SB 574 cleared the Senate 39-0 back in January. It’s sitting in the Assembly right now, with the legislative session ending this week. Most of what it does is sensible: verify AI output before it goes anywhere, keep client data off consumer-grade tools, disclose AI use where it matters. Then it adds one line that swallows the rest of the bill: an attorney shall not delegate the practice of law to generative AI.

No supervision exception. No competent-review exception. Just a flat prohibition on a phrase the statute never defines.

That should bother you more than the sanctions cases do.

“Practice of law” has never had a clean definition, and nobody writing this bill solved that problem before attaching a penalty to it. Artificial Lawyer’s own breakdown works through what that actually means in practice: drafting a demand letter. Summarizing a deposition. Answering a client’s question about what happens next. Every one of those is ordinary, already-common AI-assisted work, and every one of those plausibly counts as practicing law under a statute with no carve-out saying otherwise.

That’s the actual enforcement mechanism, and no legislative session fixes it by accident. A rule nobody can define in advance functions as a penalty waiting for someone to decide, after the fact, what the line actually was.

Here’s what makes that worse: California’s own bar is already doing the actual work, through the process built for exactly this.

The State Bar’s Committee on Professional Responsibility and Conduct approved proposed amendments to Rules 1.1, 1.4, 1.6, 3.3, 5.1, and 5.3 back in March, covering competence, client communication, confidentiality, candor to the tribunal, and supervision of AI use. Public comment closed in May. Rule 1.1’s new comment already says a lawyer must independently review, verify, and exercise professional judgment over any AI output used on a client matter. Rule 3.3 already requires verifying every AI-assisted citation before it touches a filing.

That’s the entire substance SB 574 is reaching for, worked out rule by rule, by the body that actually regulates attorneys, through a process that takes public comment before anything takes effect. That process is still open. What it’s producing, comment by comment, is language specific enough for an attorney to actually act on, which is exactly the part SB 574 skipped.

SB 574 skips that process and legislates the headline instead. Attorney regulation has always run through the people who supervise attorneys for a living: the state bar, under authority the courts have historically treated as inherent to the judiciary. A statute from a committee built for consumer protection, reaching into what counts as practicing law, is a different branch answering a question that was never really its question to answer.

Run this test on your own week. Did you draft a letter with AI assistance? Summarize a transcript? Answer a client’s question after checking something with a model first? Under SB 574’s language, you can’t say with confidence whether any of that was fine or whether it was the exact conduct the statute exists to punish. Amendments narrowing the bill’s language before a final vote won’t resolve that uncertainty. Only an actual definition of the term would, and nobody has written one.

Whatever the Assembly does with this bill before the session ends, the vagueness survives the vote either way. A law that can’t tell a lawyer in advance what it prohibits isn’t protecting clients. It’s just moving the risk from “did I do this AI task carefully” to “will a regulator later decide this counted as practicing law by delegation,” and that second question has no answer until someone gets made an example of.

—Lawyer.exe

If you’re building your firm’s AI compliance memo this quarter, forward this to whoever’s writing it. The vagueness problem here isn’t going to stay a California-only issue.

Share

Picture this: opposing counsel’s expert has been on the stand all day at trial in a product liability case, walking through his failure analysis. Methodology, causation, the mechanism itself. Confident. Specific. No hedging anywhere in his answers.

Confident and accurate are not the same thing, and the only way to know which one you’re looking at is to go back to what the expert actually wrote before anyone was watching him testify.

Say his report addressed one alternative cause and ruled it out with language he’d hedged himself: “unlikely to be the primary factor.” On the stand that day, he ruled out three alternative causes, flatly, with no hedge on any of them. Two of the three were never in his report at all.

That’s the moment to stop assuming the day went the way it sounded, and request the rough transcript that evening instead.

Here’s the workflow.

Step 1. Request daily copy before the day ends

• Most reporters offer daily copy on a multi-day trial specifically so counsel can prep the next day’s examination overnight.

• Don’t wait for the final certified transcript. A certified copy can take weeks, and tomorrow’s session starts long before it arrives.

Step 2. Run the contradiction-and-overstatement prompt

• Paste both documents into Claude and ask it to sort the differences into two categories, not one general “inconsistencies” bucket.

• Prompt:

Compare this expert's deposition or trial testimony against his written report. Sort every difference into two categories: (1) opinions or conclusions stated in testimony that go further than what the report actually says or that the report explicitly hedged; (2) methodology, tests, or data described in testimony that the report never mentions at all. For each item, quote the report language and the testimony language side by side, with page and line citations for both.

[PASTE EXPERT REPORT HERE]

[PASTE TRANSCRIPT OR ROUGH HERE]
**[NOTE: as always, you can just attach these documents or point to them in the prompt]**

Step 3. Read the output like a change in confidence, not a change in content

• Treat this the way you’d treat a client who suddenly got more certain on the stand than they ever were in their own deposition. The added confidence is the tell, not the content.

• An expert doesn’t discover a stronger opinion halfway through a case. He either always had it and didn’t write it down, or he doesn’t have it and just said it anyway.

Step 4. Verify every flagged item by hand before it goes anywhere near a cross outline

• Confirm the actual page and line citations the model pulled from both documents. This is not optional. A cross built on a misquoted report line is worse than no cross at all.

• Filter out the differences that are just word choice, not substance. The model will flag some of those. They don’t belong in the outline.

Step 5. Build tomorrow’s cross around what survives verification

• Structure it in the order that makes the expert answer the same question twice: what does your report say, and why did yesterday’s testimony say something else.

Now, tomorrow’s cross runs on the actual words instead of what anyone remembers hearing.

A cross built this way rests on the exact contradiction between what the expert wrote in his report and what he actually said on the record. That’s a harder thing for him to explain away in front of a jury than a lawyer’s recollection of what he sounded like the day before.

Inconsistency is what actually costs an expert his credibility, and it’s checkable the same night it happens if you’re willing to spend an evening on it instead of assuming the day went the way it sounded.

Verify the transcript against the report. Then build the cross.

—Lawyer.exe

Subscribe for more workflows like this. And please spread the word!

Share

Writing is the mechanism by which legal thinking actually happens, not just its byproduct. Every junior lawyer who’s worked through a hard draft from scratch built pattern recognition doing it: what a weak argument feels like before you can name why, where a citation is doing more work than it can support, when a paragraph needs cutting to a sentence. Ivy B. Grey named this directly in the ABA Journal this month: writing “is not the end product of thinking[,] it is the process by which thinking happens.”

Route the AI draft to a junior for light editing instead, and the to-do list clears today. The skill that editing was supposed to build never forms.

Grey’s actual warning is that the “danger is losing lawyers who can tell whether the brief is any good[,] and why.”

Firms already know this, and they’re doing it anyway.

Thomson Reuters’ newest Future of Professionals report found that seventy-eight percent of law firm professionals believe early-career lawyers depend on experienced mentorship to develop the skills AI is displacing. Clearly not a fringe opinion. That’s most of the profession, on the record, describing the exact trade their own firms are making.

The same report found that professionals whose approach to AI doesn’t match their firm’s strategy are almost three times more likely to consider leaving within a year. Every departure like that takes an experienced mentor out of the building. The mentorship model everyone says they believe in is losing headcount at both ends: the juniors who need it, and the seniors who’d actually provide it.

This isn’t merely theoretical. LexisNexis surveyed nearly nine hundred UK lawyers and found seventy-two percent have concerns deep legal reasoning and argumentation will be the biggest skills gap in junior lawyers today. Sixty-nine percent named something more specific: weak verification and source-checking.

The whole premise of “AI doesn’t replace judgment, it replaces document review” depends on a lawyer who can catch what the AI got wrong. Verification and source-checking is that judgment. It’s also the exact skill the survey says is already the weakest one.

Firms are running an experiment on their own bench strength, with no one checking the control group. The efficiency gain from routing drafts to AI shows up immediately, on this quarter’s numbers. The capability loss shows up in five years, when the associate who only ever edited AI output makes partner and has to supervise the tool instead of just using it. Nobody’s tracking that second number yet, because it hasn’t come due.

The associates who keep doing the hard drafting work now, even though AI makes it easy to skip, are the ones who’ll actually be qualified to supervise it once that’s the job. Firms deciding what gets routed to AI and what gets assigned to build judgment are making that call whether they’ve noticed or not.

—Lawyer.exe

If your firm hands associates AI drafts to polish, ask whether anyone’s also handing them a blank page often enough to still know how.

