The pull quote
“The support team that uses AI well is the one that treats AI as a draft generator for the agent, not a reply generator for the customer. The minute the customer is talking to AI without a human in the loop, the brand pays the bill.”
Why do most AI-driven support replies make customers angrier?
Support customers can tell when they are talking to a bot. The 2025 Forrester customer-service research found that satisfaction drops sharply when customers detect AI-only handling, particularly when the issue is non-trivial. The pattern is consistent across industries: AI front-line handling works for simple lookups (order status, password reset, hours of operation), but the harder the question, the more value the human-in-the-loop adds, and the more cost the AI-only deflection produces. Brands that ship AI as the visible reply layer often see ticket-volume drop while NPS drops faster.
The seven prompts below are designed for a different deployment: AI as the agent's draft assistant, not the customer's reply machine. The agent reads the inbound ticket, runs one of these prompts, gets a draft response in the brand's voice with the right escalation flag, edits it, and sends. Time saved per ticket: 3-5 minutes. Quality preserved. Brand voice intact. The deployment model is what separates AI that helps from AI that costs.
⚠️ Support-AI failure modes to know
The hallucinated-policy reply. AI confidently states a return policy you do not have. The tone mismatch. A customer in real distress gets a chipper “Hi there!” reply that reads as gaslighting. The deflection loop. AI sends increasingly generic responses until the customer escalates publicly on social media. The constraints in the prompts below block each of these. Before any draft ships, run the editing pass from How to Edit AI Out of Your Writing.
What are the seven support prompts?
1. Triage the ticket
Here is the customer's purchase history / account summary: [paste].
Triage this ticket. Tell me:
– The single most likely category (billing / shipping / product defect / account access / feature request / complaint / other)
– The single most likely intent (refund / fix / answer / acknowledgment / escalation)
– The emotional read (calm / frustrated / angry / panicked / formal)
– The right responder (front-line / senior agent / supervisor / engineering / legal)
– The SLA tier (standard / priority / urgent based on customer value, issue severity, public-visibility risk)
For each, cite the specific sentence from the ticket that signaled this.
Why this works: Triage is where AI saves the most time without ever touching the customer. The “right responder” output is the routing decision that humans get wrong under volume; a misrouted ticket loses 20-40 minutes. The “cite the specific sentence” output is the audit trail that makes the triage defensible if it turns out wrong.
2. Draft the reply (with escalation gate)
Here is the triage you just did: [paste from prompt 1].
Here is our brand voice brief: [paste, or “professional, warm, plain-spoken, no exclamation marks”].
Here is our policy on this issue type: [paste, even if it's 3 lines].
First, decide: should this ticket go to a human before any reply ships?
– If YES (issue is legal / safety / refund > $X / customer mentioned cancellation): output STOP, do not draft. Explain why.
– If NO: draft a reply.
The reply must:
– Acknowledge the specific issue in the first sentence (not a generic “thanks for reaching out”)
– State what we will do next, with a concrete time frame
– Avoid policy language the customer would not understand
– Match the brand voice brief
– End with an opening for the customer to push back if our read of the issue is wrong
– 5-7 sentences total
Why this works: The STOP gate is the single most important constraint in any support-AI prompt. It is the line between AI-as-assistant and AI-as-liability. The “opening for the customer to push back” sentence is the part that distinguishes a good support reply from a closed-loop one; it signals the agent is listening, not just responding. The 5-7 sentence ceiling stops the AI from over-explaining the policy.
3. The angry-customer de-escalation reply
What we got wrong (from our side, even if partially): [paste, be specific].
What we can offer them, in order of escalation: [list 3 options from minimum to maximum].
Draft a de-escalation reply. Rules:
– First sentence: name what we got wrong, specifically. No “we're sorry you feel this way.”
– Second sentence: state what we are doing about it for THIS customer (not policy changes; concrete action for them)
– Third sentence: offer one of the 3 options, the appropriate one for the severity
– Fourth sentence: invite their response
– Do NOT use: “Unfortunately,” “Per our policy,” “As stated in our terms,” “We understand your frustration”
– Do NOT promise anything not in the options list
Output 3 versions: cooler / matched / warmer than the customer's tone. Tell me which one to send.
