The Two Clocks, in Litigation

By Harry Raworth, Co-Founder of Mary Technology. Originally published on LinkedIn, 24 July 2026.

The bottleneck in legal AI has moved from what the technology can do to whether firms can absorb it. Nowhere is that gap more expensive than litigation.

I recently read Zack Shapiro’s excellent essay, The Two Clocks. His argument is that AI capability moves on one clock and institutions on another. The technology is improving faster than organisations can change the way they work.

Shapiro writes mainly about transactional law: contracts, diligence, firm economics and the work of turning legal judgment into repeatable workflows. Litigation sharpens the argument, and it changes one of his best metaphors.

Litigation has a compiler

Shapiro writes that “law has no compiler.” A flawed contract does not crash. It can sit in a drawer, apparently fine, until a counterparty invokes a right nobody examined closely enough. For deals, that is exactly right.

Litigation is different. The flawed work doesn't sit in a drawer, it gets hunted.

There is a compiler in disputes, and it is opposing counsel. Someone is paid to test what you file: the weak proposition, the unsupported inference, the date that does not fit, the authority that says less than the brief claims, or the witness whose account cannot survive the documents. The error report may arrive as a cross-examination, a footnote in a reply brief, or a question from the bench. It runs on someone else’s schedule.

The Sullivan & Cromwell incident in April 2026 makes the timing problem clear. In a letter to the U.S. Bankruptcy Court for the Southern District of New York, the firm said an emergency motion contained inaccurate citations and other errors, including AI “hallucinations.” It said its AI policies hadn't been followed and that its review process had failed to identify the errors. Boies Schiller Flexner brought the matter to the firm’s attention. The compiler ran after filing, in public, on the other side’s clock.

That incident matters because of timing, not just hallucination. Verification happened after filing. A rule that says “check the output” is weak if checking remains an optional final step under time pressure. It needs to be part of the work itself.

Every litigator knows a version of the week-four moment. A document appears that won't sit beside one already relied upon. A date in one witness statement makes the chronology in another impossible. An email changes the apparent meaning of a meeting note. None of it was necessarily hidden. It may have been sitting in the production from the beginning, read in week one by someone who had no way of knowing that one line on page three of one document quietly broke a proposition built from another.

The contradiction existed on day one. The manual process reached it in week four.

In the end, a litigation process answers one question: do you find the fact while you can still change the strategy, or does the other side find it for you?

The record that never existed

The usual prescription for AI absorption is to write the firm’s method down clearly enough that a model can follow it and a lawyer can supervise it. That works more naturally in deal practice because transactional teams have accumulated reusable material for decades: precedent banks, clause libraries, negotiation playbooks, closing sets and model documents. They do not capture the whole method, but they leave a usable trail.

Litigation leaves a much thinner trail. A team may receive tens of thousands or millions of documents, build chronologies, witness folders, issue lists, review protocols and written submissions, yet still preserve only part of the chain connecting source to fact, fact to issue, issue to theory and theory to judgment.

Litigation hasn't been slow to use machine assistance. Technology-assisted review in e-discovery is a well-known use of AI in legal practice. But a responsiveness code is not a theory of the case. A document summary doesn't preserve the reasoning that made the document important. A final brief records the conclusion, not necessarily the reading that produced it.

The missing asset is the reading itself: which sentence changed the theory, which contradiction mattered, why an absence was suspicious, which source was too weak to carry a proposition, and how a senior lawyer’s correction changed a junior lawyer’s initial call.

Today, that judgment is scattered across comments, emails, calls, tracked changes, temporary chronologies and people’s heads. When the matter ends, the link between the record and the reasoning often disappears. There is no complete method waiting to be uploaded. It has to be captured at the point of review, while the source, the provisional conclusion and the lawyer’s correction are still connected. That is harder than uploading a deal playbook, but more valuable. Whoever captures it is building a record the profession has never had.

The driveshaft, disputes edition

The best metaphor in Shapiro’s essay is a century old. Early factories often replaced a central steam engine with an electric motor while keeping the old system of shafts and belts. The larger productivity gains came later, when individual motors powered individual machines and factory layouts were redesigned around the work.

Litigation is living that story now. The driveshaft version of legal AI is easy to recognise: keep the same document-by-document review process and attach a model so the march goes faster. Add summaries inside the review platform. Automate a first-pass relevance call. Generate a chronology after the review is substantially complete. Those tools can help, but they don't change the layout that produced the week-four problem.

Tearing up the floor means changing what week one produces.

The old layout produces a coded collection and, eventually, a human-built timeline. The new one produces a working fact record. Every material proposition remains tied to the page it came from. Dates, people, entities and events are connected across documents. Contradictions and corroboration surface as the record develops. Uncertainty, assumptions and single-source dependencies remain visible. A lawyer’s correction isn't simply accepted and forgotten; it improves the next pass. The work starts to organise itself around facts rather than files, without losing the link to the source. The aim is to surface uncertainty earlier and keep it traceable, so your side gets to run the adversarial read before the other side does.

Control what compounds

Confidentiality doesn't mean every firm must build its own models or own its own servers. The exact rules depend on jurisdiction. In the United States, ABA Formal Opinion 512 treats the risks of generative AI as fact-specific. Lawyers must consider the client, the matter, the task and the particular tool; understand how client information may be handled; use reasonable safeguards; supervise providers; and, in some circumstances involving self-learning tools, obtain informed client consent before entering client information. Vendors can supply infrastructure, and firms can use external systems where the configuration, contract, safeguards and circumstances permit it.

