The Board Pack Has a New Reader: Governing AI in the Confidential Paper Cycle

Board packs have always been vulnerable to careless distribution. The difference now is that the careless distribution may be automated. A director pastes a sensitive paper into a drafting assistant to produce a summary. An executive uses a plug-in to compare two versions of a strategy document. A company secretary stores extracts in a shared workspace that happens to be connected to an AI service. None of these actions is necessarily malicious, but each can remove information from the controlled perimeter of the boardroom.

The governance question is not whether boards should use AI. It is whether the board can demonstrate that its use is deliberate, limited and recorded. A confidentiality protocol for AI-assisted board papers should sit alongside the existing information-security and conflicts regimes. It should be owned by the chair, administered by the company secretary and tested before the next pack cycle, not after an incident.

Begin with classification, not with tools

A protocol that starts by listing approved products will age quickly. A protocol that starts with information classification will remain useful as tools change. Boards should adopt a simple tiering system for pack material: public, internal, confidential and board-only or legally privileged. Each tier should state whether AI processing is permitted, and under what conditions.

For most UK boards, the practical default should be that board-only and privileged material is not entered into any AI tool unless the tool is an approved enterprise instance with a defined legal basis and the use has been authorised in advance. Confidential material may be processed only through an approved tenant with logging enabled. Public and internal material can be handled with fewer restrictions, but still within the organisation's acceptable-use policy.

The classification should be visible. A footer, header or metadata tag on each paper is more effective than a general instruction in a governance handbook. Directors should be able to see, at a glance, whether a document is eligible for AI-assisted summarisation.

Constrain the tool, not just the prompt

Prompt hygiene matters, but it is not a control on its own. Directors should not be asked to rely on their own judgement about what constitutes sensitive information in every instance. The organisation should constrain the tool.

That means an approved enterprise workspace rather than consumer accounts; single sign-on and role-based access; disabled plugins unless specifically assessed; restrictions on third-party extensions; data-loss prevention rules that recognise board-pack markers; and retention settings that match the organisation's records policy. Where the tool offers a training opt-out, the setting should be verified and recorded, not assumed.

Company secretaries should also consider the shared workspace itself. A board portal that incorporates an AI summary feature is not automatically safe because it is called a board portal. The same questions apply: where is the data processed, who can administer the workspace, what is retained, and can the organisation extract an audit trail?

Make use auditable

A confidentiality protocol without a log is a statement of intent, not a control. Boards should maintain a simple AI-use register for the paper cycle. It need not be burdensome. The register should record who used the tool, which document or category was processed, the purpose, the approval route, and where the output was stored. If an output is incorporated into a board paper, the register should link to the final version.

The register should be reviewed periodically by the company secretary and reported to the audit or governance committee in summary form. Patterns matter: repeated use of unapproved tools, summarisation of privileged material, or AI-generated content entering minutes without review are signals that the protocol is not working.

Test the contract before the next cycle

Many confidentiality failures will not be caused by directors at all. They will be caused by vendor terms that were never read. Before renewing or adopting an AI service that will touch board material, the organisation should test the contract against a short set of questions. Is board data used to train models? Which subprocessors are involved? Where is data stored and processed? What are the retention and deletion periods? What notice is given of a breach? Are audit rights available? Are UK GDPR processor terms in place, including international transfer mechanisms?

These questions should be answered in writing and retained. A vendor's marketing page is not a substitute for contractual assurance. If the answers are unclear, the tool should not be used for confidential or board-only material.

A director-level checklist

Before the next board pack is circulated, directors and the company secretary should be able to confirm the following:

• Every paper carries a classification marking, and the marking states whether AI processing is permitted.
• Only approved enterprise tools are used for confidential material, with consumer accounts and personal devices excluded.
• Plugins and extensions are disabled unless assessed and approved.
• The AI-use register is live, and entries are made at the point of use rather than reconstructed later.
• Vendor terms have been tested for training, retention, subprocessors and breach notification.
• AI-generated summaries are reviewed by a named individual before entering papers or minutes.
• Directors have received a short written instruction on the protocol, with a route to ask questions.
• The protocol has been reviewed by the audit or governance committee in the current governance cycle.

Red flags that confidentiality has slipped

Certain signals should prompt an immediate pause. These include board papers being pasted into tools that are not on the approved list; summaries appearing without a named reviewer; a vendor unable to explain whether board data is used for model training; shared workspaces with administrator access wider than the board secretariat; AI outputs stored in personal drives; and minutes containing language that no director recalls approving. Individually, each may look minor. Together, they suggest that the board has lost visibility of where its information is being processed.

The boardroom has always depended on controlled circulation. AI does not change that principle, but it does change the speed and scale of the risk. A protocol that classifies information, constrains tools, records use and tests contracts will not eliminate the risk. It will, however, give the chair and company secretary a defensible answer when the board asks who has read the pack, and where it has been.