Key findings

• Make the policy apply to embedded AI features, not only standalone chatbots.

• Tie approval depth to the effect on people, money, rights, safety, and confidential data.

• Keep authorization and high-impact decisions under deterministic and human control.

• Require an inventory, evidence record, incident route, and periodic review.

• Treat the template as operational governance, not legal advice or a compliance certificate.

Policy purpose, scope, and principles

Template text: This policy governs the selection, development, configuration, deployment, and use of artificial intelligence systems by the company, including generative AI, machine learning, AI features embedded in software, and agents that call tools or change records. It applies to employees, contractors, systems, and suppliers using company data or acting on the company's behalf.

The company will use AI for defined business purposes in a lawful, secure, privacy-conscious, fair, transparent, and accountable manner. AI use must be proportionate to the benefit and risk. The company will prefer simpler deterministic methods when they provide an adequate result. People remain accountable for work performed with AI, and the company will provide a practical way to question or correct consequential outcomes where required.

Roles, inventory, and approval authority

Template text: The executive AI owner approves policy, risk tolerance, and high-risk uses. Each AI system has a business owner responsible for purpose, users, performance, and incidents, plus a technical owner responsible for configuration, security, monitoring, and change. Privacy, legal, security, HR, or other specialists review uses within their remit. No system may operate without a named owner.

The company maintains an AI inventory recording purpose, owner, vendor and model, users, data classes, connected tools, effective permissions, hosting and transfer locations, risk tier, approval date, evaluation evidence, current version, incidents, review date, and retirement status. Staff must register new AI uses and material changes before production use. Unapproved shadow systems may be blocked or removed.

Risk tiers and required controls

Template text: Low-risk uses assist with reversible internal work using non-sensitive data and require owner approval, approved tools, and basic checking. Moderate-risk uses affect external content, customer operations, confidential data, or business records and require documented assessment, testing, permissions, monitoring, and human review. High-risk uses materially affect employment, credit, access to essential services, legal rights, health, safety, or vulnerable people and require executive, legal, privacy, and security approval before any pilot.

Prohibited uses include unlawful discrimination, deceptive impersonation, covert monitoring outside approved lawful purposes, bypassing access controls, placing credentials or restricted data into unapproved tools, autonomous high-impact decisions without required safeguards, and any practice prohibited by applicable law. Risk classification must consider intended use and reasonably foreseeable misuse, not only the vendor's product category.

Acceptable use and employee responsibilities

Template text: Staff may use only approved accounts, models, integrations, and data. They must verify material facts, calculations, citations, customer commitments, code, and other outputs before use. AI-generated content must not be represented as independently verified. Staff must follow confidentiality, intellectual property, records, security, equality, and professional obligations.

Users must not enter personal, confidential, customer, employee, financial, health, credential, or legally privileged information unless the use is explicitly approved for that data. They must not instruct a system to exceed their own authority or use output to bypass a control. Suspected leakage, harmful output, unauthorized action, unexpected access, or material error must be reported through the defined incident channel immediately.

Data, privacy, security, and permissions

Template text: Every AI use will have a documented purpose, data source, lawful basis where personal data is processed, retention rule, and approved recipients. The company will minimize input, retrieval, memory, output, and telemetry data. Sensitive data will be redacted, tokenized, or excluded where feasible. Privacy notices, rights handling, impact assessments, and transfer safeguards will be updated when required.

AI systems must use unique managed identities, least-privilege access, approved secrets management, encryption, secure configuration, and auditable actions. Read and write capabilities should be separated. High-impact or irreversible tool calls require explicit policy checks and, where appropriate, human approval. Model output is untrusted input to downstream systems and must be validated. Credentials may not be embedded in prompts, source code, or logs.

Human oversight, testing, and change management

Template text: Human review must be meaningful. The reviewer must have competence, relevant information, time, authority to reject or change the output, and a route to escalate. Review intensity will reflect consequence and uncertainty. The company will monitor for automation bias and will not treat a nominal approval click as adequate oversight.

Before release, the owner will define acceptance criteria and test representative, edge, and adversarial cases for quality, policy compliance, security, privacy, reliability, cost, and fallback. Versions of models, prompts, retrieval sources, tools, and evaluations will be recorded. Material changes require regression testing and, for moderate- or high-risk systems, renewed approval. Production monitoring and periodic human sampling continue after launch.

Vendor, contract, and lifecycle requirements

Template text: AI suppliers must pass risk-proportionate due diligence covering architecture, data use, subprocessors, security, permissions, evaluation, service operation, incidents, intellectual property, regulatory support, and financial and operational resilience. Contracts will address confidentiality, training use, retention, transfers, breach notification, change notice, audit evidence, service levels, support, data return, deletion, and exit.

The business owner will review performance, risk, access, and supplier changes at least at the cadence assigned to the risk tier. Systems that no longer have a valid purpose, owner, support path, or acceptable risk will be suspended or retired. Retirement includes revoking identities, removing connectors, exporting required records, deleting data and memory, confirming supplier deletion where applicable, and updating the inventory.

Incidents, exceptions, training, and policy review

Template text: AI incidents include unauthorized access or action, data leakage, harmful or discriminatory output, material misinformation, uncontrolled cost, unavailable fallback, policy bypass, and repeated quality failure. Staff will stop or contain the workflow when safe, preserve relevant evidence without spreading sensitive content, notify the owner, and follow security, privacy, legal, customer, and regulatory response procedures.

Exceptions must be documented with scope, reason, owner, compensating controls, approval, and expiry. Relevant staff will receive role-based AI literacy and security training. The policy owner will review this policy at least annually and after significant incidents, legal changes, new risk classes, or major technology changes. The current version, approvals, and review history will be retained.

Sources and methodology

This article synthesizes the primary sources below as of the publication date. Forecasts and recommendations are directional scenarios, not guarantees; they should be tested against your workflow, data, risk tolerance, and current vendor documentation.

National Institute of Standards and Technology: AI Risk Management Framework 1.0 (accessed 2026-07-30)

National Institute of Standards and Technology: Generative Artificial Intelligence Profile (accessed 2026-07-30)

National Institute of Standards and Technology: NIST Cybersecurity Framework 2.0 for Small Business (accessed 2026-07-30)

UK Information Commissioner's Office: AI and data protection risk toolkit (accessed 2026-07-30)

European Union: EU Artificial Intelligence Act (accessed 2026-07-30)

OWASP Foundation: OWASP AI Agent Security Cheat Sheet (accessed 2026-07-30)