A law firm's AI policy: the CAPTURE method

A useful AI policy comes down to a simple system: named uses, approved tools, an owner for every decision and retained evidence. It should neither prohibit AI in the abstract nor turn every trial into a compliance project. For a law firm, its aim is operational: enable useful applications while protecting information, requiring human review and making decisions auditable.

The signed document is therefore only the visible part. The real system includes a tool list, use-case register, data matrix, approvals, incident procedure and periodic review. Before drafting it, a firm AI maturity assessment identifies practices already in place, including personal accounts and informal experiments.

1. Set the scope: uses, data and tools

Start by listing actual uses: generating ideas, summarising, translating, preparing an outline, document analysis, assisted drafting, automation or connected agents. The same tool has different risks depending on whether it receives public text, an internal document or material relating to a matter.

The policy can establish a three-level rule:

• Green: public, fictional or previously redacted data; permitted use in allowlisted tools.
• Amber: internal information or necessary personal data; use only for an approved case, with minimisation, controlled access and approved settings.
• Red: information covered by professional secrecy, sensitive data, strategy or documents from a matter; no input by default. An exception requires an expressly approved environment and use case after technical, contractual, professional ethics and GDPR review.

This classification must specify permitted operations, not just document categories. Copying an extract, uploading a document, connecting a document workspace and allowing an agent to act create four different exposures. For more on this, see generative AI and professional secrecy.

CNIL's AI and GDPR guidance, consulted on 22 July 2026, recommends in particular defining a specified, legitimate and explicit purpose, limiting data to what is necessary, setting proportionate retention and evaluating the system in production. In practice: each approved use case fits on a sheet stating the objective, users, data, tool, expected output, duration and controls.

2. Write actionable rules, not decorative principles

A person should be able to determine in less than a minute whether they can act. The policy must therefore answer six questions: which account to use, which data to enter, which features to enable, who reviews, where to save the result and how to report a problem.

• Use only an administered professional account and the tool in its approved configuration.
• Do not connect email, a drive, browser or document database without specific authorisation and permission testing.
• Reduce data to the strict minimum; prefer fictional data, masking or pseudonymisation where the result remains useful.
• Treat every output as a draft: verify facts, citations, calculations, applicable law, consistency and bias before reuse.
• Prohibit an output alone from triggering a legal act, sending, deadline, advice or decision affecting a person.
• Immediately report disclosure, excessive access, abnormal responses, fabricated sources or unplanned actions.

Human review should be defined by risk: reviewer identity, checks and approval record. For a clause workflow, for example, the method differs from simple rewording; the article automating contract drafting with AI details this control logic.

The CNB's generative AI resource stresses controlled integration respecting ethics, confidentiality and responsibility, and provides a framework attentive to sovereignty, security, confidentiality and possible data reuse. An internal policy must translate these concerns into enabled or disabled features, permissions and acceptance criteria.

3. Assign roles and decision-making powers

In a small organisation, one person may hold several roles; decisions must nevertheless remain identifiable.

• Management or sponsoring partner: approves the accepted risk level, tools and exceptions.
• AI lead: maintains the register, organises tests, publishes instructions and monitors product changes.
• DPO or privacy adviser, where involved: assesses processing, legal basis, information, rights, retention periods and any need for an impact assessment.
• IT/CISO: validates identity, access, logging, connectors, export, deletion, backups and the incident procedure.
• Practice owner: sets quality criteria and human review for the use case.
• User: respects scope, checks output and reports any deviation.

Training is part of the system. Article 4 of the European regulation requires providers and deployers to take measures, to the best of their ability, to ensure sufficient AI literacy adapted to knowledge, experience and context. Under Article 113, Chapters I and II have applied since 2 February 2025, while general application begins on 2 August 2026, subject to the reservations in the text. It is therefore necessary to check the current version and classification of each system in Regulation (EU) 2024/1689 on EUR-Lex, without assuming all generative uses fall into one category.

Evidence of literacy may combine role-specific training, an exercise using a fictional case, a short assessment, an attendance sheet and a refresher after a significant change. CNIL recommends involving and training users, tailoring content to their functions, assessing understanding and documenting procedures kept up to date.

4. The CAPTURE method: seven decisions, seven pieces of evidence

To move from text to oversight, Initial proposes the CAPTURE sequence. Each stage produces a dated deliverable.

The supplier file must record the product, plan, configuration and contractual version reviewed. A dated example, not transferable to a consumer offering: Anthropic's commercial terms displayed when we consulted them, effective 17 June 2025, state in particular that customer content is not used to train models in the relevant services, that a DPA applies and that this content is treated as confidential. This does not remove the need to examine the actual subscribed service, its settings, processors, transfers, retention periods and changes.

5. Organise evidence without creating another leak

Retaining evidence does not mean archiving every prompt. The register may reference the use case, tool, authorised user, date, data type, check performed and any incident, without copying information from a matter. Access to evidence, retention and deletion must also be defined.

In an incident: stop the affected use, preserve necessary facts, alert designated roles, limit exposure, ask the supplier about available measures, then assess applicable obligations under the firm's procedure. Lessons learned should lead to changes in a rule, configuration or training. Secure AI deployment translates these requirements into settings and tests; oversight and continuous improvement then maintain the register and controls.

Indicators should support decisions: proportion of inventoried tools with an owner, use cases with current sheets, authorised people who completed required training, uncorrected failed tests, open incidents, time to revoke access and latest supplier review date. These are not universal targets: the firm sets thresholds and cadence according to its risks. A good AI policy proves, at a given date, why a use was authorised, how it is controlled and who can stop it.