---
title: "Write a reply that hands a case to a human agent"
description: "AI prompt and skill for support teams: a safe customer reply and an internal handoff note with reason, urgency and steps already tried."
url: https://atulsingh.co.in/ai-playbooks/handoff-reply-to-human-agent
skill: https://atulsingh.co.in/ai-playbooks/handoff-reply-to-human-agent/SKILL.md
function: support
roles: ["Customer support manager", "Support team lead", "Customer support agent", "Operations manager"]
industries: ["ecommerce", "saas", "financial-services", "hospitality"]
published: 2026-11-15T06:33:41.919+00:00
updated: 2026-11-15T15:16:16.438571+00:00
---

# Write a reply that hands a case to a human agent

This playbook helps a support team write a customer reply and an internal handoff note in one step. You paste the customer message, the policy and your escalation rules. The AI suggests whether a person is needed, writes a reply that promises nothing the policy does not allow, and records the reason, urgency and what was already tried.

**Time saved:** about 5 minutes per escalated case

## What to gather

- The customer message, pasted as received, with details you do not need (full address, card or bank details) removed.
- The policy text that applies to this case, such as returns, refunds, delivery or warranty, copied from your own document.
- Your escalation rules: when an agent must take over, who receives the case and how urgent each type is.
- Case facts from your helpdesk or order system: dates, order value, earlier contacts and anything already tried.
- The channel (email, WhatsApp or chat), the tone you want and how you sign off.

## Steps

1. Collect the customer message, the policy, the escalation rules and the case facts. Remove personal data the AI does not need, such as phone numbers, full addresses and payment details.
2. Paste each item into the matching field of the prompt below and run it in your company-approved AI workspace.
3. Read the Decision line first. Check it against your escalation rules yourself. If the AI says no handoff is needed, confirm that no rule applies.
4. If more than one rule applies, write down the owner each rule names. Decide who leads the case and who else needs to know. Put that at the top of the handoff note. The prompt does not always do this for you.
5. Read the reply for promises. Look for any refund, replacement, exception, compensation or date. Delete any that your policy text does not allow for this case.
6. Check that the reply tells the customer who will contact them next. If they asked for something by a set day, check that the reply does not agree to it unless the policy allows it.
7. Check the handoff note against the case record. Look for dates the AI worked out, such as 'tomorrow' turned into a calendar date. Look for policy readings that are not in your policy text. Mark these [ASSUMED] or delete them, and add facts the AI could not know.
8. Send the reply, paste the note into the helpdesk ticket and assign it to the lead owner.
9. If the AI keeps missing a type of case, improve the wording of your escalation rules and run the prompt again.

## The prompt

```text
You are helping the customer support team at {{business_name}}. Read the customer message, decide whether a human agent must take over, then write a reply to the customer and an internal handoff note.

Customer message:
{{customer_message}}

Case details and what has already been tried:
{{case_details}}

Policy that applies:
{{policy_text}}

Escalation rules for human agents:
{{escalation_rules}}

Tone and channel for the reply: {{reply_style}}

Rules
- Use only the policy, rules and case details above. Do not invent order facts, dates, names or policy.
- Do not promise a refund, replacement, exception, compensation or timeline unless the policy text allows it for this exact situation. If the policy is silent, or the customer asks for more than it allows, say that a colleague will review the case. Do not give a date unless the policy gives one.
- Do not admit legal fault or argue with threats. Stay calm and respectful.
- If no escalation rule applies, still write the reply and state that no handoff is needed.
- Mark anything you had to assume with [ASSUMED].

Output in this order
1. Decision: Handoff needed or No handoff needed, and the escalation rule that applies, in a few words.
2. Reply to customer: under 150 words, plain language, using the customer's name. Acknowledge the specific problem, say what happens next and who will do it, and ask only for information the agent truly needs.
3. Handoff note for the agent (skip if no handoff is needed):
Reason:
Urgency: High, Medium or Low, with one line on why
Customer wants:
Already tried:
Policy points that apply:
What the customer has been told:
Suggested next step for the agent:
Open questions:

Before finishing, check the reply against the policy line by line and remove any promise the policy does not support.
```

### Fill in

- `{{business_name}}`: Your business name and city (for example: Kaveri Home Appliances, Pune)
- `{{customer_message}}`: The customer's message, pasted as received (for example: Hi, I ordered a mixer grinder (order KHA-20418) and the jar arrived cracked. I am writing nine days after delivery because I was travelling. I want a full refund and a new jar sent by tomorrow. If this is not sorted I will go to the consumer forum. Rohan Mehta)
- `{{case_details}}`: Facts from your order or helpdesk system, including what has already been tried (for example: Customer: Rohan Mehta. Order KHA-20418, Rs 6,499, paid by UPI, delivered 6 Nov 2026. Today is 15 Nov 2026. First contact on this issue. No photo received yet. Already tried: checked the order and delivery record. No earlier reply sent.)
- `{{policy_text}}`: The policy that applies, copied from your document (for example: Damaged on delivery: report within 7 days of delivery with a photo. After inspection by the returns team we replace the item or refund to the original payment method. Refunds take 5 to 7 working days after the returned item reaches us. Reports after 7 days need manager approval. Only the returns team confirms replacement dispatch dates.)
- `{{escalation_rules}}`: Your rules for when a human agent must take over (for example: Hand to a human agent when: the customer mentions legal action, a consumer forum or a regulator; the request falls outside the policy window; the order value is above Rs 5,000 and a refund is requested; the customer has contacted us three times on the same issue; there is any injury or safety concern. Legal mentions go to the support manager and are rated High. Outside-window requests go to the returns team lead.)
- `{{reply_style}}`: Channel, tone and sign-off for the reply (for example: Email, warm and brief, signed Kaveri Support Team)

