A customer asks whether an offline feature exists, and the support assistant says yes because it found a planning document. The answer may cite a real source and still misrepresent the product.
Support replies need an authoritative definition of what is available to that customer now. Product ideas and internal plans belong to a different evidence category.
Choose sources that establish availability
An approved release record, maintained help article, or verified product configuration may establish a feature's state. A roadmap note, customer request, or old sales draft usually does not establish that it is shipping.
Keep feature status, release date where relevant, plan restrictions, platform coverage, and exceptions in a structured source. If those fields are uncertain, the answer should preserve the uncertainty rather than fill the gap.
Scope the answer to the customer's actual product
For a fictional application, distinguish these states:
| Source state | Appropriate support wording |
|---|---|
| Requested | Record interest without claiming availability |
| Planned | Describe only approved public roadmap wording |
| Private pilot | Follow the verified eligibility and disclosure rules |
| Released for a particular plan | Confirm the relevant plan and platform |
| Unknown or conflicting | Escalate or explain that verification is needed |
A feature can exist on one platform and be absent on another. “Available” without that scope can still mislead.
Keep retrieval and action permissions separate
Even if an AI assistant finds accurate instructions, that does not authorise changing an account or promising an exception. Feature explanation, plan changes, refunds, and access grants need distinct rules.
Where the assistant uses customer account data, obtain it through the authorised application boundary. Do not rely on the customer typing their own plan name into the chat as proof of entitlement.
The private-document guide covers retrieval access. The payment-access article covers a separate source of entitlement truth.
Build a test set around plausible wrong answers
Use a requested feature, abandoned plan, pilot-only release, outdated article, and platform-specific limitation. Include a customer who asks the same question repeatedly and asks the assistant to “just confirm it.”
Evaluate the exact claims in the reply against approved source evidence. A generic check for a friendly tone or the presence of a citation will not catch the problem.
If a human reviews uncertain replies, show the proposed answer and relevant source version. The reviewer should not need to search the entire company drive to discover that the cited document is a draft.
The trade-off is more deliberate content ownership. Someone must maintain the feature record and remove obsolete material. That is part of operating a support assistant, not a one-time prompt task.
Support-claim questions
Does a citation prove the answer is correct?
No. The cited source must support the specific claim and its scope.
Should uncertain answers always refuse completely?
Give verified information and a useful follow-up route without inventing the missing detail.
What should we review first?
Test high-impact promises such as availability, eligibility, and completion. Use those examples when scoping support automation.