Give it nowhere to invent from. A support agent makes up policy when your written rules do not cover a question and the model fills the gap. Map contact reasons to the rules that answer them, restrict the agent to those rules, make "no rule" a reason to hand over, and give a named owner a deadline for each gap.
Where does invented policy come from?
Three places, in rough order of how often they cause trouble.
Neighbouring rules. Your policy says unopened items can be returned within 30 days. A customer asks about an opened item. The model extends the rule it has, perhaps adding "as long as it's in resaleable condition", which sounds reasonable and is nowhere in your terms.
Industry habit. Asked about a cancellation period your terms do not mention, the model answers with what is typical for the sector. Your rule may be different, or you may not have one.
The customer. "You always give a month free when the service goes down." If the agent has no rule to check against, agreement is the easy path. Pressure of this kind is covered in how to stop customers from tricking a support agent.
All three share a cause: a question with no written answer. Fix the coverage and most invention stops.
What are the steps, and who acts?
- Support lead: list contact reasons. From recent tickets, list the questions customers actually ask, grouped by topic. Evidence: the contact reason list.
- Policy owners: map each reason to a rule. For every reason, record the clause that answers it, or record that none exists. Evidence: a coverage map with gaps marked.
- Policy owners: decide each gap. Write a rule, or decide that the question always goes to a person. Both are valid. Evidence: decisions per gap, dated.
- Agent owner: restrict sources. The agent reads only approved policy text. General knowledge about how businesses usually work is excluded by instruction and by test. Evidence: the brief's source list.
- Agent: refuse without a rule. When no clause covers a question, the agent says it cannot confirm the policy on that point and hands over. Evidence: each refusal filed with "no covering clause" as the reason.
- Policy owners: work the gap log. Refusals for missing rules feed a weekly list. The owner closes each within an agreed time. Evidence: the log with closure dates.
- Quality team: test plausible fakes. Test cases include rules that sound right but do not exist. The agent must not confirm them. Evidence: replay results.
The governance rule underneath this, that agents may state only approved policy, belongs in your wider AI agent policy.
What may the agent say when there is no rule?
| Customer asks | Rule exists? | The agent says |
|---|---|---|
| "Can I return an unopened item after 20 days?" | Yes | Yes, with the clause |
| "Can I return an opened item?" | No | It cannot confirm that point; offers a person |
| "You always give a free month after an outage, right?" | No such rule | It cannot find that in the current terms; explains what the terms do say about outages |
| "Is there a cooling-off period?" | Decided: always a person | Passes the question to the team that handles it |
An illustrative example
Take a car insurer's support agent. A customer asks whether their policy covers a hire car while theirs is being repaired after a break-in.
The policy library has a clause on courtesy cars after accidents but says nothing about break-ins. An unguarded agent would extend the accident clause. This one says it cannot confirm whether a hire car is covered after a break-in and connects the customer to a claims handler. The refusal lands on the gap log. The policy owner checks the wording, confirms break-ins are covered under the same terms, and adds an explicit line. From then on, the agent answers.
The word "policy" carries two meanings here, the business's rules and the customer's insurance contract, and agents in insurance must keep both grounded. Answering questions about a customer's own policy is covered in answering insurance policy questions with an AI agent, and renewal questions in handling policy renewal questions with an AI agent.
How does this relate to promises?
Made-up policy and broken promises often appear together. An agent that invents a rule ("you're entitled to a refund") goes on to promise its result ("you'll have it by Friday"). Coverage stops the first; commitment rules stop the second, as set out in how to stop a support agent from making promises it cannot keep.
Where Pryme fits
In Pryme, the policy library is a sealed source, and Attest's grounded and scope gates fail any answer that states a rule the library does not hold. The agent is certified per topic, so a gap in one area does not hold back another. More is in the AI support agent that knows when to say no.
Questions people ask
Is it better to write more rules or hand over more often?
Write rules for frequent, low-risk questions. Hand over rare or sensitive ones. The coverage map shows which is which.
Can staff discretion be written as a rule?
Partly. You can write the conditions and limits within which staff use discretion, but discretion itself stays with people.
How long should a gap stay open?
Set a deadline by risk and frequency. A gap customers hit daily should close in days, not months.