Home  /  Insights  /  Acceptable use policy

An artificial intelligence acceptable use policy for a $50M to $500M company: the eleven clauses

Boards approve AI budgets faster than they set governance expectations. The policy below is the one-page document that closes that gap, drawn from the governance layer of SEAS and written so counsel can mark it up in an afternoon.

By Lalit Kumar · Published 20 September 2026 · 8 minute read

Why the policy comes before the pilot

An artificial intelligence (AI) acceptable use policy is not a compliance ornament. It is the document that lets a chief financial officer approve a pilot, because it states in advance what an agent may touch, who can stop it, and what happens when it is wrong. Without it, every pilot decision is a bespoke risk conversation. In the SEAS implementation system the policy is drafted in Phase 0, before vendor selection, and the readiness checklist for Phase 1 requires it to exist. Since 2 August 2026 the European Union AI Act's transparency obligations have applied to companies operating in Europe or serving European customers, which turns "we should write one" into "we should already have one."

The seven core requirements

  1. Sanctioned tools only. AI tools are deployed through the organization's AI gateway. No shadow AI: no personal accounts, no unapproved browser extensions, no pasting company data into consumer chatbots.
  2. No personal data through unsanctioned models. Personally identifiable information is processed only in approved systems under the data protection controls the company already uses for that data.
  3. A five-minute kill switch. Any agent can be shut down within five minutes by designated personnel, with a minimum of three authorized operators per agent so the switch never depends on one person being awake.
  4. Human-in-the-loop thresholds. A human decision is required for financial transactions above $50,000, all customer-facing communications, legal and compliance decisions, and any human resources action affecting employment status.
  5. Auditable outputs. Every agent output carries a full decision trail: inputs, rule or model applied, result, and who approved it.
  6. Quarterly bias and fairness testing for every customer-facing and human-resources-related agent, with results recorded.
  7. Roles and responsibilities. A named executive sponsor, a named agent owner per deployment, and a steering committee charter with an escalation protocol.

The four cybersecurity guardrails

GuardrailRule
Price lockAgents cannot quote discounts, modify pricing or approve financial commitments beyond predefined thresholds. Any pricing decision above $500 or a 5 percent discount, whichever is lower, requires human approval.
Anti-jailbreak filterAn input validation layer screens every customer-facing agent for prompt injection. Suspicious inputs are logged, blocked and escalated to security.
Personal data egress controlA data loss prevention layer monitors agent outputs; detected personal data is redacted automatically and flagged. Monthly audits confirm zero leakage.
Risk scorecardingEvery deployment receives a composite score from 1 to 5 across data sensitivity, autonomy level, external exposure and business criticality. A minimum of 4 out of 5 is required before scaling from pilot to production.

Mapping the policy to the NIST AI Risk Management Framework

Auditors and acquirers increasingly ask whether an AI programme maps to the National Institute of Standards and Technology (NIST) AI Risk Management Framework. The eleven clauses above map to its four functions without extra paperwork:

  • Govern: the acceptable use policy itself, the roles matrix, the steering committee charter, the risk appetite statement and the escalation protocols.
  • Map: an agent inventory with scope definitions, data lineage maps, dependency mapping, threat modelling and stakeholder impact assessment.
  • Measure: the accuracy, error and exception dashboards, daily accuracy monitoring, quarterly bias testing and model drift detection.
  • Manage: the kill switch, the incident response playbook, remediation workflows and post-incident reviews.

The same dossier serves the General Data Protection Regulation and California Consumer Privacy Act obligations most mid-market companies already carry, and becomes a transferable asset at exit: a buyer's diligence team can read it instead of asking.

Non-human identities: the clause most policies miss

Agents authenticate to systems with credentials of their own. Those credentials are non-human identities, and they need the same least-privilege matrix, ownership and quarterly review that employee accounts get. The SEAS governance layer includes a separate non-human identity access policy with a registry; at minimum, the acceptable use policy should require that every agent credential has a named owner, the narrowest scope that lets it work, and a quarterly review date.

How to adopt it in a week

Day one, the sponsor circulates the eleven clauses to counsel and the chief information officer. Day three, thresholds are set for your company (the $50,000 transaction limit and the $500 or 5 percent pricing lock are the SEAS defaults, not laws). Day five, the steering committee adopts it and the kill-switch operators are named. The policy is then attached to every vendor contract, and the pilot cannot launch without it. Nothing in this page is legal advice; the clauses are an operating baseline for counsel to mark up, not a substitute for counsel.

Frequently asked questions

What should an AI acceptable use policy include?
Sanctioned tools only, no shadow AI, no personal data through unsanctioned models, a five-minute kill switch with at least three authorized operators per agent, human-in-the-loop thresholds, auditable decision trails, quarterly bias testing, named roles, and the four cybersecurity guardrails: price lock, anti-jailbreak filter, personal data egress control and risk scorecarding.
What is a kill switch for AI agents?
A documented procedure by which designated people can shut any agent down within five minutes. SEAS requires a minimum of three authorized operators per agent so the switch never depends on one person.
When must a human approve an agent's action?
Under the SEAS defaults: financial transactions above $50,000, all customer-facing communications, legal and compliance decisions, and any human resources action that affects employment status. Companies set their own thresholds.
Does the policy make us compliant with the EU AI Act?
No single policy does. It gives you the governance record the Act and its transparency obligations expect, mapped to the NIST AI Risk Management Framework, but compliance depends on your systems, your use cases and your counsel.
What is an AI risk scorecard?
A composite score from 1 to 5 across data sensitivity, autonomy level, external exposure and business criticality, assigned to every agent deployment. SEAS requires at least 4 of 5 before a pilot scales to production.

The full policy, the non-human identity access policy with its registry, the NIST AI RMF mapping and the kill-switch protocol ship in the SEAS Templates Library.

See what ships in SEAS   Get the free FLOAT diagnostic book