Legal AI confidentiality: cloud, in-house systems and supplier checks
Check legal AI confidentiality across cloud and in-house systems: data flows, permissions, retention, support access and supplier terms before a pilot.
Before a firm puts client information into an AI workflow, it needs to understand the whole route: the source system, the AI service, its supporting suppliers, the output, the logs and the people who can access them.
A security page or a promise of “no training” answers only part of that question. Assess the particular product, contract and configuration the firm intends to use.
Begin with the information and the task
Write down what the workflow needs to receive and why. A callback-routing task may need a few contact facts. A medical chronology may involve extensive health information. Do not give both the same unrestricted access simply because they use the same model.
The SRA warns that both paid and free AI tools may lack appropriate safeguards for client confidentiality. It expects firms to understand contractual, technical and organisational protections appropriate to the information involved. Confidentiality, data protection and legal professional privilege need attention; a vendor's marketing description does not resolve them. SRA warning on AI.
The firm should obtain the legal and professional assessment appropriate to its proposed use. The following is a practical evidence-gathering method, not approval to upload matter files.
Request evidence for each stage
- 01Source system
- 02AI service
- 03Supporting suppliers
- 04Output and export
- 05Logs and support access
- 06Retention and exit
| Question | Evidence to request | Configuration to demonstrate |
|---|---|---|
| Which organisations receive the information? | Applicable supplier and sub-processor details | The actual integrations enabled for this account |
| Who can see a matter? | Access model and support-access arrangements | A user with no permission cannot retrieve its contents |
| What is retained? | Terms covering inputs, outputs, logs and backups | Account retention settings and deletion behaviour |
| Is information used for other purposes? | Contractual terms for training, improvement and other processing | Relevant controls for the purchased product |
| What happens when the firm leaves? | Export, return and deletion arrangements | An export the firm can read and reconcile |
These checks are implementation recommendations. Record unanswered questions rather than converting a supplier's silence into a positive answer.
Confirm the processing roles and terms
A controller decides the purposes and means of processing personal data; a processor acts on its behalf. Determine the roles for the proposed activities rather than assuming every AI supplier is a processor for everything it does.
The ICO explains that controller–processor contracts must contain required terms, including provisions concerning instructions, confidentiality, security, sub-processors and return or deletion of personal data. Its guidance is under review following the Data (Use and Access) Act, so check the current guidance and applicable law during procurement. ICO: contracts, ICO: contracts and liabilities.
Record processing locations, international access and any transfer arrangements for review. A storage-region label alone does not describe support access, logs or every supporting service.
Compare cloud, private hosting and in-house AI
“Sovereign AI” can describe different arrangements. Specify the boundary instead: who owns the equipment, where processing happens, who administers it and which services can receive information.
| Arrangement | What to establish |
|---|---|
| Hosted AI service | Product-specific data terms, retention, processing locations and support access |
| Private hosted environment | Hosting and model providers, administrator access, backups and any external services |
| Firm-owned local server | Compatible local software, document reading, storage, updates, recovery and all external connections |
A no-training commitment is different from keeping processing on site. OpenAI's business data terms and Anthropic's commercial retention explanation illustrate why the exact product and agreement matter.
For local operation, check document text extraction, search indexes, prompts, outputs, logs, backups, remote support and telemetry—the diagnostic information software may send to its supplier. A local model with a cloud document-reading service is not a wholly local workflow. Proprietary hosted assistants cannot simply be installed on a firm-owned server.
Evaluate representative tasks before buying hardware. Include concurrent users, scan quality, review effort, licensing and ongoing support. The NCSC secure AI guidance treats operation and maintenance as continuing responsibilities. Hardware ownership alone does not establish confidentiality or legal compliance.
Demonstrate a change in permission
Use fictional matters and two test accounts. Give one account access to a document and withhold it from the other. Try search, direct links, summaries and exports. Then remove the first account's access and repeat the exercise.
This checks a particular configuration. It is not proof of the vendor's entire security programme. The supplier should explain propagation delays, cached outputs and what happens to previously exported copies. Agree how the firm will handle those limits.
Include supplier support in the discussion. An administrator's ability to retrieve content may be necessary for some tasks, but it needs an understood purpose, appropriate control and an accountable route.
Make approval specific and reviewable
Record the approved task, permitted data, product, settings, access rules, reviewer and conditions that would trigger another assessment. A new integration or a changed supplier term may alter the original decision.
Start evaluation with synthetic material. Use the supervision procedure to define how outputs are checked, and the legal AI comparison to ask competing suppliers the same questions. The legal hub connects these controls with the workflow the firm wants to improve.