Why this works: “We're sorry you feel this way” is the line that turns angry customers into public complaints. The forbidden-phrase list is the most actionable part of this prompt. The three-tone variations let the agent pick based on the specific customer's temperature rather than defaulting to the corporate-warm tone that often feels insincere to the angry customer.
4. The “fix this and never speak of it again” reply
The fix on our side: [name the action].
Draft a short reply that:
– Confirms we fixed it (one sentence)
– Names the specific thing fixed (not “the issue you reported”)
– Closes the loop without inviting further conversation
– 2-3 sentences total
– No upsell. No survey ask. No “is there anything else I can help with?”
Why this works: Most simple-fix replies are too long. The customer reported a small issue, you fixed it, and they want acknowledgment, not a follow-up conversation. The “no is there anything else” line is the constraint that respects the customer's time. This prompt is one of the highest time-savers in support volume terms because so many tickets actually deserve a 2-sentence reply.
5. The “this is actually complicated” reply
Why it is complicated: [paste, be specific].
What we need from them to make progress: [list, specifically].
Draft a reply that:
– Names the complexity (the customer should know this is not a simple lookup)
– Asks for what we need, one item at a time, with reasons
– Sets expectation on the timeline (concrete, not “soon”)
– Is warm but not breezy
– 6-8 sentences total
– Includes a “if I've misunderstood the question, here is what I'm hearing” check
Why this works: Complicated tickets get hurt by AI in two ways: oversimplification (“the answer is just…”) and over-formalization (“per our internal review process…”). The “if I've misunderstood” check is the single most useful sentence in complicated-ticket replies. It surfaces the misalignment before it costs three more rounds of back-and-forth.
6. The post-incident apology + status
The customer affected: [paste their account context + what they were trying to do during the window].
What we are doing about it for them specifically: [name the action: credit / refund / priority support / etc].
Draft the proactive reply. Rules:
– Opens with what happened, in plain words, not “we experienced a service disruption”
– Names how it affected THIS customer specifically (not how it affected everyone)
– States what we are doing for them (not future improvements; what is happening now)
– Tells them what to expect next and when
– Closes with the direct line if they need anything else
– 5-6 sentences
– No “Our team is working to ensure this does not happen again” (only say it if there's a specific change, then say what the specific change is)
Why this works: Post-incident replies have to walk the line between transparency and over-disclosure. The “name how it affected THIS customer specifically” line is what turns a templated incident notice into a personalized acknowledgment. The “no we are working to ensure” rule is the part that prevents the apology from sounding hollow; if you have a specific change to name, name it, and if you do not, do not pretend you do.
7. The end-of-shift handoff note
For each ticket, summarize for the incoming agent:
– Status (waiting on us / waiting on customer / blocked on engineering / etc)
– What was promised and by when
– The emotional read (so the next agent knows what tone to match)
– The next action and the right time to take it
– Any landmines (things that might escalate if mishandled)
Order by urgency (most-likely-to-escalate first).
Skip pleasantries; this is for the incoming agent, who has 10 minutes to read it.
Why this works: Handoff notes are where support teams lose continuity. Tickets dropped between shifts become escalations the next day. The “ordered by urgency” output is the structural difference between a handoff that protects coverage and a handoff that just lists tickets. The “landmines” output is the unwritten knowledge the outgoing agent has that AI cannot fabricate but can format usefully when supplied.
What is the worst thing you can do with AI in support?
- Let AI talk directly to customers without a human in the loop on non-trivial tickets. Simple lookups (order status) can be automated. Anything that requires judgment, empathy, or policy interpretation should never ship without agent review.
- Skip the escalation gate (prompt 2). The STOP rule for legal / safety / refund-threshold / cancellation tickets is the single most important safeguard. Removing it to save time is how customer-service AI becomes a liability event.
- Let AI invent policies. Always include the actual policy in the prompt. AI without a policy attached will confidently state a return window, refund threshold, or escalation rule that does not exist.
- Use AI to write apologies without naming what went wrong. “We apologize for any inconvenience” is the corporate-AI signature. Specific naming of the mistake is the de-escalation; the generic apology is the escalation accelerant.
