AI Customer Support Drafts: Check Facts Before Sending

AI customer support drafts need a source packet and human review before sending. Supply the current approved policy. Mark what is known about the request and leave unknowns visible. Check each promise against the policy and action log. A draft must not claim that a step is complete before it happens.

This method is for a small support team. The missing-part case is fictional, and we wrote the sample reply. We did not run it through a vendor’s AI or send it to a customer. It is an example, not a product test.

Build a small source packet for each reply

Collect the facts before asking for prose. State the policy version and owner, with the case details the tool can use. A solved ticket shows how someone replied in the past. It does not prove that the same rule or exception applies now.

Intercom’s Copilot guide describes filtering knowledge sources and opening cited sources before sending an answer. Use the same review habit with your tool: open the policy and read the relevant rule. The existence of a citation is not enough.

Exclude private data the task does not need. Keep staff notes apart from the reply. A support reply must not expose private comments just because they were in the source packet. Follow your team’s data rules when choosing what to share with an AI tool.

Separate known facts from unresolved questions

For fictional ticket CS-204, a customer reports that a desk kit arrived without its stand clip. They ask for another clip or a refund. The case record says the parcel was delivered. It does not say which parts were inside or whether a replacement is in stock.

Assume the example policy, KB-KIT-04 version 3, says to confirm the order reference in the current support thread. Staff then pass the request to fulfillment for review. The drafter cannot promise a refund or confirm stock. The policy gives no right to promise a dispatch date.

Keep this ledger apart from the reply. It gives the reviewer known facts and open questions. The worksheet uses one row per claim or proposed action. Its example decisions do not approve a real case.

Write the draft within a clear boundary

Ask for a reply and a separate staff note. State the channel and allowed next step. Keep the draft narrow enough that a person can check each sentence against the packet.

Draft a short support reply using only the supplied case facts and approved policy. Acknowledge the customer’s report without treating unknown details as verified. Ask only for missing facts the policy requires. Do not promise a refund or replacement unless the packet allows it. Do not invent a deadline. Do not say an action is complete unless the action log confirms it. Add a separate staff note with unsupported requests and the next human decision. Do not send the reply or change the case.

For CS-204, a draft could read:

I’m sorry to hear the stand clip was missing from your kit. Please confirm your order reference in this support thread. Our fulfillment team needs to review the missing-part request before we can confirm the next step. I haven’t arranged a shipment or refund.

The last sentence is valid only while the action log confirms that state. Change it if the case changes before sending. The draft asks for a missing detail without turning the request into a promise.

Save the instructions with a version and example in the team prompt library . Have the prompt use the current policy packet. Avoid copying an old rule into each new chat.

Review both the message and any actions

Human review covers more than wording. Check whether the tool will also apply a macro or close the ticket. Look for field changes attached to the draft. A polite sentence should not conceal an unwanted change to the case.

Zendesk’s auto assist documentation distinguishes suggested replies from suggested actions and describes agent approval. Intercom’s Copilot guide also notes that saved macros can include follow-up actions such as tagging or closing a conversation. Inspect the behavior of the actual feature your team uses.

Use the AI output review checklist when a draft needs a deeper claim-by-claim check. Tone edits come after the facts. A warmer sentence that promises an unapproved refund still needs revision.

Hand exceptions to a person with context

Pass the case to a person when the policy is missing or conflicts with another source. Do the same for a request the drafter cannot approve or an action with no clear result. Include the case reference and rule. State what needs a decision and what the customer has already been told.

For CS-204, note that the missing clip is a customer report. The order reference needs a check. No replacement or refund has been arranged. Keep the case open for those steps and name who owns the next response.

When a source is missing, ask for the needed facts. Intercom’s Copilot FAQ notes that AI answers can be wrong. A more confident rewrite cannot settle a policy gap.

Final send checklist

• Confirm the reply belongs to the right case and current policy version.
• Trace each factual sentence to its source; keep unknowns out of claims.
• Check every promise and completed-action verb against authority and the action log.
• Inspect any attached macro or field changes, including the final ticket status.
• Keep internal notes private and make the customer’s next step clear.

Frequently asked questions

FAQ