Product

Solutions

Security

Company

Resources

Request a Demo

Product

Solutions

Security

Company

Resources

Request a Demo

Article

2026-06-16

Document summaries are not case understanding

Early case assessment is not early data assessment, and a stack of summaries is not case understanding. Here is how we separate those jobs at Turbo Law.

Document summarization can tell you what a file says. Early Case Assessment asks a different question: what does the matter mean? At Turbo Law, we treat those as separate jobs. Summaries are useful for bounded documents. Matter-level analysis must reconcile allegations, evidence, chronology, identity, law, and contradictions across a changing record, with conclusions traceable to their sources.

Early Case Assessment is not Early Data Assessment

The distinction matters because the legal industry has spent years using similar language for two very different problems.

The Sedona Conference defines Early Case Assessment, or ECA, as assessing the merits of a case early in the litigation lifecycle to determine its viability. It may involve data collection and review, but it does not have to.

Early Data Assessment, or EDA, is different. It is the process of separating potentially relevant electronically stored information from nonrelevant ESI early in a matter, using techniques such as filtering, analytics, and human-assisted judgments. Its purpose can be to reduce the volume that must be processed and reviewed. (The Sedona Conference)

That is a useful boundary.

EDA asks questions such as:

What information do we have?
Which custodians matter?
What date ranges are relevant?
What should be collected or reviewed?

ECA asks:

What happened?
What can we prove?
What is disputed?
Which allegations are supported?
Where does causation break?
What defenses are developing?
What evidence changes exposure?
What still needs to be learned?

Culling a million records down to 100,000 can be excellent data work. It does not tell a lawyer whether the case should be tried, settled, attacked through motion practice, or investigated differently.

The object of ECA is not a smaller pile of information. It is a better understanding of the case.

Early Case Assessment is not Early Data Assessment

The distinction matters because the legal industry has spent years using similar language for two very different problems.

That is a useful boundary.

EDA asks questions such as:

What information do we have?
Which custodians matter?
What date ranges are relevant?
What should be collected or reviewed?

ECA asks:

What happened?
What can we prove?
What is disputed?
Which allegations are supported?
Where does causation break?
What defenses are developing?
What evidence changes exposure?
What still needs to be learned?

The object of ECA is not a smaller pile of information. It is a better understanding of the case.

Sometimes a summary is exactly the right tool

There is no reason to turn every legal task into matter-level analysis.

A lawyer reviewing one agreement may need the termination provision, indemnification language, governing law, or a concise account of the parties' obligations. A litigator reading a deposition may want a summary of testimony on a defined topic. A carrier may need a short digest of a newly received report.

For bounded tasks like those, summarization can be extremely useful.

ABA Formal Opinion 512 gives a helpful example. It considers a lawyer using a generative AI tool to review and summarize numerous lengthy contracts. The opinion says the lawyer would not necessarily need to manually reread the entire collection to verify the results if the lawyer had tested the tool on a smaller subset, compared its summaries with the documents, and established a reasonable basis for relying on its accuracy. (American Bar Association)

That example is important precisely because it is bounded.

The task is to summarize the contracts. The expected output is known. The lawyer can test whether the system captured the provisions it was supposed to capture. There is a practical validation method.

The mistake is taking that model of document summarization and assuming it scales directly into case understanding.

It does not.

Multi-document litigation fails in different ways

A matter is not simply a collection of documents waiting to be compressed. The significant facts are distributed across documents, people, time, procedural events, legal authorities, and competing accounts of the same event.

That creates failure modes that are different from producing a bad summary.

Omission. A medical record can be summarized accurately while omitting the fact that becomes important only when compared with the complaint, a later expert opinion, or testimony from another provider. The sentence may be faithfully summarized. The significance can still be missed.
Identity. Litigation files repeatedly refer to the same person, organization, event, condition, product, location, or communication in different ways. A system can accurately summarize every document and still fail to recognize that two references concern the same actor, or incorrectly merge two different ones. Once identity is wrong, chronology and causation can become wrong with it.
Contradiction. One pleading alleges an event occurred on a certain date. A contemporaneous record says something different. A witness remembers another sequence. An expert depends on one version. Later discovery changes the picture again. Individual summaries can all be accurate while the conflict between them remains invisible.

These are not primarily compression problems.

They are reconciliation problems.

A litigator usually does not need twenty excellent summaries sitting beside one another. The litigator needs to know that paragraph 42 of the complaint is inconsistent with a contemporaneous record, that the deposition testimony changed the timeline, that the expert's assumption depends on the disputed version, and that the governing authority makes that disputed fact material.

That is matter-level analysis.

Case understanding requires a model of the matter

At Turbo Law, this is why our Fact Graph is not an entity-extraction layer.

Entity extraction can identify names, dates, organizations, diagnoses, or other objects appearing in text. Useful, but incomplete.

An intelligent Fact Graph has to capture the developing understanding of the case.

That means representing not only what appears in individual documents, but how the facts, sources, people, allegations, events, legal issues, and competing versions of events relate to one another.

The graph is dynamic in two dimensions.

