Iowa Defense Counsel Association · 62nd Annual Meeting · September 11, 2026

Are Lawyers Obsolete?

Advances in Use of Artificial Intelligence in Defending Your Clients

What has to exist before AI is any use on a real file

Jennifer Case, J.D. · Law Tech AI

• Quick intro. I am standing between them and the coffee. • Jennifer Case. California attorney, former litigator. Eighteen years as a CEO before this. • Here because Frank Ramos got called to trial. Jon needed somebody on eleven days notice and Frank sent three names. • Credit Frank out loud.

Answering the title, first

No. And here are the three things I actually hear.

Objection one

“I don’t need it. Thirty years in.”

Objection two

“It makes things up.”

Objection three

“It forgets who I am.”

All three are fair, and all three are correct as stated. This hour is three answers.

$Jennifer’s dog, looking unconvinced$

• Answer the title in the first thirty seconds. No. • Nothing takes a deposition, reads a jury, carries a duty of loyalty, or answers to the Iowa Supreme Court. • Three things get said to me at every break, and all three are correct as stated. Put them up and agree with them. • One, I don't need it, I've been doing this thirty years. Two, it makes things up. Three, it forgets who I am. • Do not answer any of them here. Say the whole hour is three answers, in that order, and move. • Plant the fourth objection, the one nobody says out loud: should we be using this at all. Say your carriers already are, and you will come back to it with the court cases. [!] Never frame any objection as a mistake the room made. All three are fair. [DOG] Bottom right. Let it land, then move on. Do not explain it.

Objection one · stipulated

AI is a very good word generator. That is the least of it.

If you are already drafting with it, you know that, and it is genuinely impressive. It is also the only thing anybody has ever demonstrated to you, which is exactly why the objection exists.

So let me stipulate to the drafting and ask something else. Do you know what is happening in your cases right now?

• Do not open by telling them what AI cannot do. Open by agreeing with what it is very good at. • AI is an excellent word generator. Anyone in the room already drafting with it knows that, and it is genuinely impressive. Say so and mean it. • Then the observation that reframes everything: that is also the only thing anybody has ever demonstrated to you. Every AI demo you have sat through was a writing demo. Somebody typed a prompt and words came out. • So of course the reaction is I already do that, and I do it better. The objection is correct, and it is correct about the wrong subject. • Stipulate to the drafting entirely. It costs you nothing and it buys you the room. • Then move the question. All that expertise lives in one person's head, where nobody else on the matter can see it, and that person cannot see what anybody else is doing either. • Ask the question on the slide and then stop talking. Let it sit before you run the demo. [!] You are moving the subject, not winning an argument. Nothing here replaces their judgment and you should say so. • Transition: so let me show you the other half.

Objection one, answered · the half nobody mentions

Managing partners: right now, without calling anyone, what stage is every open matter at, who is on it, and what is due this week?

> /matter-update

Delgado v. Ridgeline Freight · plaintiff depo Tue 9:00

Vance Properties premises claim · site inspection Thu

Cortland Mfg product claim · expert disclosure Fri

Summit Medical Center · records affidavit due Wed

Harmon Realty dispute · demurrer hearing Fri 8:30

Ackerman adverse action claim · position statement out today

6 active matters · every deadline this week included

One saved command, written once and named · it reads the matter folders the firm already has

That is every matter, current. In the time it took me to ask the question.

• The reveal now runs about eighteen seconds, roughly one matter every two and a half seconds. Talk through them as they land. Do not wait for it to finish before you start. • If you need it again, arrow back one slide and forward again and it replays from the top. • Slow down here. This is the first thing on screen that looks like software, and half this room will decide in five seconds whether they are following you. • Tell them what they are looking at before you run it. I typed one question. It read the firm's own matter folders and wrote this back. • Explain the slash in one sentence and then move on: that is a saved command. Somebody wrote the instructions once, gave it a name, and now anyone at the firm types those two words. • That sentence is doing real work. Without it the room thinks you typed a clever question. With it they understand somebody built a reusable thing, which is the whole argument of this hour. • It also plants the playbook. When you get to the playbook slides, remind them they already watched one run. [!] Do not say terminal, script or command line. Saved command is enough, and this room has typed slash commands in Teams and Slack already. • Ask the question in the room first, out loud, to the managing partners. Right now, without calling anybody, what stage is every open matter at, who is on it, and what is due this week. • Let them sit in it. Most partners cannot answer, and that is not a criticism of how they run their firms. • The information exists. It is scattered across six people's heads and four systems, and the only way to assemble it today is a series of phone calls, which is why it mostly does not happen. [!] Say out loud that these are fictional matters. No real client files on this screen. [STORY] The managing partner walking every open case with a newer attorney, three hours at a stretch, who had reached the point of not being able to keep doing it. For a firm where supervision is an ethical duty rather than a management preference, that is the version that lands. • Point forward: by the end of the hour you will know exactly what had to exist underneath this, and it was never a product I sold anybody. • If it feels like it is landing flat, drop the mechanics entirely and just read the six lines out. The output is the point, not the box.

