What a Discovery Framework Must Do

Bottom line: most AI discovery tools will sell you a faster way to read documents. Reading documents is the part that's already getting cheap. The part that decides cases, and the part worth paying for, is what happens next: which contradiction to press, which to sit on, what's missing from the production, and which of it touches an element somebody has to prove. If a vendor can't show you that on a realistic matter, you're buying a faster way to build a second pile.

This guide is for whoever signs the purchase order, and for the outside counsel who'll be asked to use whatever gets bought. It covers why the money is moving, what discovery is actually for, the eight capabilities worth paying for, where the lawyer stays in charge, and what to demand before anyone signs.

Why the money is moving

Start with the numbers your budget already knows. Document review is 50 to 90 percent of litigation cost. Human reviewers get through roughly 45 documents an hour, and on borderline documents they agree with each other only about 70 percent of the time.

Now the other side of the ledger. In a blind run written up in We Ran a Blind eDiscovery Review with No Humans in the Loop, Legawrite.AI's open-source isResponsive engine classified 8,665 documents for about $5 of compute. At the usual $0.50 to $1.00 per document for human first-pass review, the same set runs roughly $4,300 to $8,700. The machine's recall was 0.807, with a lower confidence bound of 0.789, which clears the 75 percent defensibility floor.

The write-up is just as useful for what it admits:

• The elusion check failed. The pile the machine rejected was 5.7 percent responsive, against a 5 percent target.
• The misses clustered where you'd expect. Chats, meeting minutes and executive presentations, where responsiveness is implied rather than stated. Recall on clear documents was 0.933; on borderline ones, 0.452.
• The calibration check failed too. So automatic acceptance was locked, and all 2,038 responsive calls went to human review.

The honest posture today is machine classification with mandatory human review of what the machine flags, not fully automated production. The method is laid out in the isResponsive paper.

Two lessons for a buyer. First, first-pass review is getting cheap fast; don't pay a premium for it. Second, notice what a credible vendor sounds like: a locked rubric, thresholds fixed before the run, failures disclosed, and a defensibility certificate a court can examine. The write-up puts it this way: "Openness is not marketing. It is the mechanism of trust." Hold everyone to that standard, including the people selling you the strategy layer.

What discovery is actually for

Here's the frame that should drive every product decision. Discovery exists to turn a fight between competing stories into a record with four properties.

• Constrained. The other side's room to change its story shrinks with every sworn answer. Interrogatory answers under Rule 33 are sworn. Matters admitted under Rule 36 are conclusively established unless the court allows withdrawal. Documents produced under Rule 34 become a fixed record.
• Authenticated. The evidence is usable, not just interesting. The tool has to track who created a document, who received it, and who can authenticate it.
• Witness-linked. Documents don't testify; people do. Every material finding should map to a custodian, author, recipient or decision-maker who'll be asked about it.
• Oath-committed. Once a party answers an interrogatory, the question stops being "what happened?" and becomes "what have they sworn happened?" That shift drives everything downstream.

The buyer's test is a single question: can the tool tell you whether a fact can actually be used as evidence, and who's on the hook for it? If not, it's producing research, not litigation material.

The eight capabilities worth paying for

1. Contradiction triage, not contradiction lists

Any capable system can spot that an interrogatory answer says Tuesday, the minutes say Wednesday and the email says Thursday. That's table stakes. The question that matters is what to do about it, and there are three answers.

The hold bucket is where value compounds. A clean contradiction preserved until the declarations at summary judgment can become fatal under the sham affidavit doctrine, which lets a court disregard a declaration that contradicts the party's own earlier sworn testimony without adequate explanation. A flat list of contradictions is one more pile. A triaged list with actions is an asset.

2. Supplementation traps

When a clear falsehood shows up in the first set of interrogatory answers, the instinct is to ask them to explain it. That's usually a mistake. It gives opposing counsel time to prepare the witness and polish an answer, and you trade a strong deposition moment for a cleaned-up written response.

The better move is an adjacent question that looks routine and forces the other side either to repeat the falsehood from a new angle or to contradict its own earlier sworn answer. Either outcome helps you. The quality test is simple: would a reasonable observer think you'd ask this even if you hadn't found the lie? If yes, it's a good trap. If it only makes sense in light of the contradiction, it's a premature reveal. A tool that can draft these, and tell them apart from direct confrontation questions, is doing litigation work, not document review.

3. Batch awareness

Discovery doesn't arrive in one delivery. It moves through initial disclosures, a first round of interrogatories and document requests, responses and productions, supplemental discovery, and then depositions and dispositive motions. The same contradiction is worth a new interrogatory early and is deposition or summary judgment material only once written discovery has closed.

So the tool has to know which batch it's looking at (Batch 0 from initial disclosures, Batch 1 from the first production, and so on), remember what was asked and answered, and know which moves remain. It should also watch how the other side's positions evolve. If they supplement after Batch 1 and the fix creates a new inconsistency, they tried to repair a problem and made another one. That's a different kind of vulnerability, and the tool should say so.

4. Negative space

Gap detection asks what you need that you don't have. Negative space analysis asks something sharper: given what was produced, what else should exist that never showed up? Some examples:

• The board met monthly, and minutes exist for eleven of twelve months.
• Five of six executives on an email chain respond; one is conspicuously silent.
• A policy manual is referenced in three separate documents and never produced.
• Communications shift from email to text, or from corporate to personal accounts, at a specific point in the timeline.

