An AI policy is the document that tells your team what they can + can't do with AI tools — and the document that protects the company when something goes wrong. Most companies don't have one in 2026. The ones that do are quieter about it.
TLDR — the structure
A working AI policy has 5 sections:
- Allowed uses — what's encouraged
- Forbidden uses — what's prohibited and why
- Data classification — what data can go into which AI tool
- Customer-facing AI — disclosure + accuracy requirements
- Incident response — what to do when AI does something wrong
Plus a one-page "TL;DR for everyone" at the top. The full policy is 3-5 pages; the TL;DR is what gets read.
Section 1: Allowed uses (be explicit + permissive)
Most teams default-deny in AI policy. This is wrong. Default-deny pushes employees to use AI without telling you.
Default-allow common productivity uses, then specify the exceptions:
Examples of allowed:
- Drafting + editing internal documents (memos, emails, plans)
- Research + summarization of public information
- Code generation for personal productivity (with review)
- Meeting summaries from notes you took
- Brainstorming + ideation
If your policy doesn't allow these, employees will ignore it. Better to allow + guide than forbid + lose visibility.
Section 2: Forbidden uses (specific + justified)
For each prohibition, give the reason — not "because we said so" but "because of risk X."
Common prohibitions:
- Customer PII into public AI tools. Reason: data residency + breach risk. Use the approved tools (see Section 3).
- Source code into public AI tools without IP review. Reason: IP leakage + license compatibility. Use approved tools.
- Decisions about employees (hire/fire/promote) based on AI without human review. Reason: discrimination liability + fairness.
- Generated content presented as human-authored externally without disclosure. Reason: deception + customer trust.
- AI-generated code shipped without human review. Reason: quality + security.
- Using AI to circumvent existing policies (e.g., generating fake invoices). Reason: it's still fraud regardless of who/what created it.
Note the absence of "no ChatGPT, period." That doesn't work.
Section 3: Data classification
This is the section that does the most work. Tie AI-tool usage to data classification levels.
Sample classification:
| Tier | Examples | Allowed AI tools |
|---|---|---|
| Public | Marketing copy, public blog posts | Any AI tool |
| Internal | Internal docs, meeting notes, code | Approved AI tools (list) |
| Confidential | Customer PII, employee data, financial details | Enterprise-tier approved tools only (e.g., Claude Enterprise, ChatGPT Enterprise, your self-hosted setup) |
| Restricted | Trade secrets, security incidents, M&A | NO AI tools — human-only |
The "approved AI tools" list belongs in an appendix that you can update without rewriting the policy. Refresh quarterly.
Section 4: Customer-facing AI
Special rules for AI that interacts with customers.
Required practices:
- Disclosure. If a customer is talking to an AI (chatbot, voice agent, email auto-reply), disclose it. Required by EU AI Act and California AB-1047; recommended everywhere.
- Escalation paths. Customer can always reach a human within the same channel. No infinite-loop AI traps.
- Accuracy review. Customer-facing AI outputs (refund decisions, billing actions, advice) reviewed regularly for accuracy + bias.
- Restricted topics. Don't deploy AI to handle medical/legal/financial advice unless you've engineered for it specifically.
- Audit logging. Every customer-facing AI conversation logged + reviewable.
Section 5: Incident response
When AI does something wrong, what's the process?
The 5-step incident pattern:
- Detect — How was the issue surfaced? (User complaint, automated monitoring, employee report?)
- Contain — Disable the misbehaving agent or scope it to dry-run mode immediately
- Assess — Scope: how many users/transactions/outputs affected?
- Communicate — Internal team + affected users + (if material) external comms
- Fix + post-mortem — Root cause analysis + policy/configuration update + share learnings
Most incidents will be small. The process scales down for small incidents. Having the framework matters when you need it.
The one-page TL;DR
Put this at the top of the policy. It's what employees actually read.
AI POLICY — TL;DR
DO use AI for:
- Drafting + editing internal docs
- Research + summarization
- Code generation (with review)
- Brainstorming
- Meeting notes
DON'T put into AI:
- Customer PII (unless using [Approved Tool X])
- Source code (unless using [Approved Tool Y])
- Trade secrets / M&A info — never
DON'T use AI to:
- Make hire/fire decisions without human review
- Ship code unreviewed
- Generate customer-facing content without disclosure
When in doubt: ask [policy owner]. We'd rather you ask than guess.
Full policy: [link]
Two common policy failure modes
Mode 1: Too restrictive
"No AI tools allowed without approval." Result: employees use AI tools anyway, just without telling you. You lose visibility + control. Worse than having no policy.
Mode 2: Too vague
"Use AI responsibly." Result: every employee makes their own call. You have a policy in name but no enforcement.
The middle path: be specific about the rules that matter, generous about the rules that don't.
How to roll out the policy
- Draft internally (Ops + Legal + Engineering owners)
- Pilot with one team for 2-4 weeks; collect questions; refine
- All-hands announce with the TL;DR + 30-min Q&A
- Train managers to handle the gray-zone cases their team brings them
- Review quarterly — the AI landscape shifts every 3 months; static policies decay
Who owns the policy?
Typically:
- Drafting: Cross-functional (Ops/IT/Legal/Engineering)
- Approval: CTO or CISO + Legal
- Ongoing ownership: A specific person (often in IT or Ops) whose job includes "AI governance"
- Day-to-day enforcement: Managers + a central queue for escalations
The biggest risk is no owner — the policy gets written, posted, and forgotten.
Related governance documents
- AI governance glossary entry
- AI safety glossary entry
- EU AI Act compliance considerations
- AI agent compliance 2026
Bottom line
A working AI policy is short, specific, and revisited quarterly. It's permissive by default + restrictive where stakes are high. It has an owner. It survives contact with employees who actually use AI tools daily. If your policy is none of these, rewrite it.