Objection two, answered · where it is completely true

Hallucination is a word-prediction problem

Nothing in front of it

“Find me cases holding X.” It produces what a citation looks like, because that is the most plausible thing to come next.

Your documents in front of it

“Here are the depositions. Identify every statement about when the roof was inspected.” Pinned to the text, and checkable.

I do not recommend it for legal research. If you use it there, you read every case, every time.

• Take this objection more seriously than the first one. The risk is real, it has a bar complaint attached, and the room already knows the sanctions stories. • Explain the mechanism, because almost nobody does and it is the thing that makes the rest of the hour make sense. • It is a word predictor. That is the whole machine. It produces the most plausible next words, over and over, very fast. • So it is not remembering a case and getting it wrong. It has nothing to remember. It is producing what a citation looks like, because that is the most plausible thing to come next. Perfect format, real-sounding reporter, court that exists, case that does not. • That is why the sanctioned lawyers got caught. The output was not a bad memory. It was a well-formed guess, and nobody opened the case. • Now the flip. Put the documents in front of it and the plausible next words are constrained by the text on the page. It can still be wrong, but it is wrong about something you can check, in a document you are holding. • The difference between the two cards is whether the material is actually in front of it. • Tell it what to produce. Identify every statement about when the roof was inspected. Summarize the holding of this case, with the case loaded. Build the chronology in this format. A named output is the whole difference. [!] Do not model a lazy prompt on stage. Never say here are four hundred pages, tell me what is in them. Vague ask, vague answer, even with the documents right there. Name the output you want. • Worth flagging the rhyme: it is the same lesson as the billing slide later. Say what came out of the work. Vague in, vague out, whether you are writing a time entry or writing a prompt. • Say this plainly and do not soften it: I do not recommend AI for legal research. If you use it there, you read every case, every time, and you check the citation like you always did. [!] Do not name a specific sanctioned lawyer or case from the podium unless you are certain of it. Say the sanctions cases you have all read about. • The line that carries the rest of the hour: everything I am about to show you is the second kind. Reading your own files. That is the category where this is reliable enough to build on. • Forward reference: I will show you a tool later that quotes three pages of the record back to you verbatim before it writes a single line, so you can confirm in ten seconds that it actually read the file. • The third objection, that it forgets who you are, is a different problem with a boring fix. That is the next few slides.

The word you have heard all conference

An agent is a set of files, and permission to act on them

What you are being sold

Agentic. Autonomous. It runs the practice while you sleep.

What is actually in the box

Files it reads. An account that decides what it can open. The model underneath is the same one everybody rents.

Ask a vendor two questions. What does it read, and what is it allowed to open?

• Name the word they have heard all conference. Agentic. Autonomous. Agents. • Nobody defines it, so it sounds like something you buy and switch on. • Deflate it: the model is the same one everybody else is renting. That part is a commodity and it gets better whether you do anything or not. • What makes an agent yours is two things. The files it reads, and what its account is allowed to reach. • That is the whole talk. The first half of this hour is the files. The second half is the permissions. • So when a vendor sells you an agent, ask the two questions: what does it read, and what can it open. [!] Do not overclaim. There is real engineering underneath. The point is that the part that is yours, and the part that decides whether it is any good or any danger, is the files and the access. • Hand off: so let me show you what those files actually are.

Objection three, answered · and the fix is boring

The infrastructure is a set of plain files

About the firm

Who you are, where you practice, the rules you follow.

Writing standards

How your documents sound, including what you refuse to say.

Playbooks

How your firm runs a task, the same way every time.

A knowledge base per matter

Posture, parties, the dispute, what is open. That is the one building on the right.

