Product

Solutions

Security

Company

Resources

Request a Demo

Product

Solutions

Security

Company

Resources

Request a Demo

Article

2026-11-03

Last updated

Nov 3, 2026

10 questions to ask before choosing a litigation AI platform

Ask litigation AI vendors these 10 questions before signing. Test source citations, data protections, security evidence, fit, and exit rights.

Consider a hypothetical 45-lawyer firm: plaintiff personal injury lawyers, a mostly defense commercial group, and an innovation committee with two demos Thursday and a security questionnaire due Friday. The demo files will be clean. The committee chair is thinking about a construction matter’s fourth production and an expert report citing a missing site log. Will the tool understand that file next month, and what will the contract promise if it does not?

To choose a litigation AI platform, ask every vendor ten questions about source access, unanswered questions, whole-file coverage and updates, contractual data protections, independent security evidence, system fit and limits, human review, and testing. Judge the answers on your own closed matter, not the demo file. At Turbo Law, we expect buyers to hold us to the same questions.

The American Bar Association’s 2024 Legal Technology Survey, reported in its AI TechReport, found concerns concentrated around accuracy (74.7%), reliability (56.3%), and data privacy and security (47.2%). Those are sensible places to press.

1. What job will it do on our matters, and what will it not do?

It should perform named work your team actually repeats, with explicit limits. Start with a deposition digest or discovery response, then ask what the lawyer must supply and check. Make the presenter choose.

A useful answer identifies the deliverable, the records needed, and what remains outside the product. “We replace everything” should stop the conversation until the vendor explains exactly what that means for your practice, including work that still requires another system or a lawyer’s judgment.

This is AI vendor due diligence at the matter level. Our discussion of when a litigation platform beats a general AI assistant addresses that distinction; the demo should show whether it matters for your assignment.

2. Can every answer open to the page and line it came from?

A factual answer should open directly to its supporting passage. Ask the presenter to click through from five randomly chosen answers, including a factual assertion embedded in a longer narrative.

Read the surrounding text. A document-level footnote can leave the associate searching for support that is absent or qualified. Citation verification requires checking what the passage actually supports, which is why a citation is not verification.

District of Kansas Standing Order 26-01 (January 28, 2026) reaches factual backgrounds as well as authorities; District of Maryland Standing Order 2026-03 (August 5, 2026) likewise warns about unsupported factual assertions. These are individual courts’ orders, not national rules.

A presenter who cannot open the underlying passage has not demonstrated the checking process.

3. What does it say when the file does not answer the question?

It should identify the gap, and show conflicting records without silently choosing between them. A confident answer is no substitute for evidence.

Buyers understandably ask about a hallucination rate. For this meeting, ask something more revealing: what happens when the expert references a site log that was never supplied?

Then introduce a second date of loss. A useful response shows where each date appears and leaves the discrepancy visible for review. An unexplained choice should pause the purchase.

Our Assistant is designed to say when the file does not support an answer. Test that behavior on your records, including questions whose answers you know are missing.

1. What job will it do on our matters, and what will it not do?

2. Can every answer open to the page and line it came from?

A factual answer should open directly to its supporting passage. Ask the presenter to click through from five randomly chosen answers, including a factual assertion embedded in a longer narrative.

A presenter who cannot open the underlying passage has not demonstrated the checking process.

3. What does it say when the file does not answer the question?

It should identify the gap, and show conflicting records without silently choosing between them. A confident answer is no substitute for evidence.

Buyers understandably ask about a hallucination rate. For this meeting, ask something more revealing: what happens when the expert references a site log that was never supplied?

Then introduce a second date of loss. A useful response shows where each date appears and leaves the discrepancy visible for review. An unexplained choice should pause the purchase.

Our Assistant is designed to say when the file does not support an answer. Test that behavior on your records, including questions whose answers you know are missing.

4. Does it read the whole matter file, and stay current when records arrive?

It should account for the supplied matter file and update its understanding when records change. Ask what it excludes.

The scope should cover pleadings and productions, correspondence, transcripts with errata and exhibits, medical and expert records, and the team’s existing work product. Otherwise, the deposition outline may rest on a different account from the case chronology.

In Turbo Law’s Review, facts connect to their source pages and lines, and the connected picture updates as new records arrive. Our unified litigation workspace uses that picture across the work.

Ask separately whether earlier answers or exported drafts need refreshing. A vendor who treats new uploads as unrelated summaries has not answered the whole-matter question.

5. Is “no training on our data” written into the contract?

The promise belongs in the signed agreement and must cover downstream model providers. Read the clause.

The District of Oregon’s AI Product Evaluation Checklist (May 7, 2026), presented at its district court conference, states: “Vendor marketing materials are not contractual commitments”. It calls for enforceable terms covering the assurances. This is a conference checklist, not a court rule.

A no-training commitment should sit beside answers about client separation, retention and deletion. If the salesperson promises confidentiality but cannot identify the binding terms, stop.

ABA Formal Opinion 512 (July 29, 2024) addresses understanding tool limits, protecting confidences and checking output. Separately assess whether AI use is privileged.