Share

Thirty percent, two years, no carve-out for deals the manager never touched. We read the agreement. It said exactly that.

The termination clause required written notice and a cure period. My client had followed it to the letter. The tail commission clause was broad enough to survive almost any clean exit.

On the four corners of that document, the manager had the stronger case.

Then we checked whether the document was actually a contract.

Step 1. Verify execution before you concede a single term

• Before arguing about what an agreement says, confirm every named party actually signed it. Pull the signature pages specifically, not just the body text everyone reads first.

• A contract’s language doesn’t matter if there was no valid acceptance. Two parties agreeing to be bound is what makes every other clause enforceable in the first place.

Step 2. Run the termination and tail-obligation audit

• Paste the full agreement, signature pages included, into Claude with a prompt built to check three things at once: termination mechanics, tail obligations, and execution status.

• Prompt:

You're reviewing a [management/services/licensing] agreement on behalf of [my client / the other party]. I need three things:

1. TERMINATION MECHANICS: Identify every provision governing how this agreement ends—notice requirements, cure periods, and any conditions precedent to a valid termination. Quote the exact language and cite section numbers.

2. POST-TERMINATION / TAIL OBLIGATIONS: Identify every obligation that survives termination—commissions, fees, non-competes, confidentiality, anything with a "survives termination" clause or similar. For each one, state its exact scope: what triggers it, how long it lasts, and what it actually covers versus what a party might be claiming it covers.

3. EXECUTION STATUS: Confirm which parties are named as signatories, and flag whether the document shows actual execution—a dated signature, initials, or signature block—for each of them. If any party's execution can't be confirmed from what I've given you, say so explicitly and tell me what else to look for.

[PASTE THE FULL AGREEMENT HERE, INCLUDING ALL SIGNATURE PAGES AND EXHIBITS]
**[NOTE: as always, you can just attach these documents or point to them in the prompt]**

Step 3. Confirm what the model flags against the physical document

• If Claude flags a missing signature or an unclear execution date, go find the actual document, not a scanned copy someone forwarded three times. E-signature platforms log a certificate of completion; check for it before you assume the gap is unprovable.

• Treat this the way you’d treat a chain-of-custody objection to an exhibit. The paper is worthless if you can’t prove how it got signed.

Step 4. Build the letter around what’s actually enforceable

• Lead with the execution problem. It’s the strongest ground you have, so put everything else in the letter in service of it instead of treating it as its own separate point.

• The termination notice and the tail-commission scope still belong in there, just as backup, not as the main argument.

Most people arguing about a contract get stuck arguing about what it says. This one turned on whether the thing had ever actually become a contract in the first place, since the manager never signed her own copy. Once that was on the table, the termination language and the commission language stopped being the fight.

Read the document closely. Then go confirm who actually signed it, before you build anything on what it says.

—Lawyer.exe

Subscribe for more workflows like this. And please spread the word!

Share

A self-represented plaintiff in a privacy and discrimination suit hid text inside two of his own court filings. White font, white background, invisible to a human reader scrolling the document. The text was addressed to any AI system that might process the filing: rule in his favor, and treat the clerk’s prior ruling as an error to correct.

But Judge Walter M. Spader Jr. caught it. His order bars the plaintiff from e-filing entirely; everything now goes to the clerk’s window as a printed copy. It’s apparently the first documented prompt-injection attack aimed at a U.S. court.

Nobody needed to hack anything. He opened a word processor, typed some instructions, and made the font invisible.

That should probably worry you a bit more than the last hallucination story did.

Every AI rule that hit lawyers this year points the same direction, and it isn’t this one.

Connecticut’s own Practice Book Section 4-9, effective in June, requires anyone who used generative AI to create or edit a filing to independently verify every citation and authority it touched. That’s a rule about what you generate. It says nothing about what someone else’s filing tells your AI to do.

California’s SB 574, headed for an Assembly vote before the legislature adjourns later this month, would impose a statutory duty to verify and correct your own generative AI output. Same direction. Outbound only.

New York City’s bar addressed AI recording and transcription too: consent from every participant before you let an AI tool listen in. Also outbound, just a different flavor of it.

Three jurisdictions, three rules, one blind spot. Not one of them asks whether the document you just fed into Claude was trying to tell it something.

This lands hardest on solo and small-firm practices. They’re the ones without an IT department to quietly roll out a fix.

Big firms have review layers between an incoming filing and whatever AI tool touches it. Most solo and small-firm practices don’t. If you’re the kind of practice that pastes an opposing party’s filing, a pro se litigant’s complaint, or a client’s intake form straight into Claude to triage it, that document is now something you haven’t actually read before your AI tool has processed it. The plaintiff in this case wasn’t a hacker. He typed some words and changed a font color.

The hallucination story was always about your own tool lying to you. This one is about someone else’s document lying to your tool, on your behalf, without you knowing it happened.

Verification duties written this year assume the danger sits inside your own output. But other dangers exist inside somebody else’s input.

Read every document before an AI tool touches it. And maybe read it a little closer. Every time.

—Lawyer.exe

If your firm runs incoming filings through AI without anyone actually reading them first, forward this to whoever set that up.

Share

Amazon.com Services v. Perplexity AI, No. 26-1444, landed August 4. Perplexity’s Comet browser includes an AI assistant that navigates Amazon on a user’s behalf—browsing products, comparing prices, completing transactions the user directed. It sends browser screenshots to Perplexity’s servers for further instruction. Amazon argued Perplexity was the unauthorized accessor: their platform, their terms of service, Perplexity’s agent, Perplexity’s problem. The district court agreed and issued an injunction.

The Ninth Circuit reversed. The CFAA’s “whoever accesses” language contemplates a person. When a user directs an AI agent to act on their behalf, the user is the one who accessed the system. The agent is a tool, analogous to a browser. So, what does that mean?’

If the access is the user’s, so is the responsibility for what happens during that access. The agent browsed a system it wasn’t welcome on? Your access. The agent exceeded the scope of what the platform authorized? Your authorization problem. The CFAA’s penalties don’t land on the company that built the tool. They land on the person who pointed it at something.

The court’s browser analogy works when the agent does exactly what the user clicks. It gets harder when the agent interprets, navigates, and makes decisions the user didn’t specifically direct. A browser doesn’t decide which pages to visit. An agentic AI tool does. The more autonomous the agent, the more pressure this framework faces. The Ninth Circuit didn’t reach that question here. The next case will.

For platform operators, this narrows one weapon. The CFAA won’t police AI agent access in the Ninth Circuit. But terms-of-service breach (contract theory), trespass to chattels, and state computer fraud statutes survive. Amazon still has plays. They’re just not federal hacking statutes anymore.

For practitioners, the calculus cuts differently. Every AI research tool you deploy that touches an external system is acting under your authority. An AI agent that scrapes a database you don’t subscribe to? Your access. An agentic workflow that navigates a platform your opposing party controls? Your access. A tool that checks court dockets and hits a restricted filing? Also yours.

The supervision obligation here isn’t new. See Rule 5.3. You are responsible for the conduct of the tools you deploy, the same way you’re responsible for the work product of the paralegal you supervise. The Ninth Circuit just confirmed that this principle extends to AI agents under federal computer fraud law.

One circuit doesn’t make a rule. The Second and Fourth Circuits haven’t weighed in. Particularly in cases where the agent acts more autonomously than Perplexity’s Comet—an agent that independently decides which systems to access, what data to retrieve, and how to use it—the browser analogy will be a harder sell. That question is already being litigated in the public sphere.

Know what your tools touch. Know whose terms they’re subject to. The access is yours, at least in the Ninth Circuit.

—Lawyer.exe

If this hits close to home, forward it to whoever in your office just deployed an AI research agent without reading the terms of service.

Share

The tool is called nH Predict. Built by naviHealth, an Optum subsidiary. It takes limited patient data—cognition, mobility, ability to perform daily activities—and returns a discharge date. An algorithm’s prediction about exactly how much post-acute care a patient should need.

Family members of two deceased Medicare Advantage beneficiaries filed Estate of Gene B. Lokken v. UnitedHealth Group, Inc., No. 23-CV-3514 (D. Minn.), alleging UnitedHealth deployed nH Predict to supplant physician judgment. The algorithm told the patient when to leave. The physician’s call was secondary.

UnitedHealth pushed back on discovery the way insurers push back on everything that touches their “thought process”: proprietary algorithm, trade secret, competitive sensitivity. The standard play. It’s worked before. Courts have been reluctant to force disclosure of proprietary claims tools.