> reading about_the_firm.md 0 of 4 loaded

delgado_v_ridgeline_KB.md

• There is nothing to buy. Plain files describing how your practice works, read before the tool does anything. • Three sit above every matter: who you are and where you practice; how you write; your playbooks. • How you write changes output quality more than any prompt trick I know. [STORY] At nearly every firm the playbook has never been written down. It lives in one senior partner's head, and it walks out the door when they retire. Writing it down is worth something before any AI touches it. • Fourth: a knowledge base per matter. Posture, parties and counsel, what is in dispute, who holds the data, what is open and who owes it.

The two that matter most

One holds the case. One holds the standard.

01 · Case knowledge base

One per matter.

delgado_v_ridgeline_KB.md

Updated today · changes constantly

02 · Playbook

One per practice area.

trucking_defense_playbook.md

Unchanged since March · holds still

One moves with the case. The other one holds still on purpose.

• The two documents I build with every client. People collapse them together and it causes trouble. • Left, the case knowledge base. Matter specific, volatile by design. You touch it every time you work the file. • Right, the playbook. Firm level and stable. Opening steps, discovery standards, privilege review, production protocol, standing objections. • Bottom of every playbook: the list of things that always need an attorney. That is what makes it defensible to a client, a carrier, or a bar investigator. • When a vendor says their tool checks work against your playbook, this is the document they mean, and it has never been written down. • One moves constantly. The other holding still is the entire value.

Why this year

From one lawyer in a chat window to a team on one file

Single player

One chat history. It dies with the tab.

Multiplayer

One set of files. Every tool reads the same context.

Write it down this year. The next wave of tools will need it written.

• The one thing I would most want them to take away. • Two years of AI in law firms has been single player. One lawyer, one chat window. Close the tab and it is gone, and it never belonged to the firm. • Changing now, quickly. Multiplayer: the whole team against one shared set of files, everyone's assistant reading the same context. • The entry ticket is having written it down. Nothing else. • A firm with a context layer walks into every new tool ready. Without one it is a blank page again, now with six people instead of one. • The write-down compounds. The sooner it exists the more it compounds.

Why the location matters as much as the content

Plain text, in the shared drive the whole firm already opens

OneDrive, SharePoint, Dropbox, whatever you already pay for. Six people can work the same file on the same day.

• Where the files live matters as much as what is in them. This is where I would push back on a vendor. • Most firms buy one program and build the workflow around it, so institutional memory becomes a configuration inside somebody else's product. • They change pricing, get bought, or stop fitting your practice, and you cannot take it with you. • What I build instead is markdown. Plain text in the matter folder, in the system you already pay for, next to the documents it describes. • Four reasons: every AI tool reads it; a human reads it; it survives model and platform changes; six people can work the same file the same day. • Detail that looks cosmetic and is not: a header with date, initials, and a line saying this is the source of truth. So nobody mistakes it for a work product memo in six months.

The layer on top

Your team already has somewhere to reach the context

On your own machine

Reads a folder. The shared drive is the input.

In Slack

Tag it. It answers with firm context.

Ally, in our Teams

Tagged like a colleague. Silent until called.

Three front doors. Each one only as good as the files underneath.

• Three places, all live today. • Your own machine. Cowork, Codex and their equivalents read a folder, so the shared drive is the input and nothing gets uploaded anywhere new. • Slack. Tag the assistant in a channel and it answers in the thread with the firm's context behind it. [STORY] Ally, what I built for my own team in Teams. We tag her like a colleague and she answers with everything we have written down. [!] The design decision this room will care about: Ally does not sit and watch the channel. She answers when tagged and stays quiet otherwise. For a firm with privileged conversation moving through those channels all day, always-listening is a choice you make with your eyes open, never a default somebody flips on for you. • Common factor: none of them is smart on its own. Each is exactly as good as the files underneath it.

Before we go further

I just told you to point an agent at your matter files. So what governs the data?

• I have just spent fifteen minutes telling you to put matter files in a shared drive and point software at them. • Deal with the obvious objection before we go anywhere near the return. • There is a lot of confident nonsense about what protects that data, and some of it will cost you.

Five trial courts since early 2026 · three federal, two state

Privilege and work product turned on which tool you used

Where confidentiality failed

A consumer account, no attorney direction, under terms permitting training and disclosure to third parties.

What the court actually asked

Retention, training, and who else gets to see it.

You have had this argument about email and about the cloud. The answer was never stop. It was read the contract.