Our security commitments prohibit training on client data by us or our model vendors, enforced contractually and technically.

6. What independent security evidence can you show us?

Ask for the independent report and its scope, not a badge. The malpractice insurer Lawyers Mutual’s Generative AI Vendor Vetting Checklist asks whether the vendor is “SOC 2 Type II certified (or provides equivalent independent security audit)”.

Treat SOC 2 Type II as evidence to examine. Match the covered service to what you are buying.

Protected health information appears in plaintiff, defense and carrier files alike, including injury and employment matters. Determine whether a business associate agreement is required. Our Security page states that we execute these agreements on request and log access, queries and exports.

The Oregon checklist also addresses encryption and access controls. Ask about SSO, role-based access, and your client’s outside counsel guidelines. Refusing the firm’s security questionnaire leaves the review unfinished.

7. Does it work with the systems we already use?

The vendor should explain how documents enter and return, who maintains the connection, and where the authoritative records remain. “We integrate” is an opening answer.

Your document management system and case-management system can remain the systems of record while the litigation workspace uses their contents. Ask the presenter to distinguish a native connection from a manual export.

Then follow a production. The eDiscovery platform may be chosen by the client or its service provider; reviewed documents must arrive with their Bates numbers and useful context intact.

Good answer: A demonstrated transfer, named maintenance responsibility, and a clear account of failures.
Red flag: A connection promised for later, with no usable process today.

8. Who checks the output before it leaves the firm?

The responsible lawyer checks it, with a defined review step before external use. Under Federal Rule of Civil Procedure 11(b), signing certifies, after reasonable inquiry, that “the factual contentions have evidentiary support”, subject to the rule’s qualification for specifically identified contentions awaiting investigation or discovery.

A suggestion to the federal Advisory Committee on Civil Rules sought an additional certification about legal authorities. The committee’s October 21, 2026 agenda memo recommended dropping it because existing Rule 11 appeared sufficient; that recommendation does not establish a meeting decision.

Ask where the reviewer opens sources and the partner approves release. Our Assistant prepares work product; it does not serve, file or send.

Reject an undefined handoff. Check any applicable AI standing order and our discussion of AI disclosure rules for court filings. Follow your jurisdiction’s rules, orders and case law. Turbo Law does not give legal advice.

9. What should we keep alongside it?

Keep the services whose jobs the new product does not perform. Many firms retain an eDiscovery platform and a research service alongside a litigation workspace.

Ask about duplication, too. The Lawyers Mutual checklist asks buyers to consider whether something already in their environment can perform substantially similar work.

Turbo Law is not an eDiscovery hosting and production platform, a citator, or a case-management or billing system. We flag potential privilege and work product; we do not build or serve privilege logs.

Our explanation of the case record layer beside legal research sets out that boundary. A credible vendor names what stays. An answer that dismisses those dependencies without examining your actual work should stop the deal.

10. How do we test it, and how hard is it to leave?

Test on a closed matter with known answers, and establish exit rights before uploading. A golden dataset gives reviewers an agreed reference. Use our guide to designing a pilot on your own matter for the full plan.

Add a late production and a conflicting detail. Confirm the tool exposes both the change and the uncertainty.
Ask what setup requires from IT and the matter team. Get dependencies and responsibility in writing.
Ask how cost scales with document volume and users.
Require usable exports of data, prompts, outputs and usage history, consistent with the Oregon checklist. Refusal is a reason to stop.

A plaintiff group might emphasize demand packages; a defense group, client and carrier reporting. Both need the cited record. Require a sample export before signing.

FAQ

Which AI tool should a litigation team choose?

Choose the one that performs your defined assignment with evidence your team can check and terms your firm can accept. A polished demonstration does not establish either. Have the lawyers who will use it assess the closed-matter results.

Is SOC 2 Type II enough for a law firm?

No. Examine the report’s scope and exceptions, then compare the contract and actual controls with your firm’s obligations. A report does not answer whether a particular client permits the proposed use.

Do plaintiff and defense firms need different questions?

The same ten questions apply. Emphasis changes with the records and deliverables: an intake team may care about missing medical records, while another team needs support for a client report. Neither can dispense with source checking.

Should we ask for a hallucination rate?

Ask how the vendor evaluates errors, but do not let a headline number replace observation. Give it an unanswerable question and conflicting records. Inspect whether it exposes the problem or writes past it.

How long should a litigation AI pilot run?

Long enough to include a later production and attorney review of the resulting work. Calendar time alone proves little. Before approval, record the failures, the vendor’s response, and any limits the firm would have to accept.

Sources

10. How do we test it, and how hard is it to leave?

Add a late production and a conflicting detail. Confirm the tool exposes both the change and the uncertainty.
Ask what setup requires from IT and the matter team. Get dependencies and responsibility in writing.
Ask how cost scales with document volume and users.
Require usable exports of data, prompts, outputs and usage history, consistent with the Oregon checklist. Refusal is a reason to stop.

A plaintiff group might emphasize demand packages; a defense group, client and carrier reporting. Both need the cited record. Require a sample export before signing.