Magistrate Judge Elkins wasn’t buying it.

She granted discovery in six of seven categories. How nH Predict was built. Its development goals and anticipated benefits. Whether it was designed to supplant physician decision-making. Staff training records and protocols. What government investigators had already requested. Nearly everything about how the tool operates and who knew what about its role in coverage decisions.

What stayed out: internal investigations (privilege held) and financial incentive data. Everything else came in.

The distinction matters. The court didn’t order UnitedHealth to hand over nH Predict’s source code. It ordered disclosure of the process: how decisions were made, who made them, what role the algorithm played versus human judgment. A protective order handles proprietary code. But the question of whether an AI tool overrides physician judgment in coverage decisions isn’t a trade secret question. It’s a question about how the insurer makes decisions that affect patient care. That question is fair game.

If you’re plaintiff-side in insurance litigation, Lokken is a discovery roadmap. It tells you what to ask for and how a court is likely to evaluate those categories. Request the tool’s development purpose. Request the training materials. Request communications with regulators. The court already said yes to all of these.

If you’re defense-side, the lesson is simpler and more urgent: build the discovery response before litigation hits. If your insurer client deploys AI anywhere in the claims process, document the human oversight layer now. Create the record that shows physician judgment wasn’t overridden by the algorithm, it was informed by it. The difference between “our AI tool assists clinicians” and “our AI tool decides for clinicians” is the difference between a defensible process and a class action complaint.

For years, “proprietary algorithm” functioned as a one-line discovery objection that shut down entire categories of inquiry into how insurers make decisions. Lokken separates what’s genuinely proprietary—the code itself—from what’s merely inconvenient to disclose: the decision process, the override logic, the human oversight (or lack of it).

That distinction was always there. Courts are now enforcing it.

If your client uses AI to make coverage decisions, the decision process is discoverable. The code might be protected. The override isn’t.

—Lawyer.exe

If you’re on the insurance side, send this to your claims team before discovery hits. If you’re on the plaintiff side, save the case cite.

Share

Then you compare the agreement’s definition of “net sales” to the report’s actual deductions.

Four years of “channel partner fees.” A deduction the agreement never authorized.

“Net sales” is the most consequential defined term in any royalty-bearing license because it’s the denominator. Everything the licensee calculates royalties on flows through that definition.

Those definitions are dense, nested, and specific. The person preparing royalty reports is rarely the one who negotiated the license. Some finance analyst pulls the numbers from an ERP system that was configured once, years ago, by someone who may or may not have read the deduction carve-outs.

The audit clause exists. It’s in virtually every IP license agreement. But exercising it traditionally means engaging a forensic accountant, negotiating document access under the audit provision, and spending months comparing years of reports against a defined term that runs three paragraphs. The cost of checking often exceeds the expected recovery. So most licensors check only when the numbers look obviously wrong.

Subtle underpayment—the kind that comes from an unauthorized deduction category—compounds quietly. By the time anyone catches it, the statute of limitations has eaten the first few years.

Here’s what I run instead.

Step 1. Pull the license agreement and every royalty report in the relevant period

• The agreement first. Find the defined term for the royalty base—usually “net sales,” sometimes “gross revenue” or “licensed revenue.” Read the full definition, including all the nested terms it incorporates.

• Then pull the royalty reports. Every single one in the audit window. If the agreement’s audit clause has a lookback limit (typically two to three years), start there.

Step 2. Run the cross-check

• Feed both sets of documents to Claude. Use the full context window. This is exactly the kind of dense cross-referencing comparison where AI outperforms manual review: mechanical, definition-dependent, and spread across hundreds of pages.

• Prompt:

You are a litigation-support analyst reviewing an IP license agreement and associated royalty report(s).

Task: Cross-check the agreement's defined term for the royalty base — typically "net sales," "gross revenue," or equivalent — against the licensee's royalty report line items.

Instructions:

1. Extract the complete definition of the royalty base term from the agreement, including:
   - All inclusions (what counts as revenue)
   - All exclusions and permitted deductions (returns, allowances, taxes, etc.)
   - Any caps, floors, minimum royalties, or escalation provisions
   - Any sublicensing or channel revenue treatment

2. Map each component of that definition to the corresponding line item in the royalty report.

3. Identify and flag:
   - Revenue categories included in the definition but absent from the report
   - Deductions taken in the report that are not authorized by the definition
   - Ambiguous terms the report appears to interpret narrowly
   - Provisions (caps, floors, minimums) not reflected in the calculations

4. For each discrepancy, provide:
   - The specific contract language (with section/page reference)
   - What the report shows or omits
   - Direction of impact (overpayment or underpayment to licensor)
   - Confidence level: CLEAR (unambiguous discrepancy) or AMBIGUOUS (reasonable alternative reading exists)

Output: Numbered findings, ranked by estimated financial impact (largest first). Cite exact sections from both documents.

[PASTE LICENSE AGREEMENT AND ROYALTY REPORT(S) HERE]
**[NOTE: as always, you can just attach these documents or point to them in the prompt]**

• Read the output critically. The model will sometimes flag differences in terminology that aren’t meaningful discrepancies—”revenue” vs. “sales” when the definition treats them as equivalent. Filter for substance.

Step 3. Verify every finding against both source documents

• Treat the output like a first-year’s pre-audit memo. It tells you where to look. It does not tell you what to conclude.

• Every section reference, every deduction category, every claimed gap, confirmed against the actual contract language and the actual report line items before it goes anywhere.

• This step is not optional. A discrepancy the model flags that doesn’t hold up under manual review is worse than no finding at all, because you built an argument on it.

Step 4. Decide what to do with the findings

• If you’re on the licensor’s side: the findings are your roadmap for exercising the audit clause or, if the numbers are big enough, for a breach-of-contract demand. The cross-check tells you whether it’s worth spending real money on a formal audit.

• If you’re on the licensee’s side: run the same cross-check before the licensor does. Better to find the gap yourself and adjust than to get an audit notice and a demand letter in the same envelope.

• Either way, the prompt output is triage, not the finished analysis. The forensic accountant comes after you’ve identified which discrepancies are real and which are worth pursuing.

The audit clause in most IP licenses is a deterrent. Deterrents only work when they’re credible. For years, the economics made them hollow. Checking the numbers cost more than most licensors expected to recover, so they didn’t check. The clause sat there, technically enforceable and practically ignored.

AI makes the first-pass cross-check essentially free. When you can compare a three-paragraph “net sales” definition against four years of quarterly reports in ten minutes, the calculation changes. You don’t need to know there’s a shortfall before you look. You can just look.

That matters beyond any single recovery. A licensor who exercises the audit clause routinely changes the dynamic of the entire licensing relationship. The licensee knows you’re checking.

Read the definition. Run the numbers. The clause was always there. Now you can afford to use it.

—Lawyer.exe

Share

Subscribe for more workflows like this. And please spread the word!

Most fee agreements don’t say that out loud. They quote a rate, describe the scope in general terms, and let the hours do the explaining. That worked fine as long as the hours roughly tracked the difficulty of the work.

They don’t anymore. A motion that used to take four hours now takes forty minutes with a first draft from Claude sitting in front of you before you’ve finished your coffee. The rate on the engagement letter didn’t change. What that rate is actually buying did.

Here’s what most lawyers get wrong about what happens next.

Rule 1.5(a) lists eight factors for judging whether a fee is reasonable. Time and labor required is one of them. The others: the novelty and difficulty of the questions involved, the skill required to perform the work properly, the results obtained, the fee customarily charged in the locality for similar work, the amount involved, time limitations imposed by the client or circumstances, and the experience and ability of the lawyer performing it.

Seven of the eight factors have nothing to do with a clock. The rule was built to test value decades before anyone was drafting with a language model, and AI just made the existing gap between what the rule says and what most billing practices actually do impossible to miss.

That visibility carries a cost most lawyers aren’t bracing for.

A 2026 Integris survey of 600 law firm clients found that 35% had switched firms, or seriously considered it, over technology and operational failures, billing errors and unclear costs chief among them. Clients who feel like they weren’t told something material rarely call to complain about it. They call the next lawyer instead, and the first one finds out from a file-transfer request.

That’s the actual enforcement mechanism: a client who quietly stops being your client, for a reason you’ll never get to argue with. No bar complaint required.

