The quality of your AI customer service depends almost entirely on the instructions you give the AI, not on which platform you choose. A vague prompt like "help customers with their questions" produces generic, often useless responses. A specific prompt with your company policies, tone guidelines, and clear escalation rules produces responses your customers would not know came from an AI. This guide includes eight copy-paste prompts for the scenarios that come up in almost every support operation, with notes on what each element is doing and how to adapt it.
Before the Prompts: The One Thing Most Teams Skip
Every support prompt should start with a base system prompt that sets the context for every interaction. Think of it as the onboarding document for your AI agent. Without it, every prompt you write has to re-establish context from scratch.
Here is a base system prompt you can use as your starting template:
You are a customer support agent for [Company Name], a [brief description of what you sell, e.g., "direct-to- consumer skincare brand focused on sensitive skin"]. NAME: Use your name [Sam/Alex/Jordan] in first messages. POLICIES: - Returns: 30 days from delivery, unused condition. Free return label on orders over $60. - Shipping: Standard 5-7 business days, free over $55. Expedited 2-3 days, $12 additional. - Refunds: 3-5 business days after return is received. - Exchanges: Available for size/variant, same product only. TONE: Warm and direct. Acknowledge the issue first, then solve it. Do not start responses with "I understand your frustration" -- it sounds scripted. Be specific instead. NEVER: Invent tracking numbers, promise delivery dates you cannot confirm, or discuss competitor products. ESCALATE to a human agent if: - Customer requests it explicitly - Order value is over $300 - Customer has contacted us 3+ times about the same order - Customer mentions legal action or social media threats
Paste this into your AI platform's system instructions, fill in your specifics, and all eight prompts below will work on top of it.
Prompt 1: Damaged or Defective Item
This is one of the highest-stakes support scenarios. The customer is already frustrated and the outcome directly affects your review score and retention rate.
A customer received an order where the item arrived damaged. The outer packaging was intact. They sent a photo confirming the product is broken. Write a support reply that: 1. Opens with specific acknowledgment of THEIR situation (not generic "sorry for the inconvenience") 2. Takes full responsibility without making excuses 3. Offers TWO clear resolution options on the same line: (a) priority replacement shipped within 1 business day (b) full refund processed within 3-5 business days 4. Asks one question: do they need to return the damaged item, or can they keep or discard it? 5. Sets a clear timeline for whichever option they choose 6. Signs off with your support agent name Keep under 130 words. Tone: human and accountable. Do not use the word "unfortunately."
Prompt 2: Late or Missing Delivery
A common scenario that frustrates customers most when the response is vague or deflects blame onto the carrier.
A customer ordered 11 days ago. Shipping was quoted as 5-7 business days. Tracking shows the package has been at a regional sorting facility for 5 days with no movement. Write a response that: 1. Names the specific discrepancy (11 days vs 5-7 quoted) without minimizing it 2. Shares the specific tracking status you can see 3. States clearly what you are doing right now (opening a carrier claim or trace request on their behalf) 4. Gives a specific callback window: "I will have an update for you by [tomorrow at 3pm ET]" 5. Offers proactive compensation: 15% off their next order for the trouble, no code needed, applied to their account 6. Includes your direct extension or ticket number Tone: proactive and accountable. Avoid "the carrier is responsible" framing -- you own the customer experience.
Prompt 3: Refund Request Outside Policy Window
The goal here is not to say no. It is to find a path that retains the customer and still protects your margins.
A customer purchased 48 days ago, outside our 30-day return window. They say the product "did not work as expected" but this is their first contact with support. No prior tickets. Write a response that: 1. Does NOT open with a policy citation -- that shuts down the conversation immediately 2. Asks one genuinely curious follow-up question about what did not work (this opens the door to a solution) 3. Mentions your policy once, briefly, as context not as a verdict 4. Offers an alternative to a full refund: store credit, an exchange, or a replacement if there is a defect 5. Leaves the door open for a human agent to review exceptions if the alternative is declined Goal: retain a customer who is about to churn. Tone: genuinely interested, not defensive.
Prompt 4: Handling an Angry or Escalating Customer
When a customer is emotional, solving the problem comes second. Acknowledging their experience comes first. A response that skips to solutions without validation usually makes things worse.
