Generative AI Customer Service: A Worked Example
Generative AI customer service, worked through one task: a messy complaint email, a draft reply, where it breaks, and the checks before it ships.
"Generative AI customer service" usually means one of two very different things: a chatbot that talks to the customer directly with no person in the loop, or a model that drafts a reply for an agent to read, edit and send. The second is far more reliable right now, and it is the one this post works through — a single real-looking task, an actual draft, exactly where it goes wrong, and the checks worth running before any of it reaches a customer.
That distinction tracks a real finding, not just caution for its own sake. Microsoft's own research draws an explicit line between a tool being able to assist with a task and a tool being able to own it end to end — a model being useful for drafting a reply is a different claim from a model being safe to send one unsupervised, and the gap between those two claims is exactly where a customer service deployment goes wrong. The Stanford AI Index tracks the same gap more broadly: capability benchmarks keep climbing faster than the reliability checks needed to hand a task over unsupervised.
A worked example: turning a complaint into a draft reply
Example input — the kind of message that actually lands in a support queue, not a clean test case:
hi so i ordered the blue desk lamp on the 14th and it still hasnt arrived, tracking says delivered on the 16th but nothing showed up at my door, i've emailed twice already with no reply this is really frustrating i just want a refund or a replacement asap, order #48291
The prompt worth writing names the fields and the tone, not just "reply to this": "Read this customer message and return two things: (1) a structured ticket summary with order number, issue type, requested resolution and urgency, and (2) a draft reply in a calm, apologetic tone that acknowledges the two unanswered emails, does not promise a specific refund or replacement without confirming the account first, and asks for one piece of information needed to proceed. Do not state a policy unless it is given to you in this prompt."
A usable result from that prompt: a summary line — `Order #48291, non-delivery despite tracking showing delivered, requesting refund or replacement, high urgency, second contact` — and a draft reply that apologises for the delay and the unanswered emails, notes the tracking discrepancy, and asks the customer to confirm the delivery address before offering a resolution. Notice what the draft does not do: it does not invent a refund policy, and it does not promise an outcome the agent has not actually decided on yet. That restraint only happens because the prompt explicitly forbade stating an unconfirmed policy — without that line, a model will often fill the gap with something plausible-sounding instead of flagging that it does not know.
Where this actually breaks
- Policy invention — stating a refund window, a restocking fee or an exception that was never given to the model, confidently and in the same tone as everything else in the reply
- Tone mismatch on genuinely angry or distressed messages, where a template-shaped apology reads as dismissive rather than as actually listening
- Missing sarcasm or an implied threat (a mention of a chargeback, a review, a lawyer) that a person would flag for escalation and a model may draft past without noticing
- Sending a draft that references account details — an address, a card’s last four digits — the model was not actually given and quietly reconstructed or omitted incorrectly
The underlying reason is the same one that shows up everywhere else a model drafts something confidently: a survey of hallucination in large language models documents how a fluent, well-formed answer and a correct one are not the same claim, and the two are indistinguishable from the outside until someone actually checks the specifics. In a customer reply, the specifics that matter most are exactly the ones a policy document owns, not the model.
The checks before any of this goes live
- Give the model your actual policy text as part of the prompt or as a document it reads, and explicitly forbid stating anything not found in it — the instruction used in the worked example above.
- Run a test batch of ten real historical tickets, including at least one angry message and one ambiguous request, and read every draft before trusting the pattern on new ones.
- Keep a human in the loop before anything sends for at least the first weeks of any deployment — a draft an agent edits and approves, not an autonomous reply.
- Set an ongoing spot-check rate once it is live, not just a one-time review at launch. NIST’s AI Risk Management Framework frames exactly this: monitoring an automated system is ongoing work, not a box ticked before launch and then forgotten.
That draft-then-review shape is the same pattern behind most of what actually works with these tools right now. Checking an AI answer when you are not the expert sets out the general habit, and intelligent automation use cases walks through the equivalent worked example for reading an invoice rather than a complaint — the reading-then-checking shape is identical, only the document changes. If a question about this ever applies to Coursium itself rather than a business you run, the support page is where that goes, not a chatbot.
What to do Monday
Pick ten real tickets from last week, and write the two-part prompt shown above — summary plus draft reply, with an explicit rule against stating an unconfirmed policy. Writing a prompt that works on the first try covers the same discipline for any task, and what AI is actually bad at is worth reading before assuming the tool has covered every edge case the test batch happened not to include. If the wider goal is choosing a platform rather than writing one prompt, best AI tools for business and the AI executive assistant category cover the buying decision that usually comes before this one.
The same reading-then-drafting pattern is worth reusing elsewhere once it is working here — an AI report generator applies almost the identical brief-then-check discipline to turning that same support ticket data into a report for a manager, rather than a reply for a customer. PwC's 2026 Global AI Jobs Barometer found a 62% wage premium for workers who use AI tools well, up from 57% the year before — this kind of careful, checked drafting is closer to what that premium is actually paying for than a fully autonomous chatbot is.
Coursium teaches this precise, checked layer directly — writing a request specifically enough that the draft is actually usable, and knowing what to verify before anyone else sees it. Stay ahead of AI by learning the tools on your phone.