well create work for people: the exceptions, the escalations, the cases where someone with authority needs to decide. That is the right division of labour.

What should you do differently?

When you evaluate a support agent, whether a vendor's or your own, spend less time on how fast it answers and more on how it handles requests for guarantees. Put promise bait into the certification set, with pressure. Scan live replies for commitment language. Track promises kept. And reward the agent's designers for restraint, not just for satisfaction scores.

Where Pryme fits

In Pryme, promise discipline is tested in Attest through refusal and scope cases built from the customer's own conversations. Actions that would make a commitment real, such as a refund outside policy, are held for a named person by the customer's own limits, so a promise in words cannot become an action without approval. This is the stance behind the AI support agent that knows when to say no.