The fix people reach for first is the wrong one. Discounting an invoice after a client asks a pointed question isn’t fee reasonableness. It’s damage control, applied after the letter already promised something the firm wasn’t planning to deliver. By the time you’re adjusting a bill defensively, the client has already learned the number on the engagement letter wasn’t the real agreement.

The fix belongs earlier, in the document that sets the deal in the first place. Florida’s ethics Opinion 24-1 already requires lawyers to notify clients if they intend to charge for AI tool use, and it places that obligation at the engagement stage, not the invoice stage. Other states will land on their own versions of this, but the direction is already visible: the disclosure duty is moving earlier, not staying put at the bill.

An engagement letter that describes what you’re delivering and why it’s worth what it costs doesn’t need a defensive adjustment later, because it was never selling hours in the first place.

The ethical work here happens at signing, not at billing. A lawyer who only fixes this at invoice time is already too late, because the letter already promised to sell something it isn’t actually selling anymore.

Billing honestly and selling honestly are not the same obligation. Billing honestly means the invoice matches the hours worked. Selling honestly means the engagement letter matches what the client is actually paying for. Only one of those is what Rule 1.5 was built to test, and it was never the one most fee agreements were written to satisfy.

—Lawyer.exe

Share

If your engagement letter still describes what you sell in hours instead of outcomes, that’s worth rewriting before a client does the math for you.

Miss it and the case doesn’t come back to you. Once a valid removal window closes, it’s gone. No equitable extensions. No explaining to the judge that the docket was heavy that week.

Most removal mistakes aren’t research mistakes. They’re calendar mistakes. Someone reads the original complaint, calculates thirty days from service under 28 U.S.C. § 1446(b), and stops looking.

Then the case changes. A party gets added. A party gets dismissed. Damages get amended past the jurisdictional threshold. Each of those events can start a new thirty-day window, or close one that was still open, and nobody re-runs the calculation.

Removability isn’t a fact you check once. It’s a status that can flip more than once over a case’s life.

Here’s the check I run every time the pleadings change.

Step 1. Log every event that could be a trigger date, not just the filing date

• Original complaint, if removability was apparent on its face. Any amended pleading. Any order. Any “other paper”—a discovery response, a settlement demand, a deposition admission—from which removability first became ascertainable.

• Treat this like a statute of limitations calendar with more than one possible accrual date, not a single deadline you calculate and file away.

Step 2. Run the full procedural history through Claude

Paste the case’s procedural history, service dates, and current party citizenship. Ask it to compute every possible trigger date and the resulting deadline for each.

You are a federal civil defense litigator computing removal deadlines under 28 U.S.C. § 1446. Based on the procedural history below, identify:

1. Every event in this case's history that could count as a "trigger date" under § 1446(b) — including the original pleading if removability was apparent on its face, and any amended pleading, motion, order, or other paper from which removability first became ascertainable
2. For each trigger date identified, the resulting 30-day deadline and whether it has already run
3. Whether the one-year bar on diversity-only removals (§ 1446(c)) applies to this case, and if so, when it expires
4. Whether any exception to the one-year bar (bad-faith joinder or forum manipulation) is arguably available if the standard deadline has passed
5. A clear statement of which removal window, if any, is currently open, and the date it closes

Do not fabricate citations. Describe all statutory provisions and standards in general terms and flag for verification.

Case type: [DESCRIBE]
Date original complaint filed: [DATE]
Date each defendant was served: [DATE FOR EACH]
Amended pleadings, motions, or orders since filing: [LIST WITH DATES AND WHAT CHANGED — e.g., party added/dismissed, damages amount changed, new claim added]
Current citizenship of all parties: [LIST]

**[NOTE: as always, you can just attach these documents or point to them in the prompt]**

Step 3. Check the one-year bar separately from the 30-day clock

• If the case wasn’t originally removable on diversity grounds, the one-year bar under § 1446(c) applies regardless of when a later trigger date arises, subject to a narrow bad-faith exception. This is the deadline most likely to get missed, because it isn’t the deadline anyone’s watching for. Everyone’s watching the thirty days.

Step 4. Verify every date against the actual docket before filing

• The model’s calculation is only as good as the dates you gave it. Confirm each service date and filing date against the court’s actual docket, not your memory of the case timeline.

• This step is not optional.

I had a defense-side commercial dispute removed on diversity grounds well within the standard window. Months later, the plaintiff amended the complaint to add a non-diverse party, then voluntarily dismissed that party a few weeks after that. The removal clock didn’t run once in that case. The original removability, the destruction of diversity by the amendment, and its restoration by the dismissal each raised their own timing question, and only one of the resulting windows was still open by the time anyone stopped to check.

Removal isn’t a decision you make once at intake. It’s a status you have to keep reopening every time the pleadings move, because the event that makes a case removable and the event that makes the window closed are often the same fact pattern looked at from two different deadlines.

The lawyers who miss these deadlines aren’t bad lawyers. They’re lawyers who did the calculation once, correctly, and never did it again.

Recalculate every time something changes. The clock doesn’t wait for you to notice it reset.

Share

—Lawyer.exe

Subscribe for more workflows like this. And please spread the word!

EvenUp has raised $370 million, Eve $164 million, Supio $85 million, and Darrow $63 million, a combined $682 million, all of it purpose-built for demand-letter drafting and case valuation on the plaintiff side.

Defense-side legal AI doesn’t have an equivalent list. The same reporting that tracked those four rounds describes the defense side as “underdeveloped,” and says plainly there is still no clear, scaled, venture-backed winner built specifically around defense-side litigation intelligence.

The absence of a number is itself the finding. Nobody has funded enough defense-side companies to produce one worth reporting.

Insurance carriers and defense counsel are litigating with tools built for something else entirely: contract review, corporate due diligence, general document search, retrofitted for a fight they weren’t designed to handle.

I’ve worked both sides of this fight, and I can tell you exactly what that gap looks like from inside it.

On the plaintiff side, I’ve watched a demand go out the same week a client’s treatment ended, valued against comparable verdict and settlement data before the adjuster’s own file is even complete. On the defense side, I’ve watched a carrier run the identical fact pattern through a platform built for M&A due diligence, because nothing purpose-built for defense litigation exists inside that budget. At least not yet. Same underlying legal work. A full technology generation apart in the tooling.

That split reads directly off where venture capital already understands the math. Contingency fees look like a growth curve on a pitch deck, more cases, faster turnaround, a repeatable multiple on invested capital. Hourly defense billing looks like a cost center, and cost centers are not a story venture capital funds.

Litigation is adversarial by design. When only one side of an adversarial system gets purpose-built tools, the system doesn’t get more efficient. It gets more lopsided, one demand, one valuation, one motion at a time, and the side without the tooling is the one paying for the difference, whether that’s an insurance carrier’s loss reserves or a defense firm’s hours against a case that resolves faster than they can staff it.

If your carrier or your defense practice is still running litigation through software built for something else, the other side is actively widening that gap every week you wait to close it.

Share

—Lawyer.exe

If this was useful, forward it to whoever’s about to renew a legal AI contract on your defense desk.

Two pages later, my own expert reads a different ambiguous finding and calls it conclusively traumatic.

Same report. Same kind of ambiguous evidence. Opposite standard, and nothing in between explaining why.

I read that report twice before it almost went out and missed it both times.

Forty pages of dense medical language, written by someone else, about a methodology I didn’t train in, will hide a contradiction that lives three sections apart from a normal read, careful or not. A typo jumps off the page. A logic problem buried between page four and page eleven doesn’t.

Opposing counsel finds it, though. They read rebuttal reports looking for exactly this, and a cross-examination built on “you criticized this exact same kind of inference on page four, doctor” doesn’t need much else to work.

Here’s what I run before any expert report goes out the door.

Step 1. Gather the three documents before you open Claude

• The opposing expert’s report in full.

• Your own expert’s rebuttal report in full.

• The underlying data your expert relied on: imaging findings, records, whatever the report is actually built on.

Step 2. Run the consistency check

• Paste all three into Claude.

• Prompt:

You are a civil litigator reviewing your own retained expert's rebuttal report before service. Compare it against the opposing expert's report and the underlying data or methodology materials I'm providing, and identify:

1. Unaddressed arguments: points in the opposing expert's report that my expert's rebuttal does not respond to
2. Non-responsive rebuttal: places where my expert's rebuttal responds to an argument the opposing expert did not actually make
3. Internal methodology inconsistencies: any instance where my expert criticizes a methodology, approach, or assumption in the opposing report, then relies on the same or a comparable methodology, approach, or assumption elsewhere in their own report, without addressing the difference
4. Unsupported factual assumptions: statements in my expert's rebuttal that are not traceable to the underlying data or record materials provided
5. A short list of the specific gaps most likely to be exploited on cross-examination

Do not evaluate the underlying scientific or technical methodology itself. Flag issues for my expert to address, not conclusions about which methodology is correct.

