Guide
An AI Reply Policy for a One-Person Company
A written rule for what your AI may answer alone, what it drafts for a human to send, and what it never touches. Copy it, cut what doesn't apply, and use it in whatever tool you run support from.
Two things that happened without a written policy
In April 2025, a developer noticed that Cursor's support bot had told a customer the editor couldn't be used on more than one device, and invented a licensing policy to justify it. The company had no such policy. The thread on Hacker News reached over 1,500 points and 600 comments, mostly from developers who felt a machine had made a decision on the company's behalf without telling anyone it was a machine. Users cancelled over a policy that didn't exist.
A year earlier, a Canadian tribunal held Air Canada responsible for a refund policy its own chatbot had invented. The airline argued the chatbot was "a separate legal entity." The tribunal disagreed: a company is responsible for every answer its website gives, however that answer was produced.
Neither company intended to mislead anyone. Neither had decided, in writing, what their AI was and wasn't allowed to say before it said something costly. That decision is easiest to make before anything goes wrong.
What your customers are already worried about
A 2024 Gartner survey found that 64% of customers would rather a company not use AI in customer service at all, and 53% said they'd consider switching to a competitor over it. Asked why, the top two reasons were losing access to a human (60%) and getting a wrong answer (42%). Your customers aren't anti-AI in the abstract. They're worried about exactly the two failure modes above: no way out, and no way to trust what they're told.
A written policy answers both, before anyone has to ask.
This is a page you can point to and a rule you follow, so the first time your AI gets something wrong, you know what happens next.
The policy
Fill in the blanks for your product. Keep the structure: it's the structure that does the work.
1. Questions the AI may answer without review
Only questions covered word-for-word by a published help article: how-to steps, feature explanations, things that are true regardless of which customer is asking. If an answer requires looking at a specific account, it doesn't belong here.
2. Questions the AI drafts, a human sends
Everything about billing, refunds, cancellations, account access, anything that states a policy the company will be held to, and anything from a customer who sounds frustrated or angry. The AI can still read the thread, find the account and write a first draft. A person reads that draft before it goes out.
3. Questions the AI never touches
Legal requests and security reports, data deletion requests, and anything where being wrong costs more than a review cycle can catch. Route these straight to a human with no AI involvement at all, not even a draft.
4. Disclosure
Every AI-written reply that reaches a customer says so. Not a footnote: a visible label, at the moment they read it.
5. The handoff promise
Tell customers how to reach a person, and mean it. For example: "Reply 'human' and I'll answer personally within one working day." Then answer within one working day.
6. The first month
Read every AI-sent reply for the first month, however few there are, and check whether you drew categories 1 and 2 above where you meant to.
Copy the policy as a Markdown file and fill in your own product.
Download the templateWhen to loosen it
Move a question from category 2 to category 1 only after you've watched enough drafts about it go out unchanged that you'd be comfortable explaining the decision to a customer who asked why. There is no universal number, but there is a universal test: if an answer being wrong would make you personally apologize to that customer, review it before it sends.
How this maps to Foliosio, if that's what you use
Foliosio has three AI reply settings, and they map onto the policy above directly:
Off
No AI replies at all. The default on every new workspace.
Draft
The AI writes a reply grounded in your help articles and product context. You read it and send it, or edit it first. This is where category 2 lives.
Auto
The AI sends on its own, up to 8 replies per conversation. Turn this on only for the category-1 questions you've already watched work in Draft mode.
Two things are true regardless of which mode you choose. Every AI-generated message in the chat widget carries a visible "AI-generated" label, which is also what EU AI Act Article 50 requires when a customer is talking to a machine. And if you connect an agent (Claude Code or Cursor) to read and tag your inbox, that connection cannot send a reply to a customer under any setting. It can read, tag, change status and save a draft, nothing more, and every change it makes is written to your workspace's audit log.
You don't have to build any of that. It ships as the default. The policy above is still yours to write, because it decides which questions reach Draft and which reach Auto, and no product can make that call for you.
Turn on Draft when you're ready to see what the AI would say before any customer does.