An AI email policy template for small teams

AI AND EMAIL POLICY — [Company name]
Owner: [name]          Last reviewed: [date]          Next review: [date]

1. PURPOSE
We use AI to save time on email: sorting, summarizing, finding threads and
drafting replies. This policy keeps that safe for our customers and for us.

2. APPROVED TOOLS
You may use only these AI tools with company email:
  - [Tool 1] — approved for: [sorting, summaries, drafts]
  - [Tool 2] — approved for: [drafting in our shared mailboxes]
  - [Assistant + connector] — approved for: [read, organize, draft]
Anything not on this list needs approval from [owner] before use.
Do not connect personal AI accounts to company mailboxes.

3. WHAT DATA MAY BE PROCESSED
Allowed: routine customer and partner email in approved mailboxes.
Not allowed in any AI tool:
  - Passwords, API keys, recovery codes or other credentials
  - Payment card numbers or bank details
  - [HR, legal, health or other sensitive categories]
  - Anything a customer asked us to keep confidential
Do not paste company email into AI tools that are not on the approved list.

4. HUMAN REVIEW BEFORE SENDING
Every email that leaves our company is reviewed and sent by a person.
  - AI may draft. A human reads the full draft and clicks send.
  - The person who sends owns the content, regardless of who or what drafted it.
  - Check facts, numbers, links, promises and recipients before sending.
  - Never let a tool send, forward or delete mail automatically
    [exception: templated acknowledgments listed in section 8].

5. SENSITIVE MAILBOXES
These mailboxes are excluded from AI tools and connected assistants:
  - [billing@ / finance mailbox]
  - [legal@, security@, or similar]
  - [HR or people mailbox]
Access to these mailboxes is limited to [roles].

6. CONNECTED ASSISTANTS
If you connect an AI assistant (for example via an MCP connector):
  - Use your own account, never a shared login.
  - Grant only the access you need (read and draft, not send).
  - Review your connected apps every [quarter] and revoke what you don't use.
  - Revoke all connections on your last day.

7. DISCLOSURE
[Choose one]
  (a) We do not label individual emails as AI-assisted. Every email is
      reviewed and sent by a person, who is responsible for it.
  (b) We tell customers we use AI to help draft replies, in our
      [support page / privacy notice].
If a customer asks whether AI was involved, answer honestly.

8. AUTOMATED MESSAGES
These may be sent without human review because they are fixed templates:
  - [Auto-acknowledgment for support@]
  - [Receipts and system notifications]
AI-generated content is never sent automatically.

9. INCIDENTS
Report to [owner/channel] within [one business day] if:
  - An email was sent with incorrect AI-generated content of consequence
  - Sensitive data was put into an AI tool against this policy
  - An AI tool did something unexpected (acted on instructions in an email,
    touched the wrong mailbox, sent or deleted without approval)
We fix the issue, tell affected people where appropriate, and update this policy.
No blame for reporting honestly.

10. REVIEW
This policy is reviewed every [six months] or when we adopt a new AI tool.

Notes on each section#

Approved tools#

Keep the list short and specific about what each tool is approved for. "Approved for drafting in support@" is clearer than "approved." The requirement to ask before adding a tool matters more than the list itself; it's how you avoid finding out six months later that customer mail has been flowing through a browser extension nobody vetted.

Data that may be processed#

Most teams need to adjust the "not allowed" list. A design agency's sensitive data looks different from a fintech's. Think about what would be most damaging if it ended up somewhere unexpected, and list that. Credentials are on every list, because people paste them into email far more often than anyone admits.

Human review before sending#

This is the core of the policy. It's also the section most likely to erode quietly, as people start trusting drafts and skimming. Two lines keep it meaningful: the sender owns the content, and the specific things to check. If you want help making review fast, a short checklist works better than a general "be careful."

Choose tools that make this easy to follow. Some products enforce it by design; in Koltrix, for instance, AI drafts replies but nothing is sent without a person clicking send.

Sensitive mailboxes#

Excluding mailboxes is a stronger control than asking people to avoid certain topics. People can't accidentally process what the tool can't reach. Pair it with access control, so only the right people see those mailboxes at all.

Connected assistants#

General-purpose assistants that connect to email are becoming common. The rules are the same as for any app with access to company data: individual accounts, least access, regular review, and revocation when someone leaves. Add a line to your offboarding checklist.

Disclosure#

There's no universal rule here. Some teams prefer to state AI assistance publicly; others treat AI drafting like spellcheck, because a human reviews and sends every message. Pick one, write it down, and make sure nobody ever denies AI involvement when asked directly.

Incidents#

The incident section exists so problems surface early. The "no blame" line matters: people hide mistakes they expect to be punished for, and a hidden mistake is the one you can't fix.

Adapting it to your team's size#

Two or three people. You can collapse sections. A founder-led team might merge approved tools, connected assistants and data limits into one short list. Keep human review and incident reporting as separate, explicit sections; those are the ones that protect you when you're busy and cutting corners.

Five to fifteen people. Use the full template. This is the size where unwritten norms start to diverge, especially when a support hire, a contractor and a salesperson all handle email differently. Name a single owner for the policy so questions have somewhere to go.

Larger, or regulated. Treat this template as a draft for your compliance or security owner. You'll likely need data-processing agreements with vendors, a record of which tools touch which data, and more formal approval for new tools.

Mistakes to avoid#

• Writing a policy nobody can follow. "Never use AI with customer data" sounds safe, but if your support tool drafts replies with AI, it's already broken. Write rules that match how the team actually works, then tighten where needed.
• Banning without offering an alternative. If people find AI drafting useful and you ban unapproved tools without approving any, they'll keep using the unapproved ones quietly. Approve something.
• Forgetting contractors. Freelancers and agencies often have access to shared mailboxes. The policy applies to them too, and their tools may be the least visible.
• Never revisiting it. Tools change quickly. A policy written before your team connected an assistant to email won't cover what that assistant can do.

Rolling it out#

• Walk through the policy in a short team meeting rather than just sending a link. Questions surface gaps.
• Update tool settings to match the policy on the same day, especially excluded mailboxes and connector permissions.
• Add the policy to onboarding for new hires and contractors.
• Revisit it after the first incident report, whatever it is. That's when you learn what the policy missed.

Key takeaways#

• Your team already has an AI email policy; writing it down makes it consistent.
• The essentials: an approved tool list, data limits, human review before send, excluded sensitive mailboxes, and incident reporting.
• Excluding mailboxes from AI tools is stronger than relying on people to avoid topics.
• The person who clicks send owns the email, regardless of what drafted it.
• Review the policy on a schedule and after every incident.