Opposing expert's field and report: [PASTE OPPOSING EXPERT REPORT]
My expert's rebuttal report: [PASTE REBUTTAL REPORT]
Underlying data / methodology materials: [PASTE OR SUMMARIZE]

**[NOTE: as always, you can just attach these documents or point to them in the prompt]**

Step 3. Verify before you take anything back to your expert

• Every flagged inconsistency gets checked against the actual record before it goes anywhere. The model works from what you pasted, not the case file it hasn’t seen.

• Treat this the same way you’d treat a paralegal’s first pass at a privilege log: a good start, not a finished product.

Step 4. Give your expert the chance to fix it, not you

• Sometimes there’s a real medical distinction between the two findings your expert read differently, and your expert can explain it in one sentence. Get that sentence into the report.

• Sometimes there isn’t, and your expert needs to pick a consistent standard before this goes anywhere near a deposition transcript.

• Either way, that’s a conversation with your expert, not an edit you make yourself. Flagging the methodology question is your job but resolving it is theirs.

Treat the report the way you’d treat a fact witness’s inconsistent testimony before you put them on the stand: find the contradiction yourself, make them explain it, and do it before opposing counsel gets the chance to make the explanation happen live in front of a jury.

A rebuttal report isn’t in the file to attack the other side’s expert. It’s there to survive the exact scrutiny you’re asking the jury to apply to the other side. If your own expert can’t say why the same kind of ambiguous finding means two different things depending on which page it’s on, opposing counsel is going to ask that question first, under oath, with your expert on the stand and no time to think about the answer.

Read your own expert’s report like you’re opposing counsel. Because eventually, someone will.

Share

—Lawyer.exe

Subscribe for more workflows like this. And please spread the word!

For two years, an AI hallucination sanction has functioned like a parking ticket. Embarrassing, a little expensive, survivable. The attorney whose name is on the brief pays it, writes the mea culpa, and the firm moves on.

That model is ending.

Above the Law named the actual failure this month, and it isn’t a lawyer who doesn’t understand AI. It’s associates quietly running client work through public tools because they assume everyone else is doing the same thing, confidential material landing in places nobody vetted, and a written policy that’s either nonexistent or too vague to mean anything. That’s a governance gap: an org chart with nobody standing in it, dressed up as one lawyer’s failure to understand a tool.

Here’s what’s actually changing, and why it lands on firm leadership instead of the associate.

A written AI policy is paperwork until somebody actually reads the brief before it goes out. Above the Law’s framing surfaces exactly that failure. Gordon Rees, an Am Law 100 firm, got accused of filing an AI-hallucinated brief. Then it happened again. A policy that gets ignored twice was never operational. The control that actually stops this is whoever sits in the seat that reads the brief before it goes out, every time, and that seat was empty both times.

Regulators are writing that distinction into law. California’s ethics committee approved six proposed amendments to the Rules of Professional Conduct this spring, folding AI-specific supervision obligations directly into the rules governing how managing lawyers run their shop. Unlike the advisory guidance every other state bar has issued, these amendments would carry disciplinary authority if adopted. The comment period closed May 4. A vote is next. When it lands, “we have a policy” stops being an answer to whether a firm supervised its AI use, because the rule asks what the firm actually did, not what it wrote down.

And the money already knows where to look. I spent years on the insurance defense side before doing this full time, and I can tell you exactly how a carrier reads a sanctions order: as a finding of fact somebody else already made for them. A Rule 11 order that fines an associate for a hallucinated citation is also, quietly, a pre-built negligence exhibit. If that mistake caused client harm, the malpractice claim that follows doesn’t stop at the associate’s bar number. It runs up the chain to whoever was supposed to be supervising and wasn’t. The firm’s malpractice policy is the one actually on the hook.

None of this required a new rule to be true. The duty to supervise associates existed before any of these lawyers had a bar card. What changed is that AI made the failure fast enough, confident enough, and common enough that regulators, carriers, and reporters all started looking at the same gap at the same time. A firm that treats “we have an AI policy” as the finish line is answering a question nobody serious is asking anymore. The question is who actually checked.

The associate who filed the brief still gets sanctioned. But the next name on that order is going to belong to whoever was supposed to be watching, and didn’t.

Share

—Lawyer.exe

Forward this to the partner at your firm who thinks a one-page AI policy is the whole job.

A defendant's statement of undisputed facts (SUF) is not a neutral inventory. It's an argument dressed up as a list, and every fact on it got there because it helps the motion.

The tell is in the citations, not the sentences. A movant quoting their own expert for "pre-existing and unrelated" is quoting exactly as much of that expert's report as helps them, and not one sentence more.

I audited a rear-end collision case where the SUF made that exact move. The defense IME doctor's report was cited for the proposition that the plaintiff's cervical injury was "pre-existing and unrelated to the subject collision." Four paragraphs past the sentence the motion quoted, the same doctor's own report said the collision had aggravated the pre-existing condition.

Opposing counsel doesn't necessarily lie in a SUF. They just stop reading their own exhibit exactly where it stops helping them.

Here's how I catch it now.

Step 1. Compile the record before you touch the SUF

• Pull the full text of every exhibit the SUF cites, not just the paragraph or page cited. Depositions in full, expert reports in full, medical records in full.

• If the SUF cites page 4, you need the whole exhibit, not page 4. The part you're looking for is usually many pages later.

Step 2. Run the fact-by-fact audit

• Paste the statement of undisputed facts and the full underlying exhibits into Claude.

• Prompt:

Compare each numbered fact in this statement of undisputed facts against the full exhibit cited for it. 
For each fact: 
(1) state whether the citation actually supports the fact as written, 
(2) identify any language elsewhere in the same source document that qualifies, contradicts, or undercuts the fact, and 
(3) flag where the fact as stated is broader than what the citation actually establishes. 
Quote the qualifying or contradicting language with page and line.

• This is the step that finds the aggravation sentence. It's also the step most opposition briefs skip, because reading the movant's whole exhibit feels like the movant's job, not yours.

Step 3. Verify every flagged citation against the original document

• Treat the output like a first-year associate's memo: confirm the quote exists, confirm the page and line, confirm it means what the model says it means before it goes anywhere near your brief.

• This step is not optional. An opposition that misquotes the movant's own exhibit is worse than one that just concedes the fact.

Step 4. Sort every flagged fact into one of three buckets

• Genuinely controverted, with your own pin cite to the record.

• Conceded or distinguished, because some facts really are undisputed and pretending otherwise wastes the court's patience along with yours.

• Overstated relative to their own citation, which is its own argument even on the facts you can't fully controvert.

That audit used to mean reading every exhibit the movant cited, cover to cover, on the chance something got left out of the quote. Now the first pass takes about an hour, and it tells you exactly which of the movant's "undisputed" facts are actually just facts nobody checked.

The "undisputed" facts section wins or loses summary judgment motions more often than the argument section does, because judges read it first and it sets the frame for everything that follows. A movant who wrote "undisputed" over a fact their own witness qualified isn't asking the court to accept a fact. They're asking the court not to read the whole exhibit. Most oppositions grant that request without meaning to.

Read the whole exhibit. Every time. That's the actual opposition.

Share

—Lawyer.exe

Subscribe for more workflows like this. And spread the word.

You already know this in your hands, even if you’ve never named it. You’ve run a search in one database, come up empty, run the same search in another, and found the case sitting right there. Same query, different corpus, different result. The tool didn’t fail. It just couldn’t see what it was never given.

That’s the whole game. A legal research AI is only as good as the body of law underneath it. The model on top is increasingly a commodity—Anthropic, OpenAI, and Microsoft all launched legal offerings this year, and they are converging on the same capabilities. The caselaw corpus is the part that’s scarce. There are only about three comprehensive databases of U.S. primary law in existence.

So watch what happened while everyone was comparing chatbots.

Clio bought vLex and Fastcase for a billion dollars. That put one of those three comprehensive caselaw databases inside a single company, alongside the practice-management software running a huge share of small-firm America. The strategy has a name in the reporting, and the name is “data moat.” They didn’t buy a search tool. They bought the ground every search tool has to stand on.