## Check the result

- Search the reply for refund, replace, exception, compensation and any date or time. Each one appears only if your policy text allows it for this case.
- The Decision line names each escalation rule that applies, or states clearly that none applies.
- When several rules apply, the note says which owner leads and who else needs to know.
- The handoff note has all eight headings filled in: reason, urgency with a one-line why, what the customer wants, what was tried, policy points, what the customer was told, next step and open questions.
- Every fact in the note can be found in the customer message, the case details or the policy text. Worked-out dates, inferred policy limits and invented next steps are marked [ASSUMED] or removed.
- The reply names the customer's actual problem, says what happens next and who will contact them, and stays under 150 words.
- The reply makes no admission of legal fault and does not argue with a complaint or threat.
- A second agent could pick up the ticket from the note alone without asking the customer to repeat anything.

## Watch out for

- Over-promising: the AI may try to calm the customer by offering a refund or a date. Check every commitment in the reply against the policy and delete any the policy does not allow.
- Several rules at once: a legal mention, a late report and a high order value can all apply to one case, each with a different owner. The AI tends to list them together. A person should decide who leads so the case does not sit between two teams.
- Quiet assumptions: the AI may turn 'tomorrow' into a date or read a limit into the policy that is not written there. These look like facts in the note, so check them against the source.
- Personal data: customer messages can hold phone numbers, addresses and payment details. Remove what the AI does not need, use an approved workspace, and follow your obligations under India's DPDP Act or GDPR. This is general guidance and not legal advice.
- Wrong urgency: the AI judges from the text you give it. If a safety issue or legal threat is missing from your escalation rules, the AI will not flag it. Keep the rules complete and current.
- Rubber-stamping: if agents send replies without reading them, errors reach customers. Ask reviewers to check four things each time: the decision, the lead owner, the promises and the facts in the note.
- Outdated policy: the AI follows the policy you paste. Paste the current version each time, not an old copy from a chat history.

## Industry notes

- **ecommerce:** Damaged items, late deliveries and refund requests are the usual handoff triggers. Order value, days since delivery and payment method are the facts agents most often need in the note.
- **saas:** Outages, billing disputes and data requests need a human quickly. Include the plan, the account owner and what the customer has already tried in the product.
- **financial-services:** Complaints, disputed transactions and anything touching account security should go to a person with the right authority. Keep account numbers and identity details out of the prompt.
- **hospitality:** Booking changes, billing disagreements and guest complaints often depend on a stay date. Put the date and the cancellation terms in the case details so the reply does not promise a rate or room.

## Before and after

**Before:** An agent reads the message, looks up the policy, decides whether to pass the case on and writes a reply. The handoff note is often a single line, so the next agent asks the customer the same questions again. Replies sometimes promise a refund or date that the policy does not allow.

**After:** The agent pastes the message, policy and rules, and gets a decision, a draft reply and a structured note in one pass. The agent checks the decision, settles who leads the case and checks the promises, then sends. The next person can pick up the case from the note without going back to the customer.

## FAQ

**Can the AI decide on its own that a case needs a human?**

It suggests, and your escalation rules guide it. A person should confirm the decision before the reply goes out, especially when the AI says no handoff is needed. Cases involving safety, legal threats or large amounts should always be read by a human.

**What if the policy does not cover the customer's request?**

The prompt tells the AI to say that a colleague will review the case and to give no date. The handoff note lists the gap under open questions. The agent or manager then decides, and you may want to add the new case to your policy.

**What if a case matches more than one escalation rule?**

The AI will usually list every rule that applies. Before sending, a person picks the lead owner and notes who else needs to know. If this happens often, add a line to your escalation rules that says which team leads when rules overlap.

**Is it safe to paste customer messages into an AI tool?**

Only if your company has approved the tool and you share the minimum needed. Remove phone numbers, full addresses and payment details first. Check your obligations under the DPDP Act or GDPR with someone qualified.

## Atul's note

I see handoffs go wrong when the customer is promised something the next agent has to undo. A good reply says what happens next and who does it, and leaves out refunds and dates the policy does not allow. The note should save the agent from asking the customer the same questions again. A person reads both before sending.

## Tested on

- Claude Sonnet 5.5, 2026-11-15: pass
- GPT-6, 2026-11-15: pass
- Gemini 4, 2026-11-15: fail

[This page](https://atulsingh.co.in/ai-playbooks/handoff-reply-to-human-agent) · [Markdown copy](https://atulsingh.co.in/ai-playbooks/handoff-reply-to-human-agent.md) · [SKILL.md](https://atulsingh.co.in/ai-playbooks/handoff-reply-to-human-agent/SKILL.md)