The tool should model what a complete record would look like, flag the deviations after every batch, and attach two things to each: a recommended next step, and a call on whether that step requires a fight. Those are different postures. One is an innocent production gap a meet-and-confer letter will fix. Another is a privilege assertion that may or may not hold up. A third is a pattern that suggests documents were withheld or destroyed, where Rule 37(e) relief is worth considering. And that silent executive? Now a priority deposition.

5. Objection patterns as data

Most review tools read what was produced and ignore how the other side fought. That's a mistake. A cluster of privilege assertions over a short period often marks the stretch when something went wrong and the client called its lawyers. A relevance objection on a routine request may mean you hit a nerve. A blanket burden objection on one custodian may mean that custodian's files are large, disorganized or damaging. A redaction pattern on one topic tells you what they think is worth fighting over.

What to require is a skeptic function that tests every objection against the applicable standard. Is the relevance objection supported by the scope of the claims and defenses? Is the claimed burden actually quantified? Does the privilege log satisfy Rule 26(b)(5)? Does "not in our possession, custody or control" square with what the documents show about their access? The output should be two lists, objections to accept and objections to challenge, with draft meet-and-confer letters for the second.

The tool should also cross-reference objections against negative space. Documents missing from a heavily privileged period may be legitimately withheld. Documents missing from a period with no privilege claims are harder to explain. Keep the design conservative here: surface the patterns, flag the correlations, and let the lawyer interpret them. Opposing counsel's style and the judge's tendencies are human calls.

6. A live element map

A fact matters only if it connects to an element somebody has to prove. Negligence needs duty, breach, causation and damages. Fraud adds knowledge, materiality, reliance and intent. Breach of contract needs formation, terms, breach and damages.

The tool should keep a proof chart that updates after every batch: for each element of each claim and defense, what supports it, what undermines it, what sworn positions have been taken on it, what's unresolved, and what discovery would strengthen it. When the tool finds a contradiction, its first question should be which element that contradiction touches. One that touches no element is noise. One that undercuts an element of the other side's affirmative defense is a summary judgment exhibit.

7. Credibility, kept in its lane

Credibility is for the jury, and at summary judgment the court views the facts in the light most favorable to the non-moving party. A pre-trial tool doesn't need a trial-grade credibility profile. It needs two narrower things. The first is a tracker of every sworn position taken by every key witness, so a contradictory declaration at summary judgment can be met with the prior statements immediately. The second is a read on evasion (precise where it helps them, vague where it hurts) to aim the next round of interrogatories and admissions. Both are inputs to building the record, not arguments in the brief.

8. A full arc, not a feature

Put together, the tool should cover all of written discovery.

Note one deliberate stopping point. If a meet-and-confer fails and a motion to compel is needed, the tool flags it for the lawyer and does not draft it. That's motion practice, and the call depends on the court, the judge and the overall posture of the case.

Where the lawyer stays in charge

Five decisions belong to the attorney, and a good tool puts them in front of her instead of burying them in automated recommendations.

1. Triage overrides. "Hold" may need to become "press" because the client can't afford a longer discovery period. "Press" may need to become "hold" because pressing would reveal a line of investigation you want kept quiet.
2. Fight or accept. Whether to fight a discovery dispute turns on cost, delay, the judge, the strength of the underlying request and the risk of reciprocal scrutiny.
3. Trap review. Only someone who knows opposing counsel can judge whether a question really looks routine.
4. Case theory. Discovery sometimes shows the theory is wrong or that a stronger one exists. Dropping a claim, adding one or changing the damages theory is a strategic call.
5. Client communication. What the evidence means for case value, risk and options is counseling, not analysis.

The design principle is augmentation, not automation. The interface between tool and lawyer should be a decision queue: findings, recommendations and required approvals, with deadline-driven items separated from strategic ones, decisions tracked, and the most important items at the top.

What the screen should look like

Organize the output by what the attorney needs to decide, not by what the tool did. The primary view is the element map. Beneath it sit the contradiction triage queue, the negative space findings with recommended actions, the objection analysis with draft letters, the sworn-position tracker by witness, and the supplemental discovery drafts ready for review.

Every recommendation carries its reasoning. Not "follow up on this contradiction," but something like this: the contradiction affects Element 3 of the breach of contract claim; the current interrogatory answer is vague enough to survive summary judgment; a tighter request for admission would force a cleaner commitment; here's a draft. The lawyer reads it, agrees or doesn't, and moves on. And no critical finding should ever be buried in a long report.

Here's what I want from you

To the vendors pitching discovery tools, and to the outside counsel choosing one on a client's budget:

1. Show me the element map first, on a realistic matter, not a demo set you picked.
2. Show me triage: three buckets, a recommended action for every contradiction, and the reasoning in plain English.
3. Tell me which batch we're in and how that changed the recommendation.
4. Show me what's missing, with a "does this need a fight?" call on each item.
5. Run the skeptic on their objections and hand me draft meet-and-confer letters, not a list.
6. Publish your numbers: recall with confidence bounds, what failed, and what you routed to humans. If an open-source project can disclose a failed elusion check, so can you.
7. Put the judgment calls in a queue, and log what the lawyer decided and why.
8. Stop at the motion. Flag the motion to compel; don't draft it.

Do that and you're selling a strategy engine. Otherwise you're selling a faster pile, and piles are getting cheap.