And that ground is now contested. Fastcase, as part of Clio, is in federal court against the legal AI company Alexi. The case is pending in the D.D.C. in front of Judge Richard J. Leon. Alexi’s counterclaim is the part worth reading: it alleges the acquisition of that database is an unlawful “clog on competition” under Section 7 of the Clayton Act, because whoever controls the primary-law backfile controls what every downstream tool can be built on.

Set aside who wins. The lesson is the same either way.

When you standardize your firm on a research AI, you are not buying software. You are placing a multi-year bet on the database underneath it, and that data layer is consolidating into a few hands while lawyers argue about interface design.

Here’s what that means for how you choose, whether you’re a solo or you run procurement for a fifteen-lawyer shop.

Ask whose corpus it is. Every research tool sits on someone’s caselaw database. Find out whose. If the answer is vague, that’s the answer.

Ask what happens to your access if that owner is acquired. Licensing terms, pricing, and coverage all change after a deal closes. The access you’re quoted today is contractual, not permanent, and the contract can be assigned to someone who bought the company for its data.

Ask whether you’re locked in when the data moves. If your tool, your practice-management system, and your research corpus all live under one roof, that’s convenient right up until the roof changes hands. Convenience and dependence are the same arrangement viewed from different quarters.

Price the switch before you need it. Know what it would cost—in time, in retraining, in migrated work product—to move off the tool if its data supply changed under you. If you don’t know, you’re not standardized. You’re captured.

None of this means don’t adopt. It means adopt with your eyes on the layer that actually matters. The interface is what the vendor sells you. The corpus is what you’re actually renting, and right now the landlords are consolidating.

The model was never the moat. The caselaw is. Choose your tool like you understand the difference, because the companies spending billions certainly do.

—Lawyer.exe

If this was useful, forward it to the lawyer in your firm who signed the research-tool contract without asking whose database it runs on. And please spread the word!

Last year I had a premises-liability case, plaintiff side. The client slipped on a spill in the entryway of a regional retail chain and tore a rotator cuff. Real injury, expensive, and exactly the kind of soft-tissue claim a defense adjuster waves off as pre-existing.

The whole case turned on the store’s own procedures. How often the floors were supposed to be swept. Whether anyone logged it. Whether the written policy matched what actually happened that afternoon.

So I noticed the chain’s 30(b)(6) designee on exactly those topics. Spill-response protocol, inspection-log practice, employee training, incident reporting. Going in, the corporate rep sounded airtight on the policy in the abstract.

The opening I needed was never in what the manual said. It was whether anyone followed it that day, and whether a designee flown in from out of state could speak to that specific store’s sweep log at all.

You get one real shot at a corporate designee. Walk in with the notice and a legal pad, and you take the deposition the company prepared him to give: the clean policy recitation, the “our protocol requires hourly inspections,” the answers that sound like compliance and prove nothing about that day.

Here’s the prep I run instead. Most of it happens before I even finalize the notice.

Step 1. Map every noticed topic to the documents that prove it or break it.

• Take the deposition notice and put each topic next to the specific paper in the production that establishes it or blows it up. Sweep logs. The written cleaning policy. Training sign-offs. The incident report.

• You’re building a topic-by-topic exhibit plan, not a question list. For the slip-and-fall, that meant lining the corporate policy manual up against the actual inspection log for that store on that date—two documents the company would rather you treat separately.

• This is your element chart, one row per topic. You already build one for a summary-judgment opposition. Same discipline, pointed at a witness.

Step 2. Flag the topics the designee can’t actually speak to.

• A 30(b)(6) witness binds the company on everything noticed. The prep gap you want to find is the topic where the designee has to say “I don’t know” on the record, because that answer often does more work than anything he could tell you.

• The corporate rep knew the national policy cold. He had never set foot in the store. He could not say whether the sweep log was filled out contemporaneously or completed after the fact. Notice the topic narrowly enough that the gap becomes his problem.

• Run the notice through Claude and ask one question: which of these topics requires store-level, day-of knowledge a corporate designee is unlikely to have? Those are the pressure points.

Step 3. Run the topic list against their likely objections.

• Which topics are broad enough that opposing counsel will instruct, limit, or move for a protective order. Which of those you actually need, and which you can trade away in the meet-and-confer to keep the ones that matter.

• Going in knowing which topics you’ll concede is how you keep the ones you won’t.

Step 4. Build the outline in exhibit order, not topic order.

• Sequence the deposition so each document lands when it does the most damage. Establish the policy first, in the designee’s own words. Then hand him the log.

• The distance between what he just committed to and what the paper shows is the testimony you came for.

Here’s the actual prompt. Copy it, fill in the brackets with your own notice and production, and run it before you finalize the outline.

You are a civil litigator preparing to take a Rule 30(b)(6) deposition of a corporate party. [NOTE: You can make this much more specific to your case.] The topics have already been noticed. Based on the noticed topics, the case description, and the documents below, produce:

1. A topic-by-topic exhibit map: for each noticed topic, identify the specific documents in the production that establish it or undermine the company's position—cite them by Bates number or description
2. Designee knowledge gaps: for each topic, flag whether it requires store-level, day-of, or otherwise localized knowledge a corporate designee is unlikely to have versus knowledge a designee prepared from a policy manual could plausibly cover
3. Objection forecast: which topics are broad enough to draw an instruction, a limiting objection, or a motion for protective order—and which of those are worth conceding in a meet-and-confer to protect the topics that matter most
4. An exam sequence in exhibit order, not topic order: the order to introduce documents so each one lands when it does the most damage, building toward the exhibit that creates the sharpest contrast with the designee's own testimony

Do not fabricate documents or Bates numbers—work only from what I provide below or what is in the file, and flag any topic where I have not given you enough documents to build the map.

Case description: [What the litigation is about, and your client's position]
Noticed topics: [List each topic from the 30(b)(6) notice, or attach/link to the notice]
Key documents in your production: [List or describe them — policy manuals, logs, training records, incident reports, whatever is relevant. You can also attach/link to them]
What you most need the designee to be unable to explain: [The gap you're hoping to expose]

Watch for the model treating a policy document as if it answers a day-of question. Read the map critically and confirm each document actually reaches the topic it’s assigned to. And verify every Bates or exhibit number before it goes into your outline.

That whole build used to take the better part of two days for a document-heavy corporate deposition. Now the first pass takes an afternoon. I still run the deposition. The model doesn’t read the witness, doesn’t decide when to stop, doesn’t hear the half-second hesitation before an answer that tells you where to push. But I walk in against a map instead of a topic list.

A 30(b)(6) is the one deposition where the witness is a stand-in. You’re examining the company’s memory, and the designee is just the mouth it speaks through. The real question is never what he remembers. It’s what the company is willing to commit to under oath, and exactly where its official story runs out.

AI doesn’t take that deposition for you. It makes sure you know, before you sit down, which topic the company can’t answer. That’s the difference between a transcript that locks in liability and one full of policy recitations you can’t use.

The designee binds the company. Your job is to know, before he does, which answer binds it to you.

—Lawyer.exe

Subscribe for more workflows like this. And spread the word.

That’s the wrong time to think about it.

By the time you’re two weeks out from trial, the problems are expensive. The witness who could have authenticated the key document has been transferred, is no longer reachable by subpoena, or was never on your witness list to begin with. The business records custodian you needed is a third party and the discovery period closed six months ago. The deposition you took to establish a critical link in the chain contains a page-and-line gap that doesn’t cover the exhibits you planned to use it for.

Every one of those problems was visible. Not at trial—months before it.

Trial prep is where most litigators spend their preparation energy, and it should be. But there’s an audit that has to happen before trial prep starts, and most lawyers either skip it or run it informally, in their head, while they’re already inside the prep schedule with no time left to fix what they find.

The audit is a per-exhibit foundation map: every planned exhibit, mapped against its foundation requirements, the witness who carries those requirements, and what happens to the exhibit if that witness is unavailable.

Here’s how I run it.

Step 1. Build the exhibit list before anything else

• Pull every document, photo, video, demonstrative, and piece of expert material you expect to offer at trial.

• Don’t organize by exhibit number yet. Organize by type: documents, photos and video, expert materials, demonstratives.

• Include everything you might want; cutting is easier than finding something you missed.

• I give Claude the complaint and ask it to pull every exhibit type mentioned in the factual allegations before I start organizing. It surfaces materials I’ve noted in discovery but haven’t formally added to the exhibit list yet.

