Customer service runs on volume. Every hour an agent spends composing a de-escalation response is an hour not spent on the ticket queue. Every refund evaluation processed manually is a delay that increases the customer's frustration. Every customer whose complaint goes unacknowledged in the first hour is a churn signal you're not catching until it's too late.
AI prompts for customer service don't replace the human judgment that resolves complex complaints — but they dramatically compress the time it takes to produce a good response, and they enforce quality standards (one apology maximum, no policy language, resolution-focused close) that inconsistent human output often misses. The eight prompts below cover the full support lifecycle: triage, initial response, de-escalation, refund evaluation, sentiment tracking, onboarding, and proactive status communication. They're built from real support operations patterns — the kind of structured thinking that makes first contact resolution rates go up and handle times go down.
💡 How to use these prompts
Fill every bracket field — [CUSTOMER_MESSAGE], [CANNOT_DO], [AVAILABLE_RESOLUTIONS]. The prompt text itself is the quality control mechanism. When you leave fields blank or vague, the AI fills in defaults that sound generic. For ticket classification prompts, paste the verbatim ticket content — the AI extracts entities, sentiment signals, and escalation flags from the actual words, not your summary of them.
1
Difficult Customer De-escalation Response
Use case: Writing a response to an angry, upset, or hostile customer — without making things worse. This is the prompt for tickets that come in hot: threats, explicit frustration, complaints about repeated issues. The output enforces the "one apology maximum" rule and redirects to solutions instead of dwelling on the problem. For messages containing abusive language, add that context to [ISSUE_SUMMARY] so the tone calibrates correctly — the prompt handles it better when it knows what it's dealing with.
You are an experienced customer service manager trained in conflict resolution. Write a de-escalation response to an angry or upset customer.
Customer message:
[CUSTOMER_MESSAGE]
What the customer is angry about: [ISSUE_SUMMARY]
What you can actually do to help: [AVAILABLE_RESOLUTIONS]
What you cannot do (hard constraints): [CANNOT_DO]
Company name: [COMPANY_NAME]
Write a de-escalation response that: 1) Validates the customer's frustration without admitting liability, 2) Uses the customer's name if provided, 3) Acknowledges the specific problem they described, 4) Pivots to a concrete solution pathway, 5) If you cannot give them what they want, explains why clearly without corporate policy language, 6) Ends with a question that keeps the conversation open and moves toward resolution.
Tone: Calm, warm, firm — never defensive or dismissive
Length: 180–250 words
Do NOT: apologize more than once, repeat yourself, use phrases like 'per our policy' or 'unfortunately there's nothing I can do'
Why it works: The single-apology maximum is the structural constraint that prevents the output from sounding like a company trying to manage a PR problem rather than solve a customer's problem. Over-apologizing signals that the company is more concerned with its own image than the customer's experience — it makes the response about the company, not the customer. The "explain the reason, not the rule" instruction is what makes the hard constraint explanation credible: "I can't refund beyond 30 days because the window closes to protect both sides from processing errors that far back" is a reason. "Per our policy" is a wall.
2
Empathy-First Customer Complaint Response
Use case: Writing a complaint response that takes ownership, explains the issue in plain language, and closes with a concrete commitment — in 150–250 words. The critical input is the [INTERNAL_CAUSE] field: "system issue" is not good enough. The AI needs the actual cause so the explanation is credible rather than boilerplate. For high-value customers, run the output through a voice check before sending — the prompt produces good copy, but a good agent reviewing it produces great copy.
You are a senior customer success specialist writing a response to a customer complaint. Your goal: resolve the issue and retain the customer's trust.
Customer complaint:
[COMPLAINT_TEXT]
Company context:
- Company name: [COMPANY_NAME]
- Product/service: [PRODUCT_NAME]
- What actually happened (internal notes): [INTERNAL_CAUSE]
- What you can offer: [RESOLUTION_OPTIONS]
- Agent name: [AGENT_NAME]
Write a response that: 1) Opens with genuine empathy — acknowledge the specific frustration without generic phrases, 2) Takes ownership of the problem clearly, 3) Explains what happened in plain language (1–2 sentences max), 4) States the resolution with timeline if applicable, 5) Closes with a specific commitment.
Tone: [TONE]
Length: 150–250 words
Format: Prose, no bullet points, no subject line
Do NOT include: corporate jargon, passive voice, hedging language, empty apologies.
