An AI Acceptable Use Policy (AUP) is the first document most organizations reach for when AI comes up, and the one most often written in an afternoon. A two-page memo that says “use approved tools only” is a reasonable start. It is not yet a policy that a regulator, an auditor or an underwriter can read against what your people actually do.
This week looks inside the AUP that SanctumShield generates: what it has to answer, who reads each part, and what it is anchored to.
Start from real use, not from a template
A useful AUP begins with an inventory. Which AI tools are in use today, which business software has AI switched on inside it, which people use AI on personal or unmanaged devices, and which agents act on company systems. Those are the four layers of Shadow AI we covered in Week 4.
A policy written before that inventory tends to approve tools nobody uses and stay silent on the ones everybody does. A policy written after it can name the tools, the data each may touch, and the owner who approved each one. That is why the SanctumShield AUP is generated from the same guided assessment that produces the Executive Risk Report. The policy and the findings describe the same organization on the same date.
What the policy has to answer
The generated policy runs up to sixteen sections plus three appendices, sized to your organization. A small professional-services firm with no agents receives fewer sections than a hospital group running coding agents and clinical AI.
Grouped by the question each part answers, the policy covers:
| The question | What the policy sets down | Who reads it most closely |
|---|---|---|
| What does this policy cover, and from when? | Purpose, scope, effective date, definitions in plain language | Everyone |
| Which AI tools may we use, and for what? | Approved tools, permitted uses, the Approved AI Tools Registry (Appendix A) with a review date per tool | Employees, IT |
| Which data may go into AI, and which may not? | Data classes, prohibited inputs, sector rules such as HIPAA for health data | Employees, privacy, legal |
| Who checks the output, and how? | Human review duties, and a clause on persuasive AI output that can wear down a reviewer’s judgment | Managers, reviewers |
| What may agents do on our systems? | The Deployed Agent Policy: registration, allowlists, a named spend owner, a kill switch, rules for agent-authored code | CISO, engineering |
| Which outside providers see our data? | Vendor and third-party AI assessment | Legal, procurement |
| What happens when something goes wrong? | Reporting AI incidents and near misses without blame | Everyone |
| When does this policy change? | Review cadence, ownership, acknowledgement | Policy owner, board |
A policy with no names in it describes intentions. A policy with names in it describes a decision.
Anchored to clauses, not to adjectives
The EU AI Act asks deployers to take measures on AI literacy (Article 4, since February 2, 2025) and to meet transparency duties (Article 50, since August 2, 2026), with the high-risk regime following on December 2, 2027. Colorado SB 26-189, signed May 14, 2026, with duties from January 1, 2027, asks deployers of automated decision-making technology for notices and three years of records. HIPAA’s administrative safeguards (§164.308) already apply to health data, whatever tool touches it. The NIST AI Risk Management Framework places policy inside its GOVERN function, and ISO/IEC 42001 asks for an AI policy as part of an AI management system.
The anchor matters because it changes the conversation with a reviewer. “We have a policy” is a claim. “Section 5 answers HIPAA §164.308 for these three tools, owned by our privacy officer, reviewed on this date” is something a reviewer can check.
The section most policies leave out
Most AUPs tell people which tools to use. Fewer say anything about how a person should review what the tool produces. Research summarized in Week 20 found that human reviewers caught the problem roughly one time in four. When every regulation that asks for human oversight assumes the reviewer will catch the error, the policy should say how the review is done: independent first views, a written reason when the reviewer accepts a recommendation, and permission to say “I don’t know.”
That clause costs one paragraph. It is often the one an auditor asks about first.
Agents get their own section
When an agent acts on company systems, the organization has delegated authority, as our “Delegation Is Not Automation” essay sets out. The Deployed Agent Policy section records what each agent may do, who approved it, who pays for it, and who can stop it.
A policy that can be checked
The finished AUP carries a Verification URL. Anyone you share it with (a board member, an auditor, an underwriter) can confirm the document’s date and identity without seeing its contents. SanctumShield keeps that confirmation available for six years. The policy text itself stays with you.
Your next step
Before you draft or revise a policy, list the AI your people actually use this month. The free Shadow AI Check asks twelve plain questions and gives you a first view of that list. The policy you write after that list will be shorter where it can be and specific where it has to be.
Sources
- Regulation (EU) 2024/1689 (EU AI Act), Articles 4, 50 and the high-risk timeline as amended by the 2026 Digital Omnibus
- Colorado SB 26-189, signed May 14, 2026, with duties from January 1, 2027
- 45 CFR §164.308, HIPAA administrative safeguards
- NIST AI Risk Management Framework 1.0, GOVERN function
- ISO/IEC 42001:2023, AI management systems
- Week 20, Human Oversight Caught It Roughly One Time in Four
- Week 4, The Four Layers of Shadow AI
