What should we ask an AI litigation vendor about its data?

What are the core data questions to ask any AI litigation vendor?

Five, and they are short. What is the AI grounded in when it works on our matters. Does anything we upload train your models or anyone else's. How is our data segregated from other customers. What are the retention and deletion terms when the contract ends. And can every output be traced to the source document it came from. Good vendors answer each in a sentence.

Does our case data train the vendor's model?

This is the question that decides privilege exposure, so get it in the contract, not the demo. Require zero data retention and an explicit bar on your matter data training any model. Courts have split on whether AI use waives privilege, and the tier and the contract decide the outcome: tools bound to confidentiality sit on far safer ground than tools that train on inputs.

• Require an explicit contractual bar on your case data training the vendor's models or anyone else's.
• Require zero data retention, with named deletion terms when the engagement ends.
• Require confidentiality language that protects privilege, binding the vendor the way outside counsel is bound.
• Treat consumer-tier tools that train on inputs as disqualified for matter data, whatever the feature list says.

The privilege case law is unsettled, and that is exactly why the contract matters. Write the standard so it holds regardless of which way a given court comes out: your data does not train models, does not persist past the engagement, and stays under confidentiality terms you can enforce.

How should the vendor prove its outputs are grounded?

By showing sources, not scores. Reliable legal AI does a narrow job on your own case data and links every output back to the document it came from, so a human can verify it in seconds. Courts are sanctioning unverified AI output in real dollars, which makes an untraceable answer a liability. If you cannot click a fact and see its source, you are being asked to trust a black box.

• Every extracted fact traces to its source document, so nothing rests on the model's say-so.
• The AI works only on your own case data, never a borrowed benchmark or someone else's portfolio.
• Each AI product does one narrow, checkable job instead of one vague, unauditable one.
• Ask what the AI refuses to do: a vendor that claims its model predicts case outcomes is selling inference you cannot verify.

That last point is a useful screen. CaseGlide, for example, uses AI to read and structure the litigation record, and it does not predict verdicts or score outcome risk, by design. Grounded extraction can always be checked against a source. Outcome prediction cannot, and unverifiable inference is exactly where fabricated output hides.

What contract terms lock these answers in?

Put the data answers in the master agreement so they survive the sales cycle. Require current SOC 2 audit coverage with security documentation available on request, zero data retention, a bar on training with your data, segregation of your matter data, and confidentiality that protects privilege. A demo answer is a claim. A contract term is a control you can enforce.

1. Audit coverage: current SOC 2, with security documentation available on request.
2. Zero data retention: your matter data does not persist past the engagement, with deletion terms in writing.
3. No training on your data: an explicit bar, not an opt-out buried in settings.
4. Segregation: your data stays yours, never pooled into cross-customer benchmarks without your consent.
5. Privilege protection: confidentiality terms binding the vendor, so privileged work product keeps its protection.