Verification in Legal AI Is a Design Problem
Lawyers remain responsible. Legal AI should make its work easy to check before they sign.
The Perfect AI Is Actually a Combination of Tools: A Litigation Guide
A practical guide to matching legal research, eDiscovery, factual review and drafting tools to the task each performs best.
A question came up repeatedly in our July 2026 webinar with US litigators: which AI tool should I use for each part of a matter?
Catherine Thompson put it plainly: “In my mind, the perfect AI tool right now is actually a combination of tools.”
Trying to make one platform research the law, review a production, build the factual record, draft submissions and manage the matter equally well is a poor starting point. Those jobs use different inputs, carry different risks and need different forms of review.
Start with the task
A useful product comparison begins with a defined legal task. “Help with litigation” is too broad. “Find the current authorities on this issue”, “identify the factual record across these 8,000 pages” and “produce a first draft from an approved set of facts” are different jobs.
The questions also change with the job:
- What information must the system receive?
- What should the output contain?
- How will a lawyer check it?
- What happens when the lawyer corrects it?
- Does the result need to persist for the rest of the matter?
Legal research tools
Research products are built to find and analyze legal authority. Their value depends on coverage of the relevant law, the quality of retrieval, treatment of citators and the ability to show the authority behind a proposition.
For a research task, test whether the product finds controlling and adverse authorities, distinguishes holdings from party arguments and represents the authority accurately. A link to a real case is useful, but the proposition still needs to be checked against the case. The Bluebook study discussed here is a good example of why the source can be real while the output is still wrong.
eDiscovery and large-scale document review
eDiscovery systems are built for collection, processing, search, review and production across large document populations. They are the right environment for questions about custodians, review sets, privilege review and production status.
A product can be excellent at locating responsive documents without building a persistent factual account of the matter. Those are related capabilities, but they do not produce the same thing.
The factual record
Litigation teams also need a view across the documents: the people, events, dates, conflicts and gaps that make up the matter.
This work should remain traceable to the underlying evidence. Lawyers need to see the exact source for a factual claim, correct it when necessary and carry the corrected record into later work rather than rebuilding it for every prompt.
Mary is designed for this job. It reads the matter documents, builds a structured factual record and keeps each fact connected to its source. Chronologies, bank analysis and draft work product can then be produced from that record.
General drafting and reasoning tools
General-purpose models can be useful for outlining, reframing, editing and exploring alternative arguments, especially when the lawyer controls the inputs.
The review burden depends on what material the model is using. A draft produced from a lawyer-approved factual record is easier to assess than one produced after the model independently retrieves and reconstructs the matter.
Firms should also decide what information may be entered into each environment, whether it is retained, how access is controlled and whether the product's contractual terms match the firm's obligations.
Practice and document management systems
The system of record for files, matters, permissions and administrative work remains important when specialist AI products sit alongside it. Integrations reduce manual exports and make it easier to keep work inside approved systems.
Mary's Actionstep integration is one example. The useful question is whether each product strengthens the existing environment or creates another disconnected copy of the matter.
How to test a legal AI product
Use a closed matter that the team already knows. Give competing products the same documents and instructions.
For each task, record:
- what the product found;
- what it missed;
- whether claims were supported by the source;
- how long review took;
- whether corrections persisted;
- how much setup and support were required.
The known-matter test gives firms a more useful basis for comparison than the model name displayed underneath the product.
A workable legal AI stack
A litigation team may use one product for research, another for eDiscovery, a factual-record system for the matter and a general model for controlled drafting. The boundary between them should be deliberate.
The right stack gives each system a defined job, keeps client material inside approved environments and makes the resulting work practical to verify.