Why it works: "Genuine empathy — acknowledge the specific frustration without generic phrases" is the instruction that separates a complaint response that retains trust from one that preserves the relationship. Generic empathy ("We're sorry for the inconvenience") is something a bot would write, and customers know it. Specific empathy ("The checkout failure you hit on Tuesday — when you had a time-sensitive order — is exactly the kind of breakdown that erodes trust, and I'm genuinely sorry it happened on our watch") is something a human who understands the problem would write. The internal cause field is what makes the difference: you can't write specific empathy without knowing what actually happened.
3
Multi-Label Support Ticket Classifier
Use case: Pre-classifying incoming support tickets before routing — for queue management, SLA tracking, and priority escalation. This prompt produces a structured classification with 10 data points per ticket: category, sub-category, priority, sentiment score, urgency signals, routing recommendation, SLA target, key entities, searchable tags, and an escalation flag. Feed this into a batch automation pipeline to pre-classify high volumes. Use the sentiment score to auto-prioritize the queue — tickets scoring 1–2 on sentiment jump the line. The escalation flag is your circuit breaker: review flagged tickets before auto-routing.
You are a senior customer support operations specialist. Classify the following support ticket with precision and completeness.
Ticket content:
[TICKET_CONTENT]
Output the following structure:
**Primary Category:** [Billing / Technical / Account / Product / Shipping / Other]
**Sub-Category:** [Specific topic]
**Priority Level:** [Critical / High / Medium / Low] — with one-line rationale
**Sentiment Score:** [1–5, where 1 = extremely negative, 5 = very positive]
**Urgency Signals:** List any language indicating time pressure, repeated contact, or escalation risk
**Recommended Team:** [Tier 1 Support / Tier 2 Technical / Billing Specialist / Customer Success / Legal]
**Suggested SLA:** [< 1 hour / < 4 hours / < 24 hours / < 72 hours]
**Key Entities Mentioned:** [Product names, order numbers, account IDs, dates]
**Suggested Tags:** [3–5 lowercase tags for search indexing]
**Escalation Flag:** [Yes / No] — flag if this ticket matches escalation triggers: legal threat, media mention, [VIP_ACCOUNT_TIER], or [ESCALATION_KEYWORD_LIST]
Use only the information in the ticket. Do not infer or fabricate details. If a field cannot be determined, write "Unable to determine."
Why it works: The VIP account tier and escalation keyword list are the variables that turn a generic classifier into an operationally specific one. Tickets from enterprise accounts or containing legal/regulatory language require different handling regardless of their technical category — this prompt surfaces that distinction automatically when you feed it the right context. The urgency signal extraction is the feature most ticket classifiers miss: "I've been waiting 3 days" is a routing signal, not just a complaint. The sentiment score on a 1–5 scale (not 1–10) is calibrated to match common CSAT measurement frameworks, so the output feeds directly into your reporting dashboards.
-->
4
Refund Request Evaluation & Response
Use case: Evaluating a refund request and writing the response — in one step. This prompt does two things: it evaluates policy eligibility (fully eligible / partially eligible / not eligible, with the exact policy reason), and it writes a ready-to-send 150–200 word customer email. The business case for exception field is where you recoup high-LTV customers that policy would otherwise lose — a 3-year customer requesting their first refund is a different calculation than a new customer on their third request.
You are a customer service specialist handling a refund request. Evaluate the request and write a response that is fair, policy-compliant, and customer-first.
Refund request:
[REFUND_REQUEST]
Policy context:
- Refund policy: [REFUND_POLICY]
- Order details: [ORDER_DETAILS]
- Customer account status: [ACCOUNT_STATUS]
- Internal notes: [INTERNAL_NOTES]
Step 1 — Evaluation:
**Policy Eligibility:** [Fully eligible / Partially eligible / Not eligible] — cite exact policy reason
**Recommended Resolution:** [Full refund / Partial refund of X% / Store credit / Replacement / Denial]
**Business Case for Exception (if applicable):** If outside policy, is there a business reason to approve?
Step 2 — Write the customer response:
Tone: Empathetic, direct, no corporate-speak
Format: Email body only, 150–200 words
Content: Clear decision with reason, exact resolution terms (amount, timeline), next steps
Why it works: The two-part structure — evaluation first, then response — enforces the discipline of making a decision before writing the copy. Agents who write the response first tend to backfill the rationale to fit the response, which produces inconsistent decisions. By requiring the policy eligibility determination as a separate output, this prompt forces the agent (or the AI acting as the agent) to make an explicit call before writing the customer-facing language. The account status field is the LTV signal: it changes the recommendation for otherwise identical refund requests depending on the customer's history and value.
