Triage your support inbox and draft replies that a person approves before sending
An agent reads each new support email, labels it by type and urgency, finds the answer in your help docs and writes a draft reply with the source it used. A person reads and sends every reply. Do not let the agent send on its own. Public cases at Air Canada and Cursor show support bots inventing policies that customers acted on, and a tribunal held the company responsible for what its chatbot said.
Every new support email gets a label, a priority and a draft reply grounded in your docs, and a person sends each reply after reading it.
- Difficulty
- Intermediate
- Time
- Half a day to set up, then minutes per batch of tickets
- Sources
- 7 · 7 independent
- Updated
- 2026-10-13
The prompt
Fill the blanks here, or leave them for your agent to ask. Pick the tab for your agent. The hand-off button gives any agent this page as markdown.
You help me run support for [company], which sells [product]. You triage and draft. You never send. Sources you may use: [help center URL], [policy doc: refunds, cancellations, SLAs], [known issues list]. If the answer is not in these sources, do not answer from general knowledge. For each unread email in [inbox or label]: 1. Label it with one type: [billing, bug, how-to, account access, feature request, cancellation, other]. Set priority: urgent if [the customer cannot use the product, a payment failed, a security or data issue, legal threat], otherwise normal. 2. Escalate, and write no draft, if the email asks for a refund, discount, contract change, legal or security answer, or if the customer is very upset. Write one line on why. 3. Otherwise, write a draft reply of under 150 words in [tone]. Answer only from the sources above. Under the draft, list the source URL or section you used. 4. If you are not sure, write a draft that asks the customer one clear question, and mark it "low confidence". 5. Never promise a refund, credit, deadline, new feature or policy that is not written in the sources. Do not say a human wrote the reply. Output a table: sender, subject, type, priority, escalate (yes or no), reason, draft reply, source, confidence. Save drafts as replies in the original thread only. Do not send. STOP after 10 emails and show me the table.
Human checkpoints: the agent stops and asks you
- Approve the label set, the escalation rules and the help docs the agent may use.
- Read and send every reply yourself. The agent only writes drafts.
- Handle refunds, discounts, legal, security and angry customers yourself.
A reply that promises a refund, a policy or a feature binds your company in the customer's eyes and possibly in law. In Moffatt v. Air Canada, the tribunal held the airline liable for wrong refund advice from its website chatbot. Keep sending behind a human, and route money and policy questions to a person.
What you need
- Your help docs and written policiesdata
Refund, cancellation and SLA rules in writing. The agent may only quote these.
- Inbox access that can create draftspermission
Give draft access only, if your tool allows it. Do not give send rights to the agent.
- An escalation ownerdata
A named person for refunds, legal, security and angry customers.
Steps
Write the rules and gather the docs
List your labels, what counts as urgent and what must go to a person. Collect help docs and policies into one place. Remove old policy pages, because the agent will quote them.
Check: Every label and escalation rule fits on one page, and old policies are gone.
Run on past tickets first
Run the prompt on [20] resolved tickets and compare the drafts with what your team actually sent. Look for any promise not in the docs.
Check: No draft promises anything your docs do not say.
Connect the inbox in draft-only mode
Connect the agent so it can read the inbox and create drafts but not send. Several builders who sell draft tools describe this split as the design: the AI drafts, a person submits.
Check: A test email gets a draft in its thread and nothing is sent.
Review and send each draft
Read each draft and its source. Edit, then send. Track how many drafts you send unchanged, changed or discarded.
Check: You sent every reply yourself and logged the outcome.
Handle escalations fast
Answer escalated emails yourself the same day. In the Cursor case, the correction came hours after the wrong answer had spread online.
Check: Every escalated email has a human reply.
Update the docs from what you fix
When you correct a draft because the docs were missing or wrong, fix the doc. The agent can only be as right as its sources.
Check: Each repeated correction becomes a doc change.
What goes wrong
The bot invents a policy, such as a login rule or a refund window, and customers act on it.
Fix: Allow answers only from written sources, block promises not in them, and keep a person on every send. A Cursor co-founder said AI replies were their first filter for email support when its bot invented a device policy.
Read the lesson →The company argues the bot's answer is not its responsibility.
Fix: Assume it is. The BC tribunal in Moffatt v. Air Canada held the airline responsible for all information on its website, chatbot included.
Read the lesson →Customers do not know a bot answered, and feel deceived when it goes wrong.
Fix: Do not sign AI drafts with a made-up human name. Critics of the Cursor case faulted the lack of disclosure.
The bot gives different answers to the same question.
Fix: Ground every draft in a source and show the source to the reviewer. The Markup found New York City's business chatbot gave conflicting answers to the same housing question.
Drafts quote outdated help docs.
Fix: Keep one current source of truth and remove old pages before you connect the agent.
Evidence
Two or more public accounts from different people or sites describe the same job. We summarise each source in our own words. Open the original before you rely on it.
- Hacker News2025-04-14Cursor IDE support hallucinates lockout policy, causes user cancellations
Users report that Cursor's support bot told them a logout bug was expected behaviour under a policy that did not exist. A Cursor co-founder apologised in the thread and said AI-assisted replies were the first filter for email support. Other commenters report similar invented answers from other companies' support bots.
scaredpelican · news.ycombinator.com - Hacker News2025-04-16Cursor support bot reply by a customer who builds on it
A customer says the bot pushed back on him and pointed him back to his ticket when he asked for a human, who replied hours later. He argues AI may draft and triage but must not contact customers without review.
DonHopkins · news.ycombinator.com - News2025-04-19A customer support AI went rogue
Fortune reports that Cursor's AI support agent invented a login policy, users spread the story and some cancelled subscriptions. Critics faulted the lack of disclosure that the agent was a bot.
Sharon Goldman · fortune.com - News2024-02-19Air Canada chatbot gave a customer wrong info and the airline had to pay
Fox Business reports that Air Canada's chatbot told a customer he could claim a bereavement fare after travel, against the airline's policy, and a tribunal ordered the airline to pay about 812 Canadian dollars.
Aislinn Murphy · foxbusiness.com - News2024-03-29NYC's AI chatbot tells businesses to break the law
The Markup found the city's business chatbot gave wrong legal answers on housing vouchers, tips and cash. It gave one reporter the opposite answer it gave ten other staff.
Colin Lecher · themarkup.org - Hacker News2024-11-26· Custom agentShow HN, human-in-the-loop AI automation for ops teams' email
A founder describes a browser extension that pre-fills reply drafts and refund forms but never submits them, on the view that AI should draft and a person should submit. This is a builder's description with no failure data.
jpochtar · news.ycombinator.com - Web· n8nAI email processing autoresponder with response approval
A public n8n template drafts a reply with an LLM, emails it to a person for a yes or no approval, and only sends on yes. The page reports no outcomes.
Davide Boizza · n8n.io
- Moffatt v. Air Canada, 2024 BCCRT 149checked 2026-10-13
- Missive, answering common customer inquiries with Claude (draft-only setup)checked 2026-10-13
Report how it went
Every report makes this page better for the next person and the next agent. Reports are reviewed before they count. Agents can report with the report_run tool.
Loading reports…