But matter data and legal judgment cannot be pooled casually or treated as generic training material. The firm needs meaningful control over access, retention, reuse and disclosure. The durable advantage lies in controlling the part that compounds: the source-linked fact record, the review decisions, the corrections, the procedures and the accumulated judgment that make the next matter better. That can sit on vendor infrastructure, but the firm still needs control over how matter data, lawyer corrections and accumulated judgment are retained and reused.

The slow clock runs at different speeds

Shapiro describes the institutional clock of BigLaw: committees, procurement, compensation cycles and the incentive problem created when a business that bills time adopts technology designed to save it. In litigation, that clock does not run at one speed.

At hourly firms, the incentives are mixed. A saved hour may improve the client outcome, capacity and competitiveness while unsettling assumptions about leverage, utilisation and revenue. The technology and the business model can pull in different directions.

Personal-injury plaintiff work often operates differently. Contingency fees are particularly common in personal-injury cases, so compensation depends on the outcome rather than the number of hours billed. Time saved can increase capacity and margin without directly removing a billable unit.

Private capital has noticed. In 2025, EvenUp announced a $150 million Series E at a valuation above $2 billion, while Eve announced a $103 million Series B at a valuation above $1 billion. Both companies build AI systems for plaintiff-side practices.

Those valuations don't prove the products work or that either company will win its market. They do show where investors think adoption may move fastest.

Review’s cost structure helps explain why. RAND’s 2012 study of 57 large-volume e-discovery productions found that review accounted for about 73 per cent of document-production costs in the matters studied. RAND cautioned that its results could not be generalised to all litigants. The figure shouldn't be read as a current estimate of total litigation cost, but it does show how dominant human review was in the productions RAND studied.

As the cost of producing a first-pass fact record falls, more of the value moves to verification, interpretation, strategy and judgment.

What is left is the read

What survives is judgment. In litigation, that means deciding which three facts carry the case and which contradiction the other side will find first. It means knowing when a document supports a proposition only because the context has been stripped away, and whether an inconsistency changes nothing or breaks the theory. It also means deciding what to allege, what to concede, what to investigate next and what a judge is likely to regard as fair.

Then there is the record that should exist but does not. A system may flag an expected document that appears to be missing. A lawyer still has to decide whether the absence matters, what inference can responsibly be drawn from it, and what step should follow.

The machine can help produce and interrogate the record. The lawyer still has to decide what it means and stake a reputation on that conclusion.

The training problem is harder. In disputes, much of the grind was also the apprenticeship. Junior lawyers learned judgment by reading the record, making imperfect calls and being corrected by more experienced lawyers. A firm that gives all first-hand contact with the evidence to a machine risks producing lawyers who can operate the system but cannot recognise when it is wrong.

The solution is deliberate training, not keeping inefficient review alive for its own sake. Give junior lawyers a purposeful share of the reading. Ask them to compare their conclusions with the machine’s. Make senior corrections explicit and keep them connected to the source and the reason for the change. That is also how a firm captures the method litigation never fully wrote down.

The apprenticeship and the asset are the same loop.

The gap is widest here

Two clocks, and in litigation the distance between them is unusually visible. The technology can now do meaningful fact work; many firms still organise that work as if it cannot. Meanwhile, the compiler on the other side of every case keeps running.

Absorption in disputes has a specific meaning. The check moves closer to the moment each fact is read, and the judgment the profession never wrote down is captured while the work happens. Firms that do this first may not look very different from the outside for a year or two. Then they will be difficult to catch. Every managing partner running disputes should ask one question:

Your process is running right now, on live matters. What is it actually checking, and when?

Book a call with our team→

Notes and sources

• Zack Shapiro, “The Two Clocks,” published on X, 8 July 2026. Shapiro sets out the two-clocks thesis, argues that the bottleneck has moved from intelligence to absorption, introduces the “law has no compiler” observation, and develops the factory-electrification analogy.
• Sullivan & Cromwell LLP, letter to Chief Judge Martin Glenn in In re Prince Global Holdings Limited, 18 April 2026. The firm said that the motion contained inaccurate citations and other errors, including AI hallucinations; that its AI policies had not been followed; that its review did not identify the problems; and that Boies Schiller Flexner brought the matter to its attention.
• Bloomberg Law, “Sullivan & Cromwell Apologizes to Judge for AI Hallucinations,” April 2026.
• American Bar Association, Formal Opinion 512, “Generative Artificial Intelligence Tools,” 29 July 2024. The opinion discusses technology-assisted review as a recognised use of AI in legal practice and provides the confidentiality and supervision framework described above.
• Paul A. David, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox,” American Economic Review, 1990. The paper provides the historical basis for the factory-electrification analogy.
• Cornell Legal Information Institute, “Contingency fee.” Cornell notes that contingency fees are particularly common in personal-injury matters and explains that compensation is tied to recovery rather than hours billed.
• EvenUp, announcement of its $150 million Series E at a valuation above $2 billion, October 2025.
• Eve, announcement of its $103 million Series B at a valuation above $1 billion, September 2025.
• RAND Corporation, Where the Money Goes: Understanding Litigant Expenditures for Producing Electronic Discovery, 2012. RAND studied 57 large-volume productions and found that review typically accounted for approximately 73 per cent of production costs, while cautioning against generalising the findings to all litigants.