Step 2. Map each exhibit to its specific foundation requirements

• For each exhibit, write down the foundational elements required for admission under the applicable rule.

• Documents: FRE 803(6) business records require a custodian who can testify to regularity, ordinary course, and that the record was made at or near the time of the act by someone with knowledge. Do you have that witness designated and on your list?

• Photos and video: Who took it? Who can authenticate it? Is there a chain of custody gap that needs closing before trial?

• Expert reports: Your expert must be on the witness list and available on the day you need the report in evidence. If you’re anticipating a Daubert challenge, is the expert prepared for a separate hearing before the report comes in?

• Demonstratives: Who built it? Does it require expert sponsorship, or can a fact witness authenticate it?

Step 3. For each exhibit, identify the witness who carries the foundation

• One column: the exhibit. Next: the foundation witness. Next: that witness’s current status—retained, designated, third party, subpoenaed, or unknown.

• Any exhibit with “unknown” in the witness column is a solvable problem if you catch it now. You won’t solve it in week two of trial prep.

• Any exhibit with a third-party foundation witness who hasn’t been subpoenaed is a scheduling problem with a clock on it.

Step 4. Check for single-point-of-failure dependencies

• Identify any exhibit where one witness is the only authentication route.

• Count how many times each witness appears in the foundation column. A single witness authenticating four exhibits and carrying a key element of the theory is a structural vulnerability. If that witness becomes unavailable six days before trial, you need to know that now.

• For each single-point dependency, build a contingency before trial prep starts: a stipulation you can request, a self-authentication option, an admissions request that establishes the foundation without a live witness.

Step 5. Map deposition designations against the exhibit list

• If you’re offering deposition testimony at trial, confirm your designations cover the authentication exchanges, not just the substantive testimony.

• A witness who authenticated a document in their deposition but whose designation omits those pages cannot establish foundation for that exhibit at trial. Check the transcript. Check the designation.

• This check takes less than an hour per deponent. Missing it costs more time than that.

How I run this with Claude

I don’t build the exhibit foundation map by hand anymore. I paste the complaint and the working exhibit list into Claude with this prompt:

Here is my planned exhibit list: [paste]. 

Here are the claims and the elements I need to prove: [paste complaint or outline]. 

For each exhibit: 
(1) identify the foundational rule that applies — business records, authentication, expert sponsorship, self-authentication, or other — and what that rule specifically requires; 
(2) identify which witness must carry that foundation; 
(3) flag any exhibit with only one authentication route and no available alternative; and 
(4) identify any deposition designations that need to cover authentication exchanges and not just substantive testimony. 

Output as a table organized by exhibit type. Then list single-point-of-failure dependencies in priority order.

It returns a working foundation audit in a few minutes. Step 4, the single-point-of-failure check, is where it earns its keep. It catches concentrations you’d miss when you’re building the exhibit binders one at a time.

In a commercial breach of contract case, a key financial exhibit was a ledger from a wholly owned subsidiary of our client, not the client itself. Everybody in the room knew the subsidiary was ours. Nobody had thought through the foundation problem: the subsidiary was a separate legal entity, its records were the subsidiary’s records, and our client’s witness couldn’t authenticate them from personal knowledge of the subsidiary’s record-keeping practices without laying a specific factual basis for that knowledge.

We had the client on the witness list. The subsidiary’s records custodian was nowhere on it. The exhibit was central to the damages calculation. Running the foundation map about ten weeks before trial surfaced the gap—the foundation column for that exhibit read “unknown” the minute I tried to fill it in with an actual person who could lay FRE 803(6) foundation. The fix was simpler than it looked: we designated the subsidiary’s controller as an additional witness and prepped a narrow custodian authentication covering the specific ledgers at issue. Took half a day.

Finding the same problem during exhibit prep week would have required a motion to substitute, an amended witness list, and a fight about prejudice we didn’t need. We had time to avoid all of it because we ran the audit before we built the schedule.

The foundation requirements haven’t changed. Either the exhibit comes in or it doesn’t, and you find out which one in front of the jury if the problem surfaces after you’re already in the courtroom. What changes when you run this audit early is whether you have time to fix the gap.

Most exhibit foundation problems aren’t hard problems. They’re timing problems. The difference between a fixable gap and an unfixable one is usually measured in how early you looked.

Verify the witness list against the exhibit list before you build the trial prep schedule. The map takes a day. The gaps it finds take weeks to fix if you find them at trial.

—Lawyer.exe

Subscribe for more workflows like this.

Sometimes they don’t.

The witness who needs to testify before a key exhibit can come in isn’t scheduled until day three. The exhibit that establishes a central element of the claim rests on testimony from a witness who became unavailable six days before trial. The expert whose opinion carries the causation theory is a single point on a chain with no alternative route.

These aren’t bad luck. They’re structural. They were visible months before trial in the dependency graph of the case—which witnesses must testify before which exhibits can come in, which exhibits establish which elements, and what breaks in the case theory if a single node goes down.

The trial prep schedule addresses tasks. Most litigators never build the map that shows which tasks are load-bearing and which are contingent on other tasks completing first. Building that map before trial prep starts changes what the schedule looks like—and catches the vulnerabilities while there’s still time to build around them.

Here’s how I build it.

Step 1. Start with the elements of the claim—every claim, every defense

• Write down every element you’re carrying before you open a witness binder or an exhibit folder.

• For each element, identify the evidence you’re counting on to establish it: the specific exhibit, the specific witness, or both.

• This is the foundation layer of the dependency map. Every other node connects upward to an element of the claim.

• I give Claude the complaint and ask it to generate the full element checklist before I open anything. It catches elements buried in the boilerplate that don’t make it onto the prep board.

Step 2. Map exhibit dependencies

• For each exhibit: which element does it establish? Which witness provides foundation? Is that witness your only route to the exhibit?

• One exhibit, one witness, one element—any chain that runs through a single path with no alternative is a vulnerability the map will surface immediately.

• Build this column by column. The concentrations appear faster than you’d expect.

Step 3. Map witness dependencies

• For each witness: which elements do they establish? Which exhibits do they authenticate? What happens to the case if they’re unavailable?

• A witness who carries two primary claim elements and authenticates five exhibits is a structural concentration of risk. That isn’t visible when you’re prepping individual witness outlines in isolation. It’s visible only when the full dependency chain is on one page.

• Running the full witness list through Claude with a prompt asking it to count element and exhibit dependencies per witness surfaces the concentrations faster than building the table by hand.

• Any witness with more than two entries in the dependency column gets a contingency plan before trial prep starts, not during it.

Step 4. Identify the structural vulnerabilities

• Look for cascade points: nodes where one failure affects multiple elements simultaneously.

• Look for linear chains with no alternative route. Evidence that can only come in through one witness, via one exhibit, establishing one element is the kind of chain that doesn’t recover when the first link goes down.

• Look for conditional sequences: witnesses who must testify before certain exhibits can be offered. If the sequence breaks because of scheduling, the exhibit may be inadmissible even if the underlying foundation is otherwise sound.

Step 5. Build the contingencies before trial prep starts—not during it

• For each structural vulnerability, identify the contingency: a stipulation request to opposing counsel, an alternative authentication route, a backup witness for the same element, a hearsay exception that provides a parallel foundation path.

• The contingency doesn’t have to be your first choice. It has to exist before the prep schedule is set. A mid-prep scramble for a backup witness when every calendar slot is already allocated costs more than the time it would have taken to identify the problem in the first place.

• If a vulnerability has no available contingency, that’s information too. It changes how aggressively you protect the primary witness and primary route—and what you tell the client about the trial risk.

How I run this with Claude

I don’t build the dependency map manually anymore. I paste the complaint, a witness list, and an exhibit list into Claude with this prompt:

Build a trial dependency map from the following materials. Here is the complaint: [paste]. Here is my witness list: [paste]. Here is my exhibit list: [paste]. For each claim and defense: (1) identify which exhibits establish each element, (2) identify which witnesses are required to authenticate or sponsor each exhibit, (3) flag any element carried by only one witness or one exhibit with no alternative route, and (4) identify any conditional sequences where a witness must testify before an exhibit can be offered. Output as a structured table organized by claim element. Then list structural vulnerabilities by priority.

The output isn’t the final map, it’s the first draft of it. Claude surfaces the concentrations. The judgment about what to do with them is still yours.

