Acceptable Use Policy¶
The AUP is the single most referenced document in the program. Every engineer is expected to have read it; every auditor will ask for it; every incident begins with "what does the AUP say about this?"
This page describes what a good AUP contains, independent of any specific company's wording. For the drafting meeting that produces version 0.1, see the AUP drafting worksheet.
Structure¶
A permanent AUP (Phase 2 onward) typically has these sections. An interim AUP (Phase 0) collapses most of these into one page.
| § | Section | Length guide |
|---|---|---|
| 1 | Purpose and scope | 1 page |
| 2 | Approved tools | 1 page (updated as tools change) |
| 3 | Data handling | 2 pages |
| 4 | Developer obligations | 2 pages |
| 5 | Agent-specific rules | 1 page |
| 6 | IP and licensing | 1 page |
| 7 | Incident reporting | 1 page |
| 8 | Exceptions | 1 page |
| 9 | Review cadence & change log | 0.5 page |
Total: around 10 pages. Longer AUPs go unread.
Non-negotiables¶
Regardless of company size or sector, these clauses belong in every AUP:
- A single named incident channel with an SLA.
- An explicit "prohibited data" list: data classes that must never be sent to any AI tool.
- An approved-tools list: not "anything with SSO", but specific tools with specific conditions.
- A disallowance of consumer-tier / personal accounts for company work.
- A verification obligation. Engineers must verify output, and AI-assisted merges carry the same review obligation as any other.
- An exceptions process: named owner, timeframe, evidence required.
- A review cadence: ≤ 3 months for interim AUP, ≤ 12 months for permanent AUP.
- An acknowledgement requirement. Every in-scope staff member acknowledges each version, tracked.
- A per-tool monthly spend cap, defined and enforced (technically where possible) before broad rollout and reviewed at each contract renewal. Caps are tracked per tool independently; a combined budget pool allows a single high-consumption tool to exhaust the annual allocation before the overrun is visible.
Drafting principles¶
- Write for the engineer who will act on it. A clause that only a lawyer can parse will be ignored.
- Reference, don't duplicate. If the retention policy lives in another document, link rather than copy.
- Use roles, not names. "The CoE Lead" not "Jane Smith." Names change.
- Cite the external standard when possible. "Per NIST AI RMF Govern function" gives auditors the crosswalk for free.
- Include negative space. What is out of scope deserves a sentence as much as what is in.
Enforcement¶
- HR alignment. Acknowledgement is recorded through the HR system (or equivalent); non-acknowledgement by the deadline is a compliance issue.
- Tool-level enforcement. SSO, SCIM, and gateway rules enforce the approved-tools clause technically. Policy alone is insufficient.
- Enforcement placement. Each security control has a named enforcement layer (endpoint, network/platform, or vendor) and a named owner for each layer it spans. The S.H.I.E.L.D. enforcement placement model governs which layer is authoritative for each control type. When the approved stack includes multiple tools, a per-tool enforcement matrix documents where each control is enforced and compensates for vendor gaps at the network layer.
- Review-level enforcement. AI-adjacent PRs are tagged; tech leads are trained to check tag presence during review.
- Blame-free reporting. Good-faith incident reports are not a disciplinary matter; not reporting is.
Version and change control¶
- AUP lives in version control (git, internal wiki with history, etc.) alongside the Decision Records that drove each change.
- Every material change produces a DR with status
accepted. - AUP version bumps are recorded internally (the company's AUP is internal; this framework is only the public scaffolding).
Sizing guidance¶
Sizing guide
- Small (< 50): One-page AUP covers it all. Keep the approved-tools list to 1–3 tools. Review quarterly.
- Mid (50–500): 5–7 page AUP with per-section owners. Review every 6 months.
- Enterprise (500+): 8–10 page AUP, with BU-specific addenda for distinct sectors. Review annually, with interim patches.
- Regulated: add sector-specific addenda (HIPAA, PCI, DORA). Map every clause to a control in the regulatory crosswalk. Expect audit-level scrutiny.
Common failure modes¶
- "The AUP is six months old and nobody remembers what changed." Add a changelog section; refresh on review cadence.
- "Engineers use unapproved tools anyway." Policy without technical gates fails. Pair AUP with gateway or tenant-level controls.
- "Legal rewrote it without Engineering; nobody follows it." AUP requires a working group; see AUP drafting worksheet.
- "We have five AUPs (one per BU) that conflict." One company-wide AUP, with addenda for BU-specific clauses. Conflicts get a DR.
External references¶
- NIST AI RMF1, Govern function (G2, G3).
- ISO/IEC 42001:20232, clauses 5 and 8 (policy, operational controls).
- ICO (UK Information Commissioner's Office) guidance on AI and data protection3. Useful when the AUP interacts with UK GDPR.
Next¶
- AUP drafting worksheet: how to produce v0.1.
- Data handling: the detailed data rules the AUP references.
- Regulatory mapping: which clauses map to which external standard.
-
NIST, NIST AI 100-1, Jan 2023. ↩
-
ISO, ISO/IEC 42001:2023. ↩
-
UK ICO, Guidance on AI and data protection. ↩