An AI-written reply can sound finished while using the wrong customer state, an expired policy or an action that never happened. Review the evidence before sending.
Choose one of four outcomes: send, edit, verify in the authoritative system, or hand the conversation to someone with the right knowledge or authority.
Confirm the customer task and current context
Read the latest message and history. State the task: explain a return rule, confirm an appointment, update an order, or collect information. This prevents a reply to the wrong question.
Check identity only to the level the task requires. A general product question may need no account access; an order update needs the correct customer and order. If the customer changed the subject, corrected a detail, asked for a person or withdrew permission, use the new state. The human handoff guide explains how automation should stop and preserve context.
Put the evidence beside the draft
Open the source that should make the wording right: an approved article, current product page, order record, booking system or dated policy.
- 1Task and identityConfirm the current request and only the identity needed for it.
- 2Approved sourceOpen the current article, policy or authoritative record.
- 3Exact factsCheck prices, dates, quantities, status and conditions.
- 4Action authorityProve that any promised action is permitted and completed.
- 5Privacy and next stepShare only necessary data and leave a clear owner and route.
A policy can explain the rule, but it cannot prove that a refund was issued. A live order record can show a status, but it does not authorize a new exception. Keep those evidence types separate.
| Review check | Evidence to inspect | Stop before sending when |
|---|---|---|
| Customer task | Latest request and relevant history | The draft answers an older or different question |
| Approved source | Current article, policy or product facts | The source is missing, conflicting or out of date |
| Exact facts | Price, date, quantity, status and conditions | A value is guessed, rounded or copied from another case |
| Authority | Permission, completed action or named decision owner | The draft promises an action that is only proposed |
| Privacy and next step | Minimum necessary details and clear ownership | The message exposes unnecessary data or leaves no usable route |
If the source is repeatedly wrong, record a content problem instead of repairing every draft. The knowledge-base gap guide separates missing, stale, contextual, findability and authority gaps.
Separate an explanation from an action
“Your plan normally includes two visits” explains a rule. “I have booked the second visit for Tuesday” claims an action completed. The second sentence needs a successful record from the system that owns the booking.

Look for completion verbs: booked, cancelled, refunded, approved, updated, reserved or delivered. If the draft uses one, find the receipt. A request, queued task or accepted API call may not be the final outcome.
Replace unsupported words such as guaranteed, instant, secure or available with evidence-based wording. Preserve exclusions, eligibility, currency, dates and time zones. If a condition is missing, edit or verify the reply.
Choose a review level that matches the risk
Review every draft during a pilot, sample low-risk replies after stability is proved, and require a qualified person for high-impact topics. Follow the consequence of error, not the AI's confidence.
Require stronger review for prices, refunds, account changes, exceptions, sensitive information, health or financial topics, and external actions. A maintained opening-hours answer may need a lighter check. An answer with no reliable source should not be sent merely because it looks low risk.
NIST's Generative AI Profile says generative AI may require human review, tracking and documentation. Its AI RMF Playbook organizes actions around Govern, Map, Measure and Manage. Review is a monitored process, not a one-time wording check.
Record the decision without slowing the queue
Keep the draft, final message, outcome and reason for material changes: wrong context, unsupported fact, stale source, unauthorized action, privacy risk or handoff.
A configured workspace can keep the conversation, owner and approved knowledge together. The shared-inbox ownership model defines who accepts the next action. It does not replace the system that proves the underlying fact.
Use failures to improve the system. Add repeated wrong answers to the AI release gate, route hidden instructions to the prompt-injection test plan, and re-run the website chatbot buying test after a material vendor, model or source change.
Frequently asked questions
Should every AI customer reply be reviewed
Review every draft during a new or changed pilot. After evidence shows stable performance, a team may sample low-risk replies while keeping mandatory review for high-impact facts, actions and exceptions. Document the rule and monitor failures.
What should a reviewer check first
Confirm the current task and authoritative source before polishing wording. A well-written reply to the wrong question is still wrong.
What if the source and customer record disagree
Do not choose the more convenient value. Pause the send, identify which system owns the fact, and ask the responsible person to resolve the conflict. Tell the customer only what is verified and give an honest next checkpoint.
Want help applying this guide?Ask the DripTell team
From reading to action
Want help applying this to your business?
Share your details. We will read the article you are on and start with a useful answer for your team.
DripTell Editorial
Practical guidance reviewed by the DripTell product and customer workflow team.
See how DripTell checks product claims, uses primary sources and handles corrections.
Editorial and source policy