A customer sent this message: "I have been waiting 3 weeks for my order and nobody has helped me. This is absolutely unacceptable. I am going to dispute the charge with my bank and leave reviews everywhere. Fix this NOW." Write a de-escalation response that: 1. Validates the experience with specificity, not scripts. Do NOT write "I understand your frustration." Write something human that acknowledges 3 weeks is too long. 2. Takes ownership without making excuses 3. States ONE immediate action you are taking right now 4. Gives them control: offer 2 concrete options and let them choose 5. Provides a direct contact method (named agent, ticket number, or phone) so they do not feel lost in a queue 6. Does NOT mention the bank dispute or review threat -- do not amplify that path Tone: calm, genuine, solution-focused.
Prompt 5: Post-Resolution Review Request
The best time to ask for a review is immediately after a positive interaction, while the good feeling is fresh. A templated ask is obvious and forgettable. A specific one works.
A customer just had their issue fully resolved. Their order arrived late but we sent a replacement and they replied saying they received it and are happy with the product. Write a follow-up message that: 1. References their specific situation (late delivery, replacement sent) -- do not write a generic closing 2. Expresses genuine appreciation for their patience 3. Asks for a review using natural language, not corporate language. Link: [your review platform URL] 4. Offers a thank-you discount: 10% off their next order with code THANKS10 (no purchase minimum, 30-day expiry) 5. Keeps it to 4 sentences maximum 6. Signs off with the agent name, not "The Team" Tone: personal and warm, not transactional.
Prompt 6: Proactive Delay Notification (Batch Outreach)
This prompt is for sending an outbound email to customers whose orders are delayed before they contact you. Proactive communication dramatically reduces inbound ticket volume.
Draft an outbound email to customers whose orders will be delayed an additional 4-6 business days due to a supplier fulfillment issue. These customers have not yet contacted support. Their original delivery window was 5-7 days. Subject line: must be honest and specific. Not "An Update on Your Order" -- something that tells them immediately what the email contains. Email body must: 1. Apologize before explaining (apology first, always) 2. State the delay clearly: original estimate, new estimate 3. Explain the cause in one sentence without excessive detail 4. Offer automatic compensation: 15% off their next order, no code, already applied to their account 5. Offer a cancellation path for customers who need the item by the original date, with a clear refund timeline 6. Close with a first-name sign-off from the CX team lead, not a department name Tone: honest, accountable, and generous.
Prompt 7: Escalation Handoff to Human Agent
When the AI cannot resolve an issue, the handoff to a human should feel seamless. The worst experience is a customer who has to repeat their whole story to a new agent.
A customer has been chatting with the AI for 9 minutes about a billing dispute involving two merged accounts and a missing loyalty points balance (approximately $45 in value). The AI attempted two solutions and neither worked. The customer is still patient but showing signs of frustration. Write the AI handoff message that: 1. Summarizes the three key issues briefly so the customer does not have to repeat themselves to the human agent 2. Introduces the human agent by first name (agent name: Morgan) 3. Gives a specific wait time (estimated under 3 minutes) 4. Confirms that Morgan can see the full chat history 5. Does NOT use "I apologize for the inconvenience" 6. Keeps it to 4-5 sentences Tone: warm, organized, reassuring.
Prompt 8: CSAT Survey After a Multi-Touch Issue
CSAT surveys sent after difficult interactions are more valuable than post-routine ones. The phrasing matters a lot: corporate survey language kills response rates.
A customer had a billing error that required 3 separate contacts over 4 days to resolve. The issue is now fixed and the correct charge is reflected on their account. Send a CSAT survey request. Write an email with: Subject: a subject line that does not say "survey" or "feedback request" -- something that makes them want to open it Body: 1. Acknowledge that this took longer than it should have (do not pretend it was smooth) 2. Thank them for their patience specifically 3. Ask ONE rating question: how would they rate the overall resolution (1-5 scale) 4. Ask ONE open text question that gives you actionable data: not "how can we improve?" but something specific like "what would have made this faster to resolve?" 5. Keep the entire email under 100 words 6. Sign with a real name, not "Customer Experience Team" Tone: human, a little vulnerable, not corporate.
