# AI Reply Policy: [Your Company]

Last updated: [date]
Owner: [your name]

This is the rule we follow for what our AI may say to customers on its own,
what it drafts for a person to review, and what it never touches. It exists
so the first time the AI gets something wrong, we already know what happens
next, instead of deciding under pressure.

## 1. The AI may answer without review when:

- The question is covered word-for-word by a published help article.
- The answer is true regardless of which customer is asking (a how-to step,
  a feature explanation, a published limit).
- Nothing about the answer depends on looking at this specific customer's
  account, order, or history.

Examples for us: [list 3–5 real examples from your own product]

## 2. The AI drafts a reply, a person sends it, when:

- The question touches billing, refunds, cancellations, or account access.
- The reply states a policy the company will be held to later.
- The customer sounds frustrated, confused, or is asking for an exception.
- We haven't seen enough of this exact question answered correctly yet to
  trust it unreviewed.

Examples for us: [list 3–5 real examples from your own product]

## 3. The AI never touches:

- Legal requests and formal complaints.
- Security reports and vulnerability disclosures.
- Data deletion or privacy requests.
- Anything where a wrong answer costs more than a review cycle can catch
  in time.

Route these straight to: [name / address]

## 4. Disclosure

Every AI-written message that reaches a customer is labeled as AI-written,
visibly, at the point they read it, not tucked into a footer or a help page.

## 5. The handoff promise

Customers can always reach a person. Our promise: [e.g. "Reply 'human' and
we'll answer personally within one working day."]
We keep this promise by: [how: an alert, a queue, a person checking daily]

## 6. The review period

For the first [30] days after turning on any new category of unreviewed
AI reply, a human reads every one of them, watching for mistakes and
checking that categories 1 and 2 above are drawn in the right place.

## 7. When we move a question from category 2 to category 1

Only after watching enough drafts about it go out unchanged that we'd be
comfortable explaining the decision to a customer who asked why we trusted
a machine with it. There's no universal number, but there's a test: if a
wrong answer here would mean personally apologizing to the customer, it
stays reviewed.

---

This template is adapted from Foliosio's guide, "An AI reply policy for a
one-person company": https://www.foliosio.com/guides/ai-reply-policy
It is offered as-is, for any company to copy and edit. It is not legal advice.