The objective is controlled use, not paperwork. Every control should connect to a real system, matter, decision or accountable person.
1. Build an inventory
Record approved and unapproved tools, owners, users, data categories, jurisdictions, integrations and use cases. Shadow AI is a governance signal: if the approved environment is unusable, staff will create an informal one.
2. Tier the use cases
| Tier | Examples | Control |
|---|---|---|
| Low | Formatting, public summaries, internal brainstorming | Approved tool and basic review |
| Moderate | Contract extraction, research, translation | Source verification, restricted data and sampling |
| High | Strategy ranking, rights-impacting recommendations, client-facing advice | Named lawyer approval, validation, explanation and complete logs |
| Prohibited | Autonomous filing, settlement, waiver, disclosure or destruction | Technical block, not policy language alone |
3. Create approval gates
Require documented approval for new tools, new data categories, new integrations and material model changes. A model update can change the risk profile even if the product name remains the same.
4. Govern vendors
Assess security, data processing, subprocessors, localisation, retention, audit rights, model changes, service continuity, intellectual property, government requests and incident response. Allocate responsibility for monitoring contract and architecture changes.
5. Validate and monitor
Test with representative tasks, including edge cases and adversarial inputs. Measure legal accuracy, citation validity, omission, bias, escalation and recovery—not only speed. Revalidate after significant model, prompt, data or workflow changes.
6. Preserve evidence
For high-risk workflows, preserve system version, prompt or instruction, retrieved sources, tool actions, output, reviewer, approval and external action. Logs should enable reconstruction without becoming an uncontrolled archive of confidential information.
7. Prepare for incidents
Define when to stop a workflow, isolate a vendor, preserve evidence, notify leadership, assess client harm, correct external outputs and meet legal notification duties.
The dashboard
- Approved systems and active high-risk use cases
- Percentage of matters using AI
- Validation failures and overridden recommendations
- Confidential-data events and blocked actions
- Vendor and model changes
- Client complaints or contested AI-supported outcomes
- Time saved—reported alongside quality, not instead of quality