Written by Julien Ricciarelli-Bonnal
11 November 2026
The Essentials
An internal artificial intelligence policy should do more than warn employees not to share confidential information. A useful document identifies approved tools, permitted uses by function, data that can be entered, review requirements and decisions that must remain under human control. It should also set permissions for agents capable of acting inside company systems, establish accountability and explain how incidents, staff training and policy updates will be managed. A written policy is not a substitute for data protection law or sector-specific obligations. Its purpose is to turn applicable requirements and management choices into instructions that people can follow while doing their actual jobs.
A company may have introduced an AI assistant into its office software, allowed sales staff to experiment with transcription and discovered that employees also use public tools through personal accounts. Nobody necessarily believes they are breaking a rule. There may simply be no common rule to break. Management then searches for an AI policy template, hoping a document will bring order to practices that have already taken hold.
A workable policy is therefore neither an ethical manifesto nor a catalogue of warnings. It is an operating agreement: who may use which system, for which task, with what information, subject to whose approval, and with what response if something goes wrong? Its quality depends on whether those questions have answers outside the boardroom.
Start with what people actually do, not a list of abstract bans
The document needs to define its scope: employees, contractors, approved platforms, covered activities and exceptions. It should distinguish a standalone conversational assistant from AI embedded in existing business software and from systems deliberately deployed by the organisation.
Consider a rule employees can act upon: “The approved assistant may be used to draft a presentation outline from public information. Customer records may only be analysed in the authorised environment. Individual hiring decisions must not be delegated to an unassessed system.” This is more useful than several pages of commitments to fairness and innovation.
There must also be a process for proposing new tools. Without a clear route to approval, blanket prohibition can encourage unofficial accounts and unreported experiments. Management needs to understand which needs its current software fails to meet. Silence about AI use is not convincing evidence that employees have stopped using it.
Data rules must tell employees exactly what can cross the boundary
Confidentiality appears in almost every AI policy template, yet employees still need to recognise the categories involved. Public information, routine internal documents, commercial secrets, client records, personal data, health details, HR files, access credentials and material received under contractual confidentiality obligations do not belong in one undifferentiated category. The policy should supply examples taken from the company’s actual work.
Nor are all technical environments equivalent. A consumer service, a managed business subscription and a system hosted under the organisation’s control may have different arrangements for processing information. The company should examine retention, possible reuse of prompts and outputs, administrator access, subcontractors, security measures and international transfers.
France’s data protection authority, the CNIL, recommends defining authorised and prohibited uses and assessing the risks of generative AI. Where personal data are involved, the policy must operate alongside the organisation’s data protection procedures. Depending on the processing, questions about lawful basis, responsibilities between customer and provider, retention and security will require proper examination. A statement that employees must obey the GDPR does not perform that work.
Separate drafting assistance from decisions that require approval
Review should reflect consequences. Brainstorming possible headings for an internal presentation should not require the same process as publishing a financial claim, issuing a customer quote or evaluating a candidate. A practical distinction is between low-impact assistance, material that requires a named reviewer and sensitive decisions that remain with an accountable person. The distinction should be reflected in workflows, not left as a vague expectation of vigilance.
We previously examined why some businesses restrict ChatGPT use among their employees. The underlying problem is not simply the choice of software. It is the gap between how convincing a generated response appears and the checks actually performed before someone relies upon it.
“Users remain responsible” is not a sufficient operational instruction. Who checks the figures? Who signs off a public statement? Does a contract draft need legal review? Does a customer-facing recommendation need a second approval? The policy has to name the checks and the roles involved.
Autonomous agents need permissions, not just editorial guidelines
A new boundary appears when AI moves beyond producing material and starts taking action. An assistant that drafts a reply is not equivalent to an agent that sends it to a hundred customers. A policy should identify what systems may do without further approval: retrieve records, amend CRM entries, create orders, publish content, contact third parties or initiate financial operations.
It should distinguish reading from writing and execution, then put explicit limits around each. An agent may prepare an email without permission to send it. It may analyse an invoice without the authority to pay it. Technical access rights, financial thresholds, human checkpoints and activity logs must reinforce the policy. An instruction written into a prompt is not necessarily a reliable security boundary.
The organisation also needs a stopping procedure. Who can revoke access or suspend an agent when a sequence becomes abnormal? Can actions be reconstructed, corrected or reversed? Who monitors activity, and what triggers escalation?
An internal policy does not automatically establish AI Act compliance
It is tempting to present an AI policy as the document that makes a company compliant with Europe’s AI Act. That is misleading. The Regulation does not impose one universal charter on every company. Its obligations depend on the organisation’s role, the type of system, the way it is deployed and any relevant risk classification. Some businesses face requirements that cannot be satisfied through staff guidance alone.
Certain rules already apply. The AI literacy provisions began applying on 2 February 2025 and require providers and deployers to take appropriate measures to develop relevant understanding among people working with AI systems on their behalf. Further provisions, including important transparency rules, became applicable in August 2026. Different dates apply to various high-risk categories. A company must assess its specific position rather than assume a standard paragraph covers its duties.
The essential distinction is between legal obligations and internal management choices. Data protection, cybersecurity, employment law, professional confidentiality and applicable sectoral rules set constraints that the organisation must respect. The policy translates them into instructions people can use: when additional legal review is needed, when AI involvement must be disclosed and when a deployment must be stopped pending assessment.
Give the document an owner and a way to change
A policy with no operating owner soon becomes a forgotten file. Someone must manage revisions, handle requests for exceptions, receive incident reports and ensure that decisions reach the relevant teams. Senior management should retain ownership of major choices, supported by data protection, legal, security and operational expertise.
Training is part of the operating model, not a ceremonial presentation of the rules. Staff need practice recognising sensitive information, testing factual claims and knowing when to escalate. That is also the purpose of AI Business Governance built around practical uses, responsibilities and adoption: connecting executive decisions with the routine work employees actually perform, rather than accumulating policies detached from practice.
The document should be revisited when suppliers change their terms, new data are introduced, an agent gains broader access or an incident exposes a weakness. Useful indicators might include exceptions requested, review failures, incidents reported and workflows improved.
A good AI use policy does not attempt to eliminate every risk through increasingly elaborate bans. It makes clear which freedom employees have and where supervision is justified. Its immediate test is ordinary: someone hesitates before submitting a document, using a tool or approving a result, and can find an answer that works in the circumstances.
The deeper test is whether the organisation can explain its important AI uses: what it authorises, why, who is accountable and how a decision can be changed. At that point, the policy stops being a defensive statement of intentions and becomes part of the way the company governs AI-assisted work.
Written by Julien Ricciarelli-Bonnal
11 November 2026