5
Customer Sentiment & CSAT Predictor
Use case: Predicting customer satisfaction outcomes after a support interaction — to identify churn risk before it surfaces in a cancellation request. The output is a 10-field sentiment report: score, CSAT prediction, churn risk classification, key positive and negative phrase extraction, and a recommended follow-up action. Actionable in under 2 minutes. The "Key Phrases (Negative)" field is your early warning system — phrases like "last time," "already tried," or "considering canceling" are high-churn signals that deserve a same-day review from your CS manager.
You are a customer experience analyst specializing in sentiment analysis. Analyze the following customer interaction and predict the customer's satisfaction outcome.
Customer interaction:
[INTERACTION_CONTENT]
Context:
- Channel: [CHANNEL]
- Product/service: [PRODUCT_NAME]
- Interaction stage: [STAGE]
Deliver a structured sentiment report:
**Sentiment Score:** [1–10, 1 = highly negative]
**Confidence Level:** [High / Medium / Low]
**Primary Emotion:** [Frustrated / Angry / Disappointed / Confused / Neutral / Relieved / Satisfied / Delighted]
**Secondary Emotions:** List up to 3 additional emotional signals
**Key Phrases (Negative):** Quote specific language indicating negative sentiment
**Key Phrases (Positive):** Quote specific language indicating positive sentiment
**Predicted CSAT Score:** [1–5] with brief rationale
**Churn Risk:** [High / Medium / Low] — with one-sentence justification
**Recommended Follow-up Action:** Specific next step
**Agent Performance Signal:** Note any agent behaviors that influenced sentiment
Why it works: The churn risk classification with a one-sentence justification forces a directional call rather than a neutral score. A score of 3.2 with "customer used the phrase 'considering canceling' twice and referenced three prior unresolved contacts" is actionable. A score of 3.2 without context is a number in a dashboard that no one acts on. The agent performance signal is the feedback loop that most support operations miss — if the same agent keeps appearing in negative-sentiment reports with high churn risk, that's a coaching opportunity that scales with the AI analysis, not despite it.
-->
6
First Contact Resolution Response Template
Use case: Writing a complete support response that closes the ticket in one reply — no back-and-forth. First Contact Resolution is the highest-leverage support metric: every 10% improvement in FCR reduces handle time by approximately 15%. The verification question step is the difference between FCR and a false-positive close: if you don't ask the customer to confirm the issue is resolved, they'll reopen the ticket 20 minutes later thinking it wasn't fixed.
You are a Tier 1 support specialist optimizing for first contact resolution (FCR). Write a complete response to the following support request that resolves the issue in a single reply — no back-and-forth.
Support request:
[SUPPORT_REQUEST]
Product: [PRODUCT_NAME]
Resolution steps available: [RESOLUTION_STEPS]
Relevant documentation links: [DOC_LINKS]
Agent name: [AGENT_NAME]
Write a response structured as follows:
1. Greeting and acknowledgment (1 sentence) — address the request specifically
2. Solution — numbered steps, clear and actionable, assumes no technical expertise
3. Verification question — ask one question that confirms resolution without requiring a follow-up ticket
4. Alternative path — if the primary solution doesn't work, what should they try next
5. Closing — brief, warm, specific to the issue resolved
Format: Minimal markdown, readable in plain email. No headers. No corporate sign-off boilerplate.
Length: 200–300 words
Tone: [TONE]
Why it works: The verification question is the step most agents skip because it feels presumptuous — "what if the solution doesn't work and I asked them to confirm it's fixed?" The answer: if the solution didn't work, the customer will tell you. The verification question is not a trap, it's a shortcut to FCR confirmation. An agent who closes a ticket without verifying is guessing at FCR. An agent who asks "Did that resolve the issue for you?" gets a data point. Over time, the aggregate verification data tells you exactly which resolution paths are working and which are producing false-positive closes. Without the question, you have none of that data.
7
New Customer Onboarding Email Sequence
Use case: Designing a 5-email onboarding sequence that drives activation, prevents early churn, and reduces support tickets in the first 30 days. The Day 1 email's only job is to get one action done in under 10 minutes — if you have more than one CTA, you're reducing activation. The activation milestone is the one action that predicts long-term retention: if users complete it, they stay. The churn reasons should come from exit survey data, not guesses — the prompt uses them to preempt the top dropout points.
You are a customer success specialist designing an onboarding email sequence for new customers. Create a 5-email sequence that drives activation, reduces churn in the first 30 days, and prevents common support issues before they occur.
