Why Generic AI Tools Fail Law Firms

A promising demo, a disappointing rollout, a tool nobody opens three months later. The usual cause is not the technology. It is what the tool is built to work from.

The TeamPublished 5 min read

Many law firms that pilot an AI tool go through the same arc: an impressive demo, a disappointing rollout, and a tool that quietly stops being used within a few months.

The problem is rarely that AI can’t help with legal work. It is that many tools are built around a general-purpose language model, not around the materials, legal rules and verification that make legal work reliable. A plausible answer is not the same as a usable work product. To be useful, AI has to work from the actual record, connect its conclusions to sources an attorney can check, and leave the judgment with the attorney.

What a generic model knows, and what it doesn’t

A general-purpose language model has broad knowledge of language and information. It can summarize, draft and answer questions. But it doesn’t know the facts of a particular matter, the documents in a client’s file, or how a specific legal requirement applies to the evidence in front of it.

A legal-looking interface doesn’t change that. Without the record and a structured way to evaluate it, the model can produce an answer that sounds convincing without establishing whether the evidence supports it.

Where it breaks down in practice

  • It lacks the case. A model can’t evaluate evidence it has never seen. It may offer a plausible assessment without considering the actual documents, the missing corroboration or the conflicting facts in the file.
  • It confuses a claim with evidence. A recommendation letter may describe an achievement; that doesn’t establish its significance. A document that mentions an award doesn’t show how the award is recognized.
  • It can’t show where a conclusion came from. A fluent explanation is not a citation to the document and passage that support it. If the attorney can’t check it, the attorney has to redo it.
  • It misses contradictions across documents. Dates, job titles, affiliations and descriptions of achievements can differ between a CV, recommendation letters and supporting records. Reading each document on its own won’t reveal them.
  • It raises confidentiality questions. Lawyers need to know how a provider handles client information: who can access it, whether it is kept, and whether it is used to train models. Under ABA Formal Opinion 512, a lawyer needs the client’s informed consent before entering information about a representation into a self-learning tool, because what goes in can surface for other users; boilerplate in an engagement letter is not enough.

What a grounded system changes

A system built around the case record works differently. It reads the supporting documents, extracts the relevant facts, evaluates the evidence against the applicable legal framework and ties each finding to its source. That changes the work:

  • Evidence is mapped to specific legal criteria, not just summarized.
  • Every finding can be checked against the page and passage it relies on.
  • Missing corroboration and conflicting information become items to review, not surprises.
  • The system separates what the record supports from what still has to be established.
  • The attorney decides what the assessment says.

Retrieving the right documents is not the same as showing that a conclusion is supported by them. The second is the work that matters.

What grounding does not fix

A Stanford study published in the Journal of Empirical Legal Studies tested the leading AI legal research tools from LexisNexis and Thomson Reuters, as they stood in 2024, which retrieve from legal sources before they answer. They made fewer errors than general-purpose GPT-4, but still produced incorrect or misgrounded answers between 17% and 33% of the time.

The study tested legal research questions, not the review of a client’s case file. Its lesson still applies: connecting an AI system to source material doesn’t guarantee that its conclusions are right.

That is why traceability matters. A system should tie each finding to its source, check that quoted passages really appear there, and make verification part of the workflow. A citation doesn’t prove an interpretation is correct. It lets the attorney examine the evidence behind it.

Why demos hide the problem

A generic model in a polished interface can look impressive in a scripted demo built on clean, simple examples. The problems appear later, with a large file, incomplete documents, conflicting facts and evidence that has to be measured against specific legal requirements.

An attorney doesn’t need another chatbot that can summarize a document. The attorney needs to know what the record establishes, what it doesn’t, and where to look to verify each finding.

A test worth running

When you evaluate an AI tool, ask the vendor to run it on a closed case your firm already knows, and compare its findings with your own assessment:

  • Did it find the relevant evidence?
  • Did it map the evidence to the right legal elements?
  • Did it notice missing corroboration?
  • Did it catch contradictions across documents?
  • Can every finding be traced to the original page?
  • Does it separate supported conclusions from open questions?

Watch what happens where the evidence is thin. A system that says clearly what it can’t establish is more useful than one that fills every gap with a confident answer.

The principle

AI is useful in legal practice when it shortens the distance between the evidence in a file and the work an attorney has to do to evaluate it. That takes more than a language model in a legal interface: it takes the actual record, a structured understanding of the legal framework, citations that can be checked, and a workflow that keeps the attorney in control.

This is how we build Case Intelligence for immigration law firms. It reads an applicant’s documents, maps the evidence to the criteria of the visa, finds gaps and contradictions, and cites every finding to its source page. Each firm’s data is isolated and never used to train AI models, and nothing is final until an attorney accepts it.

Sources

  1. ABA Formal Opinion 512 — Generative Artificial Intelligence Tools (July 29, 2024)
  2. Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D., & Ho, D. E. (2025). “Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools.” Journal of Empirical Legal Studies, 22, 216–242

Related notes