• Name the objection nobody says out loud: should we be using this at all. • Then the fact that reframes the room: your carriers already are. Look at the top of your own deduction reports. Those reason codes come out of an automated bill review system. The people cutting your bills are not reading them by hand. • You have had this argument twice before. Email in the nineties. The cloud about ten years ago. Same question every time: is client information safe on somebody else's server. • Both times the answer from the bar was the same. You may use it, with reasonable care, and you have to understand what you signed. • Five trial courts have now looked at AI and privilege specifically. Three federal district courts, two state courts, all since early 2026. This is the part the room has not heard. • Heppner, Southern District of New York, February 2026. A defendant got a federal grand jury subpoena and used a consumer version of an AI tool, on his own, with no direction from his attorneys. The consumer privacy terms let the company train on his data and disclose activity to third parties. The court found no reasonable expectation of confidentiality, which is the foundation the privilege claim rests on. Work product was the closer call. [!] Say clearly what Heppner does NOT hold. It does not say using AI destroys privilege. It holds that a consumer tool, used without attorney direction, under terms permitting training and disclosure, does not support a reasonable expectation of confidentiality. That distinction is the entire point of this slide. • The other four, Warner, Morgan, Tate and Assini, found work product protection preserved in various ways. So the courts are not hostile to the technology. They are looking at the tier and the terms. [!] Be precise. Heppner does not hold that using AI waives privilege. It holds that a consumer tool, used without attorney direction, under terms permitting training and disclosure, does not support a reasonable expectation of confidentiality. Overstating this is the one way to lose credibility on this slide. • The rule to hand them: you cannot put medical records, or any client confidential information, into a free or consumer tool. Your duty covers everything you obtain in litigation, not only what the client told you directly. [!] Cite only these five by name. They are in the white paper with the primary orders linked. Do not add cases from memory or from the podium. • Offer the white paper. It is a lead magnet and this is the room for it.

What the context layer cannot do for you

Teach the agent your policy. It still inherits your permissions.

You would never screen a lateral by asking them not to open the file. Do not screen an agent that way either.

• Every firm here has at least one matter only certain people should open. A lateral came in with a conflict, or a carrier requires named-user access as a panel condition. • You can teach the agent that rule, and you should. That is what the playbook is for. • Be precise about what it buys. An instruction is a request. It shapes what normally happens. It is not enforcement. • I have written exactly that into a client's own record: a soft control, shapes normal use, enforces nothing. • How you actually screen a lateral: you do not ask them politely to stay out. You take the access away. [STORY] Firm where the matter share was open read-write to all staff with no segmentation, while the firm carried a carrier restriction on paper limiting certain files to assigned staff. We wrote the rule into the tool as a stopgap and I said so in writing. • Real fix: the account the agent runs under cannot reach the folder. Coming in a few slides.

B A A · business associate agreement

HIPAA has a paperwork requirement. Plenty of firms never need it.

You need one when

Your client is a covered entity. Med-mal, hospital, clinic work.

Everyone else

No BAA. Rule 1.6 still covers every file you hold.

A floor under one kind of data. How far above it you sit is yours.

• Define the jargon first. B, A, A. Business associate agreement. Assume nobody in the room knows it. • A specific contract under HIPAA. The client is a covered entity, a hospital, clinic or health plan. If you handle their protected health information on their behalf you are their business associate. • Applies squarely to med-mal, hospital and clinic work, and some employment files. • If none of your clients are covered entities, you may not need one at all. [!] That does not soften confidentiality by one degree. Rule 1.6 has no interest in what your vendor signed. • A floor under one category of data. Where your firm sits above it is a judgment you make and write down. • I can tell you what vendors will sign. I cannot tell you how conservative your firm should be, and anyone with a single answer has not looked at your client list.

Verified September 2026 · re-check before you rely on it

All four will sign. Each draws the line somewhere different.

Microsoft

E3, E5, Business Premium. Web search sits outside.

OpenAI

Enterprise, yes. Business, despite the name, no.

Google

Gemini in Workspace, on certain plans.

Anthropic

Chat and projects. Cowork never.

They all sign. Ask for the covered list, then count what your people actually use.

• All four will sign. The differences are in what the signature covers, and they are not small. • Microsoft: Copilot folds into the agreement already covering Exchange, SharePoint and OneDrive, on E3, E5 or Business Premium. Your administrator knows which. Web search sits outside it. • Google: Gemini inside Gmail, Docs, Sheets, Slides and Meet on certain Workspace plans, once an admin signs in the console. • OpenAI, the trap worth naming: Enterprise is covered. Business, which is exactly what a small firm buys on the name, has no agreement at all. Two different products. • Anthropic: self-serve since July, no seat minimum, so a twenty-seat firm accepts from a settings page. Chat and projects covered. Cowork excluded permanently. • Close on this: do not ask whether your vendor signs. They all do. Ask for the covered list and count how many things your people actually use fall outside it. [!] Verified September 2026. Tell them to re-check before relying on it.