In an IP interference case where our client was asserting ownership rights, the chain of title ran through five prior assignments. We had an expert to trace the chain and calculate damages using industry licensing benchmarks. The dependency I missed on first pass: the expert’s chain of title opinion was only as good as the underlying assignment documents in evidence, and the second transfer in the chain had been executed by a company that had since dissolved. The only person who could authenticate that assignment was the former principal of the dissolved entity—alive, reachable, and not on anyone’s witness list.

Without that document coming in, the chain of title had a gap in it, and the expert’s ownership opinion had nothing to stand on at that link. The defendant’s entire “you have no right to this IP” theory would have had a gift-wrapped evidentiary argument handed to it. I caught it building the dependency map about eight weeks out. We subpoenaed the former principal, got the authentication on the record, and designated the testimony. If I’d been running exhibit prep the way most litigators do—binder by binder in the two weeks before trial—the deposition window would have been closed by the time that gap surfaced.

The case theory is only as strong as its weakest dependency. The right witnesses, the right exhibits, and the right arguments can all be in place while the dependency structure underneath them contains a single-point failure that doesn’t become visible until the wrong moment in front of the jury.

The map doesn’t change what trial looks like. It changes whether you walk into it knowing where the vulnerabilities are—and whether you’ve already decided what to do about them.

Build the map before you build the prep schedule. It takes a day. Finding the same problem during trial preparation costs two weeks. Finding it at trial costs more than that.

—Lawyer.exe

Subscribe for more frameworks like this.

You've read the brief. You have opinions about it. Some are right. Some are reflexive. You won't know which is which until you've been through it twice and actually given it some thought.

The manual process: annotate as you read, build a response outline from the annotations, discover at draft hour that you missed the argument buried in footnote 9. That's how response briefs have always been written, right?

What the model does that you don't: it doesn't get anchored. It maps the document systematically—what the brief argued, what it should have argued, where the reasoning doesn't close. It also surfaces what the opposing party chose not to argue, which is usually more useful than what they did.

The argument a brief doesn't make is usually stronger than the one it does.

Step 1. Pull the opposing brief

• Full text—not a summary. PDF copy-paste into the context window, or attach the file directly.

• Give the model your theory in one sentence before pasting the brief. "Our position is [X]" focuses the vulnerability analysis on what actually matters to the outcome of the motion—not every arguable weakness, the ones that move the needle.

Step 2. Run the Opposition Brief Weaknesses Identification prompt

The prompt:

```

You are a litigator preparing a response brief. Analyze the opposing brief below and identify:

1. Legal arguments that misstate or misapply the applicable standard — be specific about what the correct standard is and how the brief departs from it

2. Factual assertions that appear to contradict the record or that are unsupported by cited evidence

3. Internal inconsistencies — places where the brief's arguments conflict with each other

4. Arguments the brief conspicuously avoids making that would have been stronger than what it did argue

5. The single weakest argument in the brief, and the single strongest argument I will need to address seriously

For each issue identified, note the specific page/section of the brief.

Do not fabricate case citations. Describe any legal standards you reference generally and flag them for verification.

My client's position: [Plaintiff / Defendant — brief description of our theory]

[PASTE/ATTACH OPPOSING BRIEF HERE]

```

• Section 4—arguments the brief conspicuously avoided—is the most important output. The opposing party mapped their own case before filing. What they chose not to argue tells you what they knew was weak. Those arguments typically surface in reply briefs when they're most damaging to you. Find them before the reply.

Usage Notes:

• Watch for: The model may flag legal standards as misapplied without reading your circuit's current authority. "The correct standard is X" needs verification before it goes anywhere near your brief.

• Verify before use: Every legal standard the model identifies as misstated—confirm from primary sources in your jurisdiction before you draft the response.

• Time saved: Replaces 2–3 hours of annotated read-through. The output is a working brief attack plan.

PRO TIP: Sonnet 5 handles this well—brief analysis is more about systematic reading than deep reasoning. Step up to Opus or Fable if the brief involves a complex multi-factor standard or a particularly dense procedural argument. Effort level: high. Run one brief at a time; multi-brief sessions produce muddled output.

Step 3. Map the response, verify the standards, then draft

• Lead your response with their weakest argument, not your strongest. Weaken them before you strengthen yourself—a court that's already skeptical of their position is more receptive to yours.

• Section 5 (single strongest argument) is not optional reading. It tells you where to concentrate your rebuttal energy before you've written 400 words on something the court will skip.

• If the output is unfocused: "Focus specifically on the [X] claim. What is the legal standard the brief had to meet for that claim, and where specifically does the argument fall short of it?" One claim at a time produces cleaner analysis than asking for the whole brief at once.

I tested this prompt this morning on a case in which I already finished the Reply brief. The output was obviously not exactly the same, but it was quite similar and as a first pass output that took less than one minute to set up, it was very good. And another thing to consider is that the output had *literally zero context* other than the Opposition. That's when you know you have a good prompt.

Once you run this thing, you have a documented brief attack plan before you've written a word of your response.

Verify every misapplied standard against primary sources. Make the strategic call on which holes are worth pressing—the model maps them, you pick the fights. Then draft.

—Lawyer.exe

Subscribe for more workflows like this.

The complaint lands. Your answer is due in 21 days. Somewhere in that window, you need to decide whether there’s a 12(b) motion worth filing—or whether you’re answering and moving on.

That decision is worth making with 10 minutes of AI triage and 20 minutes of verification. The alternative is 90 minutes of manual read-through, or skipping the analysis and hoping you didn’t miss a standing issue.

The triage question is simple: is there a motion worth filing? Answer that first. Build the brief later if yes. Miss a 12(b)(1) standing problem on the face of the complaint and you’ve either waived it or discovered it at an inconvenient time.

The model won’t read your jurisdiction’s case law. That’s your job. What it does—fast and reliably—is map the arguments, rate their strength, and tell you where to spend your 20 minutes of verification time.

Step 1. Pull the full complaint text

• Full text, not a summary. Same rule as every other document analysis prompt.

• Specify the jurisdiction and court. The output is only as good as the context you give it.

Step 2. Run the Motion to Dismiss Argument Identifier

The prompt:

```

You are a civil defense litigator analyzing a complaint for 

motion to dismiss potential. Review the complaint below 

and identify:

1. Any Rule 12(b)(6) arguments — specifically: claims that 

   lack sufficient factual allegations under Twombly/Iqbal, 

   elements that are conclusorily pleaded, or legal theories 

   that are not cognizable in this jurisdiction

2. Any Rule 12(b)(1) arguments — subject matter jurisdiction 

   issues including standing, mootness, or ripeness problems 

   apparent from the face of the complaint

3. Any Rule 12(b)(5) service of process defects if 

   discernible from the complaint

4. For each argument identified: rate it Strong / Moderate / 

   Weak and explain in one sentence why

Do not fabricate case citations. Describe legal standards in 

general terms and flag them for my verification.

Jurisdiction: [Federal / State — specify]

Court: [District or circuit, or state court]

[PASTE COMPLAINT HERE, OR UPLOAD IT]

```

Any argument the model rates Strong still needs verification against your circuit’s current standard before it gets anywhere near a brief.

Usage Notes:

• Watch for: The model knows the federal 12(b)(6) standard well. State court pleading standards vary—it may misstate the applicable standard for your jurisdiction. Verify before you rely on it.

• Verify before use: Every legal standard the model flags as misapplied. Confirm the correct standard from primary sources before drafting anything.

• Time saved: Replaces 90 minutes of initial analysis. The real value is surfacing weak arguments fast so you can rule them out—and getting there in 10 minutes instead of an hour.

PRO TIP: Run this even when you’re confident there’s no motion to file. The model occasionally surfaces a 12(b)(1) standing issue or a service defect that isn’t obvious on a first read. Ten minutes. Worth it every time.

Step 3. Verify and decide

• Take every Strong or Moderate argument and confirm the governing standard from primary sources in your jurisdiction.

• The model hasn’t read your judge’s recent opinions on 12(b)(6) motions. If you know the judge disfavors them, that’s your call.

• If the output is clean—nothing rated above Weak—you have a documented analysis. File an answer and move on.

I ran this on a recent file of mine and it surfaced jurisdictional issues that we are still in the middle of motion practice on, six months later.

You spent 10 minutes on the triage. You know whether the motion is worth filing. That’s the only decision this prompt is for.

Verify the standards. Check your jurisdiction. Then decide.

—Lawyer.exe

Subscribe for more workflows like this.