Legal AI Vendor Due Diligence: A Checklist for Legal Platforms
A practical guide for legal services platforms and law firm teams on how to vet an AI vendor for security, data access, governance and production readiness before you sign.
Co-Founder & Director of Engineering
· 9 min read

On this page · 8 sections
Key takeaways
- Due diligence on a legal AI vendor covers four areas: security, client data access, governance and production readiness. Check each one in writing before you sign.
- ABA Formal Opinion 512 says lawyers must understand the tools they use, protect client information and review the output. Your vendor has to make those duties possible.
- Ask whether client data can train the vendor's model, where it is stored and who can see it. Put the answers in the contract.
- Test the tool on your own documents with a fixed set of questions. Stanford researchers found leading legal research tools wrong on more than 17% of test queries.
- Decide who owns the code, data and prompts at handover. A platform you cannot leave is a risk your clients carry too.
A legal AI tool reads privileged documents, drafts advice and cites the law. If it leaks a file or invents a case, the harm lands on the firm and its clients, not on the vendor. So the review you run before signing matters more than the demo.
This guide is for legal services platforms, legal tech teams and law firms choosing an AI vendor or build partner. It turns the questions buyers ask into a checklist you can use in a live selection. It covers four areas: security, client data access, governance and production readiness. If you are still deciding whether to buy a product or build your own system, start with our build vs buy vs partner decision matrix.
Why legal AI needs its own due diligence
Most software reviews ask whether the tool works and whether it is secure. Legal work adds two more questions. Can the firm still meet its duties to clients? And can a lawyer check every answer before it is used?
The American Bar Association answered part of this in Formal Opinion 512, issued on July 29, 2024. It applies the existing rules to generative AI tools. Lawyers need a reasonable understanding of what a tool can and cannot do. They must protect client information, which may mean getting the client's informed consent before putting their data into a tool that learns from it. They must review AI output, including citations, before relying on it. And they should read a tool's terms of use and privacy policy, or ask someone with security skills to do it.
None of this bans AI. It means the vendor you pick has to make those duties possible. A tool that hides its sources or keeps client data on unclear terms makes them harder to meet.
Area 1: Security
Start with how the vendor protects the system and the data in it.
- Independent audit. Ask for a recent SOC 2 report or ISO 27001 certificate, and read the scope. An audit that covers the vendor's sales tools but not the AI platform tells you little.
- Access control. Ask how users sign in, whether single sign-on is supported and how access is removed when someone leaves.
- Encryption. Data should be encrypted in transit and at rest. Ask who holds the keys.
- Logging. Every query, every document opened and every answer given should be logged with a time and a user. You will need this when a client or a court asks what happened.
- Incident response. Ask how fast the vendor tells you about a breach and what the contract says about it.
Area 2: Client data access
This is where legal buyers should spend the most time. The questions are simple. The answers often are not.
- Training. Can your data be used to train or improve the vendor's model? Get the answer in the contract, not in a sales email.
- Location. Where is the data stored and processed? Firms with clients in the EU, the UK or the Middle East may need it kept in a given region.
- Retention. How long are prompts, documents and answers kept, and can you delete them?
- Sub-processors. Which model providers and cloud services sit behind the tool? Each one is another party that touches client data.
- Matter walls. Can the tool keep one client's documents away from another team's questions? Ask to see it, not just hear about it.
If the answers do not fit your obligations, there are other options. A system can run in your own cloud account, or fully on your own servers, so client data never leaves your control. Our post on HIPAA and CCPA engineering problems shows how these rules turn into design choices.
Area 3: Governance
Governance is how the firm keeps control of the tool after launch. The NIST AI Risk Management Framework, released on January 26, 2023, is a useful neutral checklist. It is voluntary, and NIST added a profile for generative AI on July 26, 2024. Use it to ask the vendor:
- Who decides which tasks the tool may do, and who signs off on changes?
- How are model updates tested before they reach your users?
- Which steps always need a person, such as sending advice to a client or filing a document?