First, it changes with the practice area. The relationships that matter in medical malpractice are not the same relationships that matter in product liability, premises liability, toxic tort, workers' compensation, or coverage litigation. The underlying case model has to reflect the records, rules, proof structures, and strategic questions of the litigation itself.

Second, it changes with the individual matter. A case is not static on the day the file is opened. New medical records arrive. Written discovery is exchanged. Depositions create new testimony. Experts introduce opinions. Pleadings change. Research identifies controlling authority. A document produced six months into discovery can change the significance of something received on day one.

The graph must change with that record.

That is why the Turbo Law Fact Graph connects the matter rather than treating each workflow as an isolated AI interaction. Pleadings, discovery, case law, case-specific facts, records, timelines, drafts, research, and strategy can operate against the same developing understanding of the matter. Our law-firm platform is designed around one matter-level intelligence layer feeding review, drafting, research, strategy, and reporting. (Turbo Law)

The difference is fundamental.

A summary says, "Here is what this document contains."

Case understanding asks, "How does this change what we believe about the matter?"

The source of a fact matters as much as the fact

Litigators rarely work with facts in the abstract.

They work with facts and provenance.

There is a major difference between an allegation in a complaint, an entry in a contemporaneous business record, testimony from a party, an admission in written discovery, and an expert's characterization of the evidence.

They may all contain sentences about the same event. They do not carry the same meaning.

That is why matter-level analysis should preserve the path back to evidence. In Turbo Law, the platform's factual conclusions are designed to remain connected to their underlying sources, including page and line references where applicable.

This serves a practical purpose beyond citation convenience. A lawyer evaluating a conclusion should be able to move from synthesis back to evidence quickly enough to answer:

Is this alleged or established?
Who said it?
When?
What source supports it?
Is there another source that conflicts with it?
Has anything received later changed the analysis?

Without provenance, sophisticated synthesis can become another assertion that the lawyer has to independently reconstruct.

With provenance, the system can help the lawyer inspect the reasoning from the case record outward.

The right unit of analysis is the legal question

There is another reason document-by-document analysis reaches a ceiling.

Lawyers do not ultimately make document-level decisions.

They make decisions about claims, defenses, exposure, discovery, motions, experts, settlement, reporting, and trial strategy.

Consider a causation question. The relevant answer may depend on the complaint, records predating the alleged event, treatment after it, deposition testimony, expert assumptions, discovery responses, and jurisdiction-specific case law.

No single document contains the answer.

The useful unit of work is therefore not always, "Summarize these records."

It may be, "What evidence supports and undermines this causation theory, where are the contradictions, and what remains unresolved?"

That shift sounds subtle, but it changes the architecture required to do the work.

Document summarization starts with a document and produces compressed text.

Matter-level analysis starts with the question and moves through the connected record required to answer it.

This is also why drafting should not begin from a blank prompt when the matter itself already contains the relevant intelligence. In Turbo Law, Draft can work from the matter, firm priors, and applicable jurisdictional rules. Research can work across the case file and public record. Strategy & Settlement can use the same matter context for evaluation, mediation preparation, settlement analysis, and reporting.

The outputs differ. The underlying case should not.

Verification should increase with the consequence of the task

Better matter context does not eliminate lawyer review.

It makes review more useful.

ABA Formal Opinion 512 emphasizes that the appropriate degree of independent verification depends on the tool and the task. It also makes clear that lawyers cannot surrender professional judgment to a generative system and remain responsible for the legal work performed for the client. (American Bar Association)

Florida Bar Ethics Opinion 24-1 reaches the same basic operating principle from the perspective of competence, confidentiality, billing, and professional responsibility: lawyers remain responsible for their work product and judgment and should establish practices for verifying AI-assisted work. (The Florida Bar)

The National Center for State Courts has likewise recommended that the rigor applied to AI tools should be proportional to the intended use and the risk involved. Its broader guidance emphasizes human oversight and accuracy review, especially as AI is used for more consequential legal work. (National Center for State Courts)

That matches how we build Turbo Law.

Human review remains part of the process. The system is designed to organize evidence, surface connections, and keep conclusions tied back to the underlying record so attorneys can apply judgment where it matters.

A bounded document summary and a settlement recommendation should not be treated as equivalent tasks. Neither should a chronology and a dispositive argument. The consequences are different, so the depth of source review and professional judgment should be different.

The goal is not to remove the lawyer from the loop.

It is to give the lawyer a much better loop.

A stack of summaries is still a stack

Document summaries solve a real problem. We use them where they are the right abstraction.

But complex litigation does not become understandable merely because every file has been summarized.

The case exists in the relationships between those files: what supports what, what contradicts what, which person is which, how the chronology changed, which allegation lacks evidence, how later discovery changes earlier assumptions, and how the developing record interacts with the law.

That is the line between document intelligence and matter intelligence.

Early Data Assessment helps teams understand and control the information universe.

Early Case Assessment helps lawyers understand the case.

At Turbo Law, our objective is to build for the second without losing the rigor of the first. The result should not be a shorter version of the file. It should be a connected, source-grounded understanding of the matter that becomes more useful as the litigation evolves.