Product context:
- Product name: [PRODUCT_NAME]
- Core value delivered: [CORE_VALUE]
- Primary activation milestone: [ACTIVATION_MILESTONE]
- Top 3 reasons new users churn in first 30 days: [CHURN_REASONS]
- Top 3 features that retain users long-term: [RETENTION_FEATURES]
- Sender name: [SENDER_NAME]
Write 5 emails:
**Email 1 — Immediately after signup (Welcome + Quick Win)**
Subject: | Body: 120–150 words | Goal: First action in < 10 minutes
**Email 2 — Day 3 (Activation Check-in)**
Subject: | Body: 100–130 words | Goal: Drive to [ACTIVATION_MILESTONE]
**Email 3 — Day 7 (Value Reinforcement)**
Subject: | Body: 110–140 words | Goal: Show ROI proof, address top churn reason 1
**Email 4 — Day 14 (Feature Discovery)**
Subject: | Body: 120–150 words | Goal: Introduce [RETENTION_FEATURES][0]
**Email 5 — Day 28 (Check-in + Next Step)**
Subject: | Body: 100–130 words | Goal: Invite feedback, upsell path if appropriate
Tone: [TONE]
Each email: No more than one CTA, plain text format, no images.
Why it works: The one-CTA rule per email is the constraint that most onboarding sequences violate, and it's why Day 1 activation rates are low. When you give new users two options ("explore the dashboard and invite your team"), neither gets done. When you give them one — "complete [ACTIVATION_MILESTONE] — it'll take under 10 minutes" — the activation rate jumps. The Day 28 email is the touchpoint most onboarding sequences skip, and it's the highest-value moment: customers who've made it to Day 28 are approaching the first major churn decision point. A simple feedback invitation at this stage catches churn signals before they become cancellations.
8
Proactive Status Update Message Generator
Use case: Writing outbound status communications to customers affected by a known issue — before they contact support. Proactive communication cuts inbound ticket volume by 30–60% during incidents. This prompt generates two ready-to-use formats: a full email (subject + 150–200 word body) and an SMS/push notification under 160 characters. Both lead with customer impact, not technical explanation. If the issue is resolved, the first sentence says so. If it's ongoing, the ETA is stated clearly.
You are a customer communications specialist. Write a proactive outbound message to customers affected by a known issue or delay — before they contact support.
Situation:
- Issue type: [ISSUE_TYPE]
- Scope of impact: [AFFECTED_CUSTOMERS]
- Current status: [CURRENT_STATUS]
- Root cause (if sharable): [ROOT_CAUSE_SUMMARY]
- What customers need to do (if anything): [CUSTOMER_ACTION_REQUIRED]
- Compensation offered (if any): [COMPENSATION]
- Contact for questions: [SUPPORT_CONTACT]
- Brand name: [BRAND_NAME]
Write the message in two formats:
**Format 1 — Email (subject + body, 150–200 words)**
**Format 2 — SMS/Push Notification (under 160 characters)**
Tone: [TONE]
Requirements:
- Lead with what customers care about (the impact on them), not the technical cause
- Be specific about timelines — no vague "as soon as possible"
- If the issue is resolved, say so in the first sentence
- If the issue is ongoing, state the ETA clearly
Why it works: "Lead with the impact on them, not the technical cause" is the instruction that prevents status updates from sounding like engineering post-mortems. Customers don't care that a database cluster failed — they care that their order is delayed. The SMS/push format under 160 characters forces ruthless prioritization: if you can't say what the customer needs to know in 160 characters, the message has too much in it. For ongoing incidents, the prompt instructs the AI to send updates every 2–4 hours even if status hasn't changed — silence during an active incident is the fastest way to turn a manageable issue into a ticket flood.
Need a custom customer service prompt?
Use the PromptSonar AI Generator to build tailored prompts for your specific support workflows — ticket classification, de-escalation, refund handling, and more.
Try the AI Generator →
💡 Principles for AI-Powered Customer Service
Run every complaint response through a voice check for high-value customers. The AI produces good copy; a good agent reviewing it produces great copy. For enterprise accounts and churn-risk customers, the 90-second review is worth it.
Use sentiment scoring to auto-prioritize the queue. Tickets scoring 1–2 on the sentiment scale should jump the line — they're your highest-risk interactions and the ones where fast response has the most impact on retention.
Send proactive status updates before customers contact you. Every proactive update you send prevents 3–6 inbound tickets. Run this prompt the moment an incident is confirmed, not after the ticket volume spikes.
Feed the onboarding sequence data back into churn analysis. If Email 3 (Day 7) consistently has the lowest open rate, that churn reason is not being addressed adequately — update the sequence to address it more directly.
Browse the full Customer Service prompt library on PromptSonar — 20 production-ready prompts for support teams, CS managers, and customer success specialists. All prompts are bilingual (EN/FR) and ready to copy and customize.
Also worth exploring: the Business AI prompts guide covers customer success strategy and retention frameworks, and the Article Writing prompts guide includes prompts for customer-facing content and knowledge base articles.