In a commercial dispute of ordinary size, the facts that decide the case rarely sit in one place. A breach-of-contract claim may turn on an email from a vice president of sales, a line in a revenue ledger, a chat thread among engineers, and what the chief financial officer says months later at deposition. Each of these arrives through a different channel, from a different custodian, and on a different schedule. Litigation teams are practiced at collecting and reviewing such material, but they often cannot see how the pieces bear on one another until late in discovery, when the deposition calendar is fixed and the chances to act on what they learn have narrowed.
Software that uses machine learning and large language models to connect records across sources is designed to move that moment earlier. This post argues that the strategic value of these tools lies chiefly in timing: a team that sees the relationships, contradictions, and patterns in its record during the first months of a case makes better decisions about discovery, depositions, motions, and settlement than a team that assembles the same picture just before trial. That value depends entirely on lawyers treating machine-identified connections as leads to be verified, and the professional obligations that attach to the work do not change because software produced the first draft.
Document review sorts the record but leaves the connections to the reviewers
Conventional review, whether linear or technology-assisted, answers questions about individual documents: whether an email is responsive, whether it is privileged, and whether it bears on a contested issue. Technology-assisted review has gained broad acceptance for the responsiveness question in particular [1], and it has made large productions manageable. What it does not do is tell a team how a given document relates to the ledger entry produced from a different custodian or to testimony that has not yet been given.
That relational work has traditionally been done by hand. Associates maintain chronologies in spreadsheets, reviewers leave notes in coding fields, and much of the case's structure lives in the memory of whoever has read the most. When first-level review is divided among a dozen contract attorneys, no single reader may see both the email that mentions a payment and the spreadsheet that records it. Financial records compound the difficulty, since they are often produced in native format and reviewed separately, while deposition transcripts arrive weeks or months after the documents they discuss. The underlying problem is structural: the record is organized by source and by custodian, while the case is organized by claim, defense, and witness.
Structured case intelligence organizes the record around people, events, and claims
The newer generation of litigation tools attempts to close that gap by extracting the elements a lawyer would track and linking them across the record. These include named individuals, companies, dates, dollar amounts, contract provisions, and product names. Once extracted, those elements can be placed on a common index, so that a reference to a payment in an email can be matched to the corresponding ledger line, the board minutes that approved it, and the deposition passage in which a witness describes it. The result, when it works, is a chronology in which every entry points back to a Bates-numbered source.
The qualification is substantial. Extraction is least reliable precisely where litigation records are messiest: nicknames and initials, shared mailboxes, long forwarded chains, scanned handwriting, and spreadsheets whose meaning depends on formulas and hidden columns. Tools differ considerably in how they handle these problems, and there is no widely adopted benchmark against which a firm can compare their accuracy on legal records [citation needed]. A structured view of the record is therefore best understood as an organized hypothesis about what the documents show, one that becomes useful only when someone tests it against the sources.
Inconsistencies found early shape the discovery plan and the order of depositions
An illustration shows how timing changes decisions. Consider a hypothetical matter, Meridian Analytics, Inc. v. Halvorsen Data Partners, LLC, in which Meridian alleges that Halvorsen breached a data-licensing agreement by sublicensing Meridian's datasets to a third-party customer. Halvorsen's answer denies any sublicense, and its chief financial officer submits a declaration stating that Halvorsen derived no revenue from Meridian data. In Halvorsen's first production, linked analysis surfaces three items that no single reviewer flagged together: invoices to the third party whose line items track the names of Meridian's datasets, an internal chat message in which a sales engineer refers to "the Meridian feed," and a revenue schedule that allocates income to a product code associated with that feed.
Seen together and early, those documents reshape Meridian's discovery plan. Counsel can serve targeted requests under Fed. R. Civ. P. 34 for the third-party contract and related correspondence, an interrogatory under Rule 33 asking Halvorsen to identify every product incorporating Meridian data, and a notice under Rule 30(b)(6) that designates revenue attribution as a topic. The order of depositions changes as well. Because Rule 30(a)(2)(A)(i) limits each side to 10 depositions without leave or stipulation, and Rule 30(d)(1) limits each to one day of seven hours, Meridian has reason to depose the sales engineer before the chief financial officer, so that the financial witness is examined against a developed record rather than a set of documents counsel has only half absorbed.
Specificity also helps if Halvorsen resists. Under Rule 26(b)(1), discovery must be proportional to the needs of the case, and a request tied to identified documents is easier to defend in a motion to compel than a broad demand for "all documents concerning Meridian [2]." The same discipline applies in the other direction. The earliest use of linked analysis is often on a firm's own client's records, before production, where finding an inconsistency in the client's account allows counsel to address it with the client, prepare witnesses accordingly, and shape positions at the Rule 26(f) conference with fewer surprises in reserve.
Earlier knowledge of the record changes motion practice and settlement
The benefits extend beyond discovery. A party that understands its own record early can amend its pleadings while leave is still "freely give[n] when justice so requires" under Rule 15(a)(2), rather than seeking amendment after the scheduling order's deadline has passed and the more demanding good-cause standard of Rule 16(b)(4) applies [3][4]. It can also meet its duty under Rule 26(e) to supplement disclosures and responses in a timely manner. Preservation decisions improve for the same reason, because identifying the custodians and data sources that actually matter informs the scope of a litigation hold and reduces exposure under Rule 37(e) for lost electronically stored information [5].
