Four questions that decide where client data ends up
Before you sign for a contract-review assistant, research copilot or drafting tool, you should be able to say exactly what happens to a client's documents after someone uploads them. Most diligence on legal AI tools confidentiality comes down to four questions. Does the vendor, or the model provider behind it, train on your data? How long are prompts, files and outputs kept? Which third parties touch them? And in which countries are they stored and processed?
Vendor marketing tends to fold all four into one reassuring line, usually "enterprise-grade security." Pull them apart. A vendor can truthfully say it never trains on customer data and still keep your prompts for weeks, send them to a model provider on another continent and record them in a third-party monitoring tool. Most generative AI products for law firms are built in layers: the app you buy runs on a cloud host and calls a foundation model. Each layer has its own retention and location settings, and you need answers for every one of them.
Looking to streamline vendor RFP selection?
Access pre-negotiated tier discounts through our independent institutional network.
Model training: what "we don't train on your data" actually covers
Training is the one risk you can't undo. Once a client's material has shaped a model, fragments could in principle show up in outputs for other customers, and you can't get them back. That's why the major model providers' business and API terms generally leave customer content out of training by default. Consumer chat apps work differently: on personal accounts, training is often on unless the user turns it off.
The clause you need is narrower and more exact than the slogan. It should bind the vendor and every model provider it uses. It should cover fine-tuning, evaluation datasets and "product improvement," not only foundation-model training. And it should say what happens to feedback data, such as ratings and corrected outputs, because some vendors treat that as their own. If the vendor offers a custom model trained on your firm's precedents, get it in writing that only your tenant uses that model and that it's deleted when the contract ends.
Retention, logs and the limits of zero data retention
Most of the real exposure is in retention. Prompts and outputs can sit in four places: the vendor's application database (chat history, workspaces), the model provider's abuse-monitoring logs, backups, and observability or support tools. Model providers commonly keep API inputs and outputs for a limited period for abuse monitoring, often up to about 30 days, unless the customer has been approved for an exception. Check the current terms for each provider, because they change.
Zero data retention (ZDR) usually means the model provider processes each request without writing the content to persistent storage. It applies to the model layer only. It doesn't cover the legal app's own history, document store or backups. So ask two separate questions. Is ZDR in place with each model provider? And what does the vendor itself keep, for how long, and can you shorten it? For most firms, a sensible target is workspace retention you can configure to match the matter lifecycle, backups that expire within roughly 30–90 days, and written confirmation of deletion when the contract ends.
Sub-processors: mapping the chain behind the app
A typical legal AI product might rely on one or two model providers, a hyperscale cloud host, a vector database, an OCR or transcription service, an error-tracking tool and a support desk platform. Under GDPR-style laws, every one of them that touches personal data is a sub-processor. Your contract should give you a current list, advance notice of changes (often 14–30 days) and a right to object, or to terminate, if a new sub-processor isn't acceptable.
Ask for the list broken down by data type. A support tool that only sees account emails carries far less risk than a service that receives full document text. Also find out whether the vendor calls the model directly from its developer or through a cloud platform such as Azure, AWS or Google Cloud. That choice often decides which regions are available and which retention terms apply.
Looking to streamline vendor RFP selection?
Access pre-negotiated tier discounts through our independent institutional network.
Data residency is not the same as no transfer
Data residency means data is stored at rest in a named region. It doesn't mean processing happens there. A vendor can keep your files in Frankfurt and still send each prompt to a model endpoint elsewhere, or let support engineers in another country access the system remotely. Under data-protection law, both of those count as transfers.
Get three things in writing: the storage region, the inference (processing) region, and every location from which staff and sub-processors can reach content. In-region inference is widely available in the EU, the UK and parts of Asia-Pacific. Coverage is patchier in the Gulf, Africa and Latin America, where the newest models can arrive in local cloud regions later than elsewhere.
Cost matters here too. Typical 2026 market pricing for legal AI seats runs very roughly from USD 50 to several hundred per user per month, depending on the product and the scope. Dedicated regional hosting or single-tenant deployments usually come with higher seat minimums or a premium, sometimes in the range of 10–30%. These figures vary widely by contract and country, so treat them as indicative and confirm with vendors.
How secrecy and data-protection rules apply by region
Two sets of rules apply together. Professional-secrecy duties protect the client relationship. Data-protection laws protect the individuals named in the files. The table below is a general orientation only. Local counsel or your bar regulator should confirm how these rules apply to your firm.
| Jurisdiction | Data-protection framework | Professional-secrecy angle | What it means for AI vendors |
|---|---|---|---|
| EU | GDPR: Art. 28 processor terms; Chapter V transfer rules (SCCs, adequacy decisions) | National secrecy laws, some carrying criminal penalties (e.g., Section 203 of Germany's Criminal Code) | Processor agreement, a transfer mechanism and, in some member states, written secrecy undertakings from service providers |
| UK | UK GDPR and Data Protection Act 2018; IDTA or UK Addendum for transfers | SRA confidentiality duty; legal professional privilege | Similar to the EU; confirm that third-party access doesn't put privilege at risk |
| India | Digital Personal Data Protection Act 2023, with implementing Rules being phased in | Bar Council rules; statutory protection of professional communications | Contract terms with processors; watch for any transfer destinations the government restricts |
| UAE (onshore, DIFC, ADGM) | Federal PDPL; separate DIFC and ADGM data protection regimes | Lawyer confidentiality under federal and free-zone rules | Establish which regime applies; free zones have their own transfer rules |
| Saudi Arabia | PDPL and its transfer regulations | Confidentiality duties under the legal-profession law | Transfers outside the Kingdom need a legal basis and safeguards |
| Singapore, Brazil, South Africa | PDPA, LGPD, POPIA | Bar and law-society rules | Processor contracts, plus comparable protection for cross-border transfers |
The same rule holds in every jurisdiction: you remain responsible for how the vendor handles client data. Anything the contract doesn't cover is your risk.
Nine contract terms to insist on
- No-training clause that binds the vendor and all model providers, and covers fine-tuning, evaluation and feedback data.
- Retention schedule for every store (app, model logs, backups, support tools), with ZDR at the model layer where it's offered.
- Named regions for storage, inference and staff access, with an obligation to notify you before any change.
- Sub-processor list showing data types, with advance notice of changes and a right to terminate.
- Confidentiality undertaking that meets professional-secrecy standards. A generic NDA isn't enough.
- Breach notification within 24–72 hours, so you can meet your own regulatory deadlines.
- Audit evidence: ISO 27001 certification and/or a SOC 2 Type II report, plus a penetration-test summary.
- Deletion and export at exit, certified in writing within a fixed period.
- Government-access clause requiring the vendor to notify you of requests and challenge them where it lawfully can.
Red flags, and a shortlisting rule for 10–200-lawyer firms
Walk away, or escalate, if the vendor can't name its model providers, describes retention only as "industry standard," offers residency for storage but won't commit on inference, or limits its no-training promise to the foundation model. Staff putting client work into consumer accounts is a separate risk. Handle it through internal policy and training. A vendor contract won't fix it.
For most firms with 10–200 lawyers, the rule is simple. Shortlist only vendors that put all nine terms in the signed contract, not just on a trust page. Then check each practice group's major clients: would they, or their regulator, accept the named inference region? If even one major client wouldn't, make that a negotiating requirement before you discuss price.