- How does a user report a wrong answer, and what happens next?
Rules like "a lawyer approves every client-facing answer" should sit in the workflow, outside the model, so they run every time. Our guide on securing AI agents covers how to set those limits for tools that can take actions.
Area 4: Production readiness
A demo on clean sample files proves very little. Before you sign, test the tool on your own work.
Accuracy on your documents. Build a test set of real questions with known answers from your own files. Run it, check every citation and count the wrong answers. This matters because even leading tools get things wrong. In a 2024 Stanford study, legal research tools from major vendors gave wrong or unsupported answers on more than 17% of test queries, and one tool was above 34%.
Citations you can check. Every answer should show the source passage it came from, so a lawyer can verify it in seconds.
Escalation. Ask what the tool does when it should not answer. A good system says so and hands the question to a person.
Integration. Ask which systems it connects to, such as document management, practice management or billing, and who fixes the link when one of them changes.
Ownership and exit. Ask who owns the code, prompts and data at the end. If you cannot export your data or move the system, you are locked in.
Pilot to production. Ask for a written plan from pilot to live use, with a date and a measure of success. Our post on proof of concept vs production explains why so many pilots stop there.
A one-page checklist
Send the same questions to every vendor and compare the written answers.
| Area | Ask for | Red flag |
|---|---|---|
| Security | Audit report with scope, access control, logs | No audit, or one that excludes the AI platform |
| Client data | Training, location, retention and sub-processors in the contract | "We don't train on your data" only in a sales email |
| Governance | Change process, human approval steps, error reporting | Model updates pushed with no testing or notice |
| Accuracy | Results on your own test set, with citations | Only a demo on the vendor's sample files |
| Ownership | Code, prompts and data handed over, export path | Data you cannot export, a platform you cannot leave |
Buying a product or building your own
Ready-made legal AI products are quick to start and suit common tasks like research and contract review. A custom build suits platforms with their own workflow, their own data or strict rules on where data lives. The checklist above applies to both. With a build partner, add three more questions: who writes the code, how the price is set and what you own at the end.
EdgeFirm is a build partner for legal AI development. The technical founders who scope the work also write the code, the price is fixed before coding begins and you receive all source code at go-live. Pricing starts from $20K, and the pricing page explains how scoping works.
Two of our legal builds show what the checklist looks like in practice. For LAWEP, a legislative drafting system, we built a citation engine that traces every clause back to its source, and the build ran for 5 months. The client reports use across 8 jurisdictions. For Inpro, a consumer legal guidance platform, we built ethical safeguards and an attorney escalation system, so cases the tool should not handle go to a person. The legal industry page covers more examples.
What to do next
Write one page that names the task, the documents the tool will read, the people who will use it and the steps that need a lawyer. Send it with the checklist to three vendors and compare the questions they ask back. A vendor that asks about your data, your duties and your systems is thinking about delivery.
If you want a second view on a shortlist or a build plan, book a 30-minute strategy call with the founders. You leave with a scope, a timeline and a straight answer.
Frequently asked questions
It is the review a legal team runs before it signs with an AI vendor. It checks security, who can see client data, how the tool is governed and whether it works on your own documents.
The opinion, issued on July 29, 2024, says lawyers need a reasonable understanding of the tools they use, must protect client information and must review the output. Pick a vendor that shows you how the tool works and lets you check every answer.
Only if the contract allows it, so read the terms. Ask the vendor to state in writing that your data is not used for training, where it is stored, how long it is kept and who can see it.
Build a set of real questions with known answers from your own documents. Run them through the tool, check every source it cites and count the wrong answers. Repeat the test after each model change.
You should. Ask for all source code, prompts, data and access at handover, so you can change vendors or run the system yourself. EdgeFirm hands over all source code at go-live.
Topics
- legal AI
- AI vendor due diligence
- AI governance
- law firm technology
Written by
Co-Founder & Director of Engineering
Usman brings 8+ years of experience building enterprise systems. He specializes in system architecture, DevOps, and data pipelines that power production AI.


