Start with the policy, not the prompt
A support reply has two jobs: address the customer’s question and accurately represent what your business will do. An AI assistant can make a reply easier to read, but it may also fill gaps with plausible-sounding promises. Before drafting, identify the published policy that applies, such as a returns, cancellation or delivery policy. Use the approved text as the source for the answer rather than asking the model to recall your rules.
Keep policies in a place your team can maintain, and make it clear which version is in use. If different products, regions or purchase channels have different terms, label those distinctions. A short policy excerpt that matches the case is usually more useful than a large collection of loosely related documents. When no policy covers the request, the draft should say that the team needs to check, not invent a rule.
Separate case facts from policy rules
Give the drafting assistant two clearly marked inputs: the facts known about this customer’s case and the relevant policy text. Case facts might include what the customer asked, the product involved and an order status confirmed in your system. Do not treat the customer’s description as proof of details you have not verified. For example, a customer may say a parcel is lost, while your team knows only that it has not arrived.
This separation makes the draft easier to review. A reviewer can ask whether each statement comes from the case record, the policy or a proposed next step. If a fact is missing, the reply can request the specific information needed to proceed. It should not imply that a refund has been approved, a replacement has shipped or an exception has been granted unless someone has confirmed that action.
Use a prompt that sets useful boundaries
A reusable prompt can tell the assistant to write in your support voice while limiting what it may claim. For example: “Draft a concise, courteous reply using only the case facts and policy excerpt below. Explain the applicable policy in plain language. Do not promise an exception, refund or timeline unless it is explicitly supported. If information is missing or the policy does not resolve the request, identify what a support agent should check. Do not include internal notes in the customer-facing reply.”
Ask for two outputs: a proposed customer reply and a short review note listing the policy passage used, missing facts and any proposed action that needs approval. The review note is for your team, not the customer. This format helps a busy agent see where the draft needs attention. It also discourages a fluent reply from concealing uncertainty, although it does not guarantee that the assistant has followed every instruction.
Review the draft before it reaches the customer
Check the draft against the policy itself, especially dates, eligibility conditions, exclusions and the steps a customer must take. Look for small wording changes that alter a commitment: “we can review your request” is different from “we will approve your request.” Confirm that the reply answers the actual question, uses an appropriate tone and gives the customer a clear next step. Remove explanations that are technically accurate but irrelevant to the case.
Some cases need more than an edited draft. Escalate requests involving policy exceptions, disputed facts, sensitive personal information or an action the agent cannot authorize. A customer who describes an urgent or potentially harmful situation may also need a different response process. AI can help prepare a summary for the person handling the case, but the decision and customer-facing commitment should remain with an authorized team member.
Protect customer information and improve the workflow
Before entering a case into an AI tool, follow your organization’s rules for customer data and the tool’s permitted use. Include only the information needed to draft the reply; an order status may be useful, while a full address or payment details may not be. If your team cannot use the tool with real customer records, practice with fictional examples or remove identifying details where appropriate.
Save approved examples of effective replies alongside the policies they reflect, and update them when a policy changes. During routine review, note recurring problems: unsupported promises, missing questions, overly firm language or confusing explanations. Use those observations to improve the source material, prompt and review checklist. The goal is not to automate away judgment. It is to give agents a sound starting point while making policy checks and uncertainty visible.