- Forget that the customer can paste your reply on social media. Every reply ships to a public-comms audit. Treat each draft as if it were going on the screenshot quote-tweet.
What if your team handles hundreds of tickets a day?
Save the seven prompts as Claude skills, one per situation. Bundle them as the support-stack plugin that an agent invokes from the ticket interface with one shortcut. Per-ticket time saved: 3-5 minutes on average; some triage cases save 15. The plugin pays for itself within the first week for any team with a queue depth of more than 50 tickets a day. The bigger structural shift is qualitative: agents stop spending time on triage decisions and start spending it on the judgment calls those tickets actually need.
📊 The Prompt-to-Workflow Ladder
Tier 1: the prompts (this post). Tier 2: the skill (one per ticket type). Tier 3: the plugin (support-stack bundle). Tier 4: the workflow (triage fires automatically when a ticket lands, draft is queued for agent review). When to climb →
What are common questions about AI in customer support?
Should we tell customers we use AI for drafts?
Most companies do not need to disclose draft-level use; the agent is still the one sending the reply. Where disclosure matters is automated chatbot front-line interactions, where the customer should know they are not yet talking to a human. The deception cost in undisclosed bot interactions is high; the cost in undisclosed draft assistance is low.
What kinds of tickets should never touch AI?
Anything legal (subpoena, regulatory complaint, ADA accommodation request), anything safety-related (product injury, security breach affecting their account), anything involving a death or grief, and anything where the customer has mentioned cancellation as a bargaining chip. These get routed to humans who handle them in their own voice; AI involvement at any stage creates risk that outweighs the time savings.
How do we keep AI from drifting away from our brand voice?
Maintain a 1-page voice brief and include it in every prompt as a constraint. Refresh the brief every 6 months by analyzing the actual agent replies that earned the best CSAT scores. Those are the patterns to mirror.
Will AI handle CSAT surveys for us?
It can summarize them in bulk, surface themes, and draft the response to negative survey feedback. It should not generate the survey questions themselves (those need product / leadership ownership), and it should not auto-reply to negative CSAT without a human pass. That is the moment a customer is closest to public complaint and the moment a templated AI response does the most damage.
Where do these prompts come from?
They are the support-team section of the larger AI Prompt Library. The Library has over 500 prompts across 33+ categories with three difficulty levels, including the full operations stack: support, ops, recruiting, finance close, vendor management, and customer-success outreach.
Sources to read next?
- Forrester customer-experience research
- Harvard Business Review on customer service
- Zendesk CX Trends Report
- Gartner customer service research
- Customer service (Grokipedia)
✏️ Before any draft goes out
A reply that pattern-matches as AI loses customer trust faster than a slower human reply. Our pre-send editing pass for any agent-facing draft: How to Edit AI Out of Your Writing → The full 29-pattern catalog from Wikipedia's “Signs of AI writing” guide is documented there. For support teams editing AI drafts at volume, the open-source Humanizer skill for Claude Code is the fastest way to run the pass on every draft.
The AI Prompt Library · $39
over 500 tested prompts including the full operations section
The seven support prompts above are a free preview. The full Library has the operations stack: refund-decision frameworks, churn-save scripts, voice-of-customer summarization, agent-coaching feedback prompts, and the integration patterns for the most common support platforms.
1-on-1 Claude Crash Course with James · $75
Train your support team on the support-stack workflow
A focused 1-hour session. We pick your two highest-volume ticket types, build skills for each, and your team walks out with the prompt structure they can use on tomorrow's tickets. Fastest path from prompt to production-ready workflow.
Get Smarter About AI Every Morning
Free daily newsletter. Built for people who want to use AI well, not chase every model.
Free forever. Unsubscribe anytime.
You may also like
- Best Claude Prompts for Sales — cold outreach + objection scripts
- Best AI Prompts for Email Writing — the everyday inbox workhorse
- Best AI Prompts for Meeting Notes — decisions that survive the meeting
- How to Edit AI Out of Your Writing — strip AI patterns before you publish
- Prompt to Workflow: The AI Ladder — when to climb from prompt to workflow