When to Use Which Prompt Type
| Scenario | Primary Goal | Key Element | Expected Outcome |
|---|---|---|---|
| Damaged item | Retain and recover | Two clear resolution options | Replacement or refund, customer retained |
| Late delivery | Reduce escalation | Specific timeline and proactive compensation | Reduced churn, ticket closed same day |
| Out-of-policy refund | Find middle ground | Clarifying question before policy citation | Exchange or credit accepted, customer retained |
| Angry customer | De-escalate first | Specific acknowledgment before solutions | Tone shift, resolution path opened |
| Review request | Capture social proof | Specific context, not generic ask | Higher response rate than templated requests |
| Proactive delay notice | Prevent inbound tickets | Honest subject line, compensation upfront | 30-50% reduction in delay-related tickets |
| Escalation handoff | Seamless transfer | Issue summary so customer does not repeat | Human agent can act immediately |
| CSAT after trouble | Actionable feedback | Specific question, not generic "how did we do" | Higher response rate, more useful data |
Adapting These Prompts for Your Brand
The prompts above work as written for a general e-commerce context. To adapt them for your brand, make three types of changes. First, update the policy details (return window, shipping times, compensation amounts) to match your actual policies. Second, adjust the tone descriptor to match your brand voice: "professional and reassuring" for a medical or financial brand, "casual and energetic" for a consumer lifestyle brand. Third, add any escalation triggers specific to your business, such as a specific order value threshold or complaint type that always needs a human review.
Save a copy of each adapted prompt with a version date. When your policies change, update the prompt. Stale prompts are one of the most common causes of chatbot accuracy drift over time.
Frequently Asked Questions
What AI platform should I use these prompts with?
These prompts work with any conversational AI platform: ChatGPT (GPT-4o or GPT-4o mini), Claude 3.5 Sonnet or Haiku, Gemini 1.5 Flash, or any platform that accepts a system prompt and a user message. For high-volume support automation, Claude 3 Haiku and Gemini 1.5 Flash offer the best cost-quality balance for routine support scenarios at roughly $0.00025-$0.000075 per 1,000 input tokens.
How do I keep the AI from making up policies or guarantees?
Include your exact policies in the base system prompt and tell the AI explicitly to give ranges for anything that varies by case. Add a line like "Never guarantee delivery dates you cannot confirm from the tracking system" and "Never quote a refund amount unless it appears in the customer's order record." The more specific your prohibitions, the less room there is for the AI to improvise in ways that create liability.
Should I tell customers they are talking to an AI?
In most jurisdictions, you are required to disclose if asked directly. The better practice is to be upfront about it in the chat window itself, such as "Chatting with our AI assistant, Maya" rather than implying it is a human named Maya. Customers who know they are talking to AI are not necessarily more frustrated; they are frustrated when the AI fails them and they cannot get to a human easily. Make the human escalation path obvious.
How do I measure whether my AI support prompts are working?
Track three metrics: first-contact resolution rate (issues resolved without escalation or follow-up), CSAT score after AI-handled conversations versus human-handled ones, and escalation rate. A well-configured AI support setup should resolve 55-70% of tickets without human involvement. If your escalation rate is above 40%, the prompts or the knowledge base need more specific information. If CSAT on AI interactions is more than 10 points below human interactions, the tone or empathy language needs work.
Can these prompts replace my human support team?
No, and the goal should not be to replace them. AI handles the routine, the repeatable, and the off-hours volume well. Human agents handle the complex, the emotional, the high-value, and the unusual cases better. The right setup uses AI to handle 60-70% of volume so human agents can focus on the 30-40% of interactions that actually require human judgment, empathy, or authority to resolve.
Measuring and Improving Over Time
Run each of these prompts on 10-20 real tickets from your support queue before deploying them. Compare the AI-generated responses to what a good human agent would say. Where the AI response is generic, add more specificity to the prompt. Where it is too long, add a word limit. Where it misses an emotional beat, adjust the tone instruction.
Good prompts get better through iteration. The first version you deploy will not be the best version. The version you have after six months of refinement based on real outcomes will be significantly better and will represent genuine institutional knowledge about how to handle your customers well.