PrymeIntelligence

Customer support·Answer

Can an AI agent make exceptions to company policy?

An AI agent should not make exceptions to company policy. It can apply written discretion, spot a case for an exception and draft it, but a person decides.

No. An exception is a decision to step outside the rule, and an agent's authority is the rule. What an agent can do is apply discretion the policy already writes down, recognise when a customer has a real case for an exception, and draft it for a person to approve. If the same exception keeps being approved, change the policy.

Why can an agent not grant exceptions?

Because its bounds are the policy. An agent that can step outside its limits when the story is sad enough does not have limits; it has suggestions. Customers learn quickly what works, and so do fraudsters. "Your manager told me last week you'd make an exception" is a line that has persuaded people for decades. Agents are easier to persuade, and they can be targeted with crafted text that tries to rewrite their instructions.

Fairness matters too. One reason to use an agent is that it treats similar customers alike. Exceptions granted on the agent's own judgement undo that.

What is the difference between discretion and an exception?

Many policies contain written discretion: "Late fees may be waived once in twelve months where the customer has a good payment history." Applying that is following the policy, not breaking it. The agent checks the conditions and acts within them.

An exception is a request the policy does not cover, or one it rules out. The test is whether the agent can point to the clause that permits the outcome. If it can, it is a rule. If it cannot, it is an exception, and a person decides.

What may the agent do, and when must it hold or refuse?

Situation The agent A person
Outcome permitted by written discretion Applies it, citing the clause Samples
Customer asks for something the policy rules out Refuses, gives the reason, offers a person Handles any challenge
Customer has a plausible case the policy did not foresee Drafts the exception with evidence and holds Decides
Customer claims someone already promised an exception Checks the record; if absent, holds Decides
Same exception approved repeatedly Reports the pattern Updates the policy

How refunds, credits and exceptions are drafted and held is covered in refunds, credits and exceptions: draft and hold. The same distinction appears in finance, where expense agents enforce policy consistently and people handle the exceptions.

What does this look like in practice?

Take a furniture retailer with a 30-day returns window. A customer asks to return a chair on day 34, explaining they were in hospital for two weeks after it arrived.

The policy has no clause for this. The agent does not refuse flatly and does not approve. It says the window has passed, that it can ask the returns team to consider an exception, and asks whether the customer can share a discharge letter or similar. It drafts the exception with the dates, the reason and the evidence, and holds it. The team approves. Over the next quarter, the team approves eleven similar requests, and the policy is changed to extend the window for documented hospital stays. From then on, the agent applies it as a rule.

Which controls make it safe?

  • Bounds enforced at the moment of action, not by instructions the customer can argue with.
  • Written discretion held in the policy library, with clause references the agent cites.
  • Claims of earlier promises checked against the record before anything else.
  • A register of held exceptions, reviewed for patterns.
  • A list of things the agent must never say, such as "I'll make an exception just this once"; see the promises your support agent must never make.

Retention offers are a common area where discretion is written into limits, as in how to make retention offers within set limits with an AI agent. The wider approach is in the guide to AI customer support agents.

Where Pryme fits

In Pryme an exception outside the agent's bounds is refused or held at the moment of action, whatever the conversation says. The verdict names the bound that decided, so you can see whether the block was role scope, the customer's own limits or the certificate. Only a person can approve the held draft.

Questions people ask

Can the agent apologise when it cannot make an exception?

Yes, for the customer's situation, without implying the decision might change. See whether an AI agent can apologise on behalf of the company.

Will customers be annoyed that the agent cannot bend?

Some will. A clear reason and a quick route to a person usually matter more than flexibility.

Should the agent tell customers exceptions are possible?

Only where a person really does consider them. Raising hopes it cannot meet is its own kind of broken promise.

Decide what your agents may do. Prove it. Enforce it.

Pryme Intelligence certifies each agent on your own cases, checks every action against five bounds at the moment it happens, and files every verdict.