Support replies that de-escalate
- support
- tone
An upset customer, a busy agent and a fluent AI is a three-way recipe for a reply that overshoots: over-apologising, over-promising, or arguing a point nobody asked about. The reliable pattern splits the job into three gates — routing, drafting and the policy gate — and keeps the model inside the one it can actually do.
Three gates before send
- Route. Is this a how-to question (an AI first draft is fine), a billing or policy edge (a human decides first), or something involving harm, legal exposure or a public threat (escalate now, draft nothing)? Triage before text.
- Draft with constraints. The facts of the case are fixed; the draft moves inside them. Tone rules, a length cap, and — most important — the list of things that must not be promised.
- Policy gate. Before sending, check every commitment against the actual policy. The model does not know your policy, your margins or your exceptions. You do.
Drafting with constraints
Weak prompt
Write a reply to this angry customer and make them happy
Stronger prompt
Draft a 120-word reply. Facts: order #4821 arrived six days late; our delivery guarantee covers a full refund of shipping only. Acknowledge the frustration once, plainly. Offer the shipping refund and one month of free delivery as goodwill. Do not promise anything else — no product refund, no delivery dates beyond “within three business days”. Tone: calm and direct; no grovelling, no jargon, no exclamation marks.
Why the stronger one works
- “Make them happy” invites invented discounts and hollow promises — the model optimises for the mood, not the policy.
- The strong prompt fixes the facts, names the one allowed remedy, and forbids the rest explicitly.
- A length cap and tone rules keep the draft sendable with light editing instead of a rewrite.
A bad example
Send: “We are so terribly sorry for this experience! We completely understand and to make it right we will refund your order and expedite a replacement immediately.”
A kind, fluent reply that promises two things you may not be authorised to give — and the customer now has it in writing. De-escalation failed in the direction of generosity.
A better example
Send: “You are right to be annoyed — the delay was ours. I have refunded the shipping cost in full and added a month of free delivery to your account. If the package has not arrived by Thursday, reply here and I will investigate the courier case personally, with an answer within three business days.”
Acknowledged once, fixed what policy allows, and the only commitment made is one you can keep.
Why it works
Drafting is language; promising is policy — two different jobs, and flattery blurs them. Constraints keep the reply inside what is true, which is exactly what makes it de-escalating: angry customers are not soothed by warmth, they are soothed by specific, kept promises. Keeping the model on the drafting side of the line also protects the agent: a wrong promise made by a person in an AI-worded email is still a wrong promise by the company. And with the facts frozen and the remedies fixed, the model is genuinely fast — a first draft that already passes the policy gate.
Practice
Build the support reply
A customer's order (#4821) arrived six days late. Your delivery guarantee covers a refund of shipping only, and you are authorised to add one month of free delivery as goodwill. The customer is angry and has mentioned writing a review. You will ask an AI assistant for the reply draft.
Fill the four slots — facts, allowed remedy, tone rules and forbidden promises — then compare with the reference version.
A strong example (self-assessment — not machine-graded)
Facts: order #4821 arrived six days late; courier records show it left the depot on schedule and stalled in transit; this is the first contact about the delay. Allowed remedy: refund of the shipping cost in full, plus one month of free delivery as goodwill. Tone rules: calm and direct, acknowledge the frustration once and plainly, no grovelling, no jargon, no exclamation marks, about 120 words, end with a concrete next step. Forbidden: no refund of the product price, no discount codes beyond the stated goodwill, no delivery guarantees beyond 'within three business days' for any follow-up answer.
Hint
The facts slot is frozen material — the draft may rephrase it, never extend it.
The forbidden list is what keeps a warm draft sendable: name every promise the model might reach for.
Why this is the answer
A constrained draft passes the policy gate by construction. The facts slot stops invention, the remedy slot names the one allowed outcome, the tone rules keep it sendable with light editing, and the forbidden list closes the gaps a model would otherwise fill with generosity. The result is a reply that is warm where it is true and silent where the policy is strict — which is exactly what de-escalates: specific, kept promises.
Transfer
- Templates: keep the constraints block — allowed remedies, tone rules, forbidden promises — ready to paste.
- Reviews and social: reply publicly with facts and a private channel, never with concessions made under visible pressure.
- Escalation cues: legal language, safety, repeat contacts. Route first; drafting can wait three minutes.
- After any AI draft: the policy gate is a human reading every promise, sentence by sentence.
Key takeaways
- Three gates: route the case, draft with constraints, gate every promise against policy.
- Angry customers are soothed by specific kept promises, not warmth.
- The model drafts; the policy decides. Never send a commitment it invented.
Next
The work track keeps growing — ten lessons live now. Keep the reps going on the practice floor, or see where you stand with the skills check. The job-search workflow continues the track with CVs and cover letters. The coding track picks up the same discipline next: What AI can (and can’t) do with code. The placeholder entry covers keeping personal data out of drafts.