∞

The enforcement version

Split the work by folder, rather than by platform

The records folder

One assistant’s account can open it. The other one cannot.

Everything else

No health information, so it stays open to whichever program does that job best.

An agent cannot read what its account cannot open. That is a wall. An instruction is not.

• Callback: a few slides ago I said an instruction is a request. This is the enforcement version. • Because coverage lines differ by vendor, it is normal to run chronologies in one program and the rest of the practice in another. [STORY] A client doing exactly that. Records under active review sit in a locked site one vendor's tools can read: four named people, read-only, full logging. Everything else runs somewhere else with no path to the records at all. • The file system decides it, so there is nothing for the agent to disregard. • It only works because the knowledge files live in the document system rather than inside either vendor's product. • Draw the boundary at the folder. Every knowledge file with no health information in it stays available to whichever program does that job best.

Part four

That is the foundation, and the whole team can reach it. So what does a firm actually build on top of it?

• That is the foundation, and the whole team can reach it. • Now what firms actually build on top of it. • The drafting getting better is what everyone expects, and it is the smallest part of it. • What follows comes out of real engagements: billing, coverage work, medical records, intake, and one that is simply about making a better argument. • Different practices, different fee structures, same foundation underneath all of them. [DOG] Second photo. Mid-hour reset before the build stories.

Two inputs, and only one of them is published

The standards they cut you on never appear in the guidelines they gave you

Carrier billing guidelines

Published, and different at every carrier.

Your own billing records

What they actually cut. Where the unwritten rules live.

We searched all thirty-eight pages for the words they cut us on. Boilerplate, aggregate, nominal: zero hits.

• Take this away even if you forget everything else: there is no one-size-fits-all answer to carrier billing, and anyone selling you one has not looked at your bills. • Two inputs we analyze. The carrier's published billing guidelines, and the firm's own billing records. • Only one of those two is published. The other is the only place the real rules exist. • The proof: we took the reasons the carrier actually gave for cutting, and searched their own thirty-eight pages for those words. Boilerplate, zero hits. Aggregate, zero hits. Nominal, zero hits. Excessive appears once, in a section about unauthorized access to their computer systems, which is not this. • So they are cutting on standards that are real, applied consistently, and written down nowhere you can read. • You cannot read your way to those. The only place they exist is the pattern of what has already been cut on your own bills. • Six carriers, six different rule sets. The version we built for one does not transfer to the next. [!] Say guidelines, not Defense Counsel Manual. If pressed: State Farm Legal Services Program Defense Counsel Manual, April 2025, plus a separate State Farm Guidelines excerpt. [!] Only if somebody presses for a number: at that one firm those unwritten standards were 36.8 percent of the dollars cut. One client's books, so frame it that way and do not generalize it.

What we built · two gates

Read the pre-bill against that carrier’s rules before anyone sees it

Likely to pass

Nothing fires.

Worth a second look

Turns on facts only the biller knows.

Unlikely to be paid as written

This carrier has cut this pattern before.

It never tells you whether to bill. That stays yours.

• Two gates. Gate one is the billing manager running the whole batch and getting a workbook back, broken out by biller. • Gate two is the person who wrote the entry, checking it before billing ever sees it. Answer in about two seconds. • Three verdicts. Likely to pass, nothing fires. Worth a second look, it turns on facts only you know. Unlikely to be paid as written, meaning this carrier has cut this exact pattern before. • On the third it names the rule, asks what the time actually covered, and rewrites the description around your answer, in language that carrier has historically paid. [!] Be precise: it is not inventing work and it is not padding. Most of the time the work was completely real and the description undersold it by about six words. • It never tells you whether to bill. That is the firm's call and the lawyer's call, and we decided that deliberately. • Callback to the broken loop: the staff entering time never see the rejections, so nobody learns. Gate two closes that before the bill goes out instead of six weeks after.

Both attorney work · illustrative, from real deduction reports

The carrier wants to know what came out of the work

Zeroed · called clerical

“Review file for objections to subpoena.”

Not adjusted

“Review/analyze claim file, identify relevant parties and information, to support preparation of trial