Motion practice benefits from the same organization of the record. Rule 56(c)(1)(A) requires a party asserting that a fact is or is not genuinely disputed to cite particular parts of materials in the record, and a chronology in which each fact is already tied to a Bates number or transcript page is close to the form that standard demands. Settlement analysis changes too. A client deciding whether to resolve a matter before incurring the cost of full discovery is better served by an assessment built on cross-referenced evidence than by one built on the documents that happened to be coded as significant in the first review pass. Early assessments have always been possible, of course, but they have usually rested on a smaller and less connected portion of the record.
Machine-identified connections remain leads until a lawyer verifies them
None of this alters the lawyer's responsibility for the result. Language models can produce links that are plausible and wrong, and courts have sanctioned attorneys who filed briefs containing citations that generative tools fabricated [6]. Discovery carries its own exposure: a signature under Rule 26(g) certifies, after a reasonable inquiry, that a disclosure is complete and correct when made and that a discovery response is consistent with the rules, and Rule 11 imposes a comparable obligation on factual contentions in pleadings and motions [2]. Comment 8 to ABA Model Rule 1.1 extends the duty of competence to the benefits and risks of relevant technology, and ABA Formal Opinion 512 addresses the use of generative AI tools under the Model Rules [7]. Confidentiality under ABA Model Rule 1.6 and the duty to supervise nonlawyer assistance under ABA Model Rule 5.3 also bear on which vendors a firm uses and on what terms [7].
Firms adopting these tools therefore need a verification practice that is written down and followed. The certifications under Rules 26(g) and 11 attach to the lawyer who signs, whatever tool produced the draft, and a reviewer who relies on a tool's summary of a document may not have made the reasonable inquiry those rules contemplate. A written practice also gives a supervising lawyer a way to confirm that the checks were actually performed, which bears on the supervisory obligations described above. At a minimum, it should include the following steps.
Every fact drawn from a machine-generated chronology is checked against the underlying document or transcript before it appears in a filing, discovery response, or deposition outline.
Each claimed link between two records is confirmed by a reviewer who reads both sources rather than the tool's summary of them.
The team records which tool, version, and data set produced a given analysis, so the work can be reconstructed if it is later questioned.
The engagement terms with the vendor are reviewed for data retention, model training on client material, and access by vendor personnel before any client data is loaded.
These steps cost time, and some of the efficiency the tools offer is consumed by them. That trade is the correct one. The strategic advantage described here comes from knowing the record sooner, and a mistaken connection acted upon early can do more harm than a correct one found late.
Work product protection for AI-generated analysis is not yet settled
One question deserves attention from any firm building these tools into its practice. A chronology or relationship map prepared by counsel in anticipation of litigation would ordinarily qualify as work product under Rule 26(b)(3), and much of it would reflect opinion work product entitled to the strongest protection [2]. Whether the same is true of the prompts a lawyer enters, the raw outputs a tool generates, and the configuration choices that shape its analysis has not been resolved consistently across jurisdictions [8]. Requests for information about how an opponent used such tools, particularly where the tool also informed search terms or responsiveness decisions, are likely to test those boundaries.
The practical course is to pilot these tools on a closed matter where the decisive evidence is already known, and to measure whether the team would have found it earlier and with what rate of error. That exercise yields a defensible workflow and an honest sense of where the tools help. It also gives the firm a record of its process, which may matter if the scope of work product protection is litigated in one of its own cases before the courts have settled it.
References
Rio Tinto PLC v. Vale S.A., 306 F.R.D. 125, 127 (S.D.N.Y. 2015). The fetched page does not show the reporter volume or the decision date. The page numbers it shows (*127 to *134) match 306 F.R.D. 125, and secondary sources date the opinion March 2015. Check both before publishing. https://www.courtlistener.com/opinion/8788861/rio-tinto-plc-v-vale-sa/
Fed. R. Civ. P. 26(b)(3)(A)–(B). https://www.law.cornell.edu/rules/frcp/rule_26
Fed. R. Civ. P. 15(a)(2). https://www.law.cornell.edu/rules/frcp/rule_15
Fed. R. Civ. P. 16(b)(3)(A), 16(b)(4). https://www.law.cornell.edu/rules/frcp/rule_16#rule_16_b
Fed. R. Civ. P. 37(e). https://www.law.cornell.edu/rules/frcp/rule_37#rule_37_a_5
Mata v. Avianca, Inc., No. 22-cv-1461 (PKC) (S.D.N.Y. June 22, 2023) (opinion and order on sanctions). A secondary source gives the reporter cite as 678 F. Supp. 3d 443. Verify it before use. https://caselaw.findlaw.com/court/us-dis-crt-sd-new-yor/2335142.html
ABA Comm. on Ethics & Prof'l Responsibility, Formal Op. 512, at 2, 14 (2024) (quoting Model Rules of Prof'l Conduct r. 1.1 cmt. 8). https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf
Melissa Weberman & L. Michel Marchand, The Emerging Framework on AI Prompts, Privilege, and Discovery, Arnold & Porter eData Edge (May 19, 2026). The post discusses four decisions, which are better to cite directly once you have pulled and checked them; I could not fetch them. As the post cites them: Tremblay v. OpenAI, Inc., No. 23-cv-03223, 2024 WL 3748003 (N.D. Cal. Aug. 8, 2024); Concord… https://www.arnoldporter.com/en/perspectives/blogs/edata-edge/2026/05/emerging-framework-on-ai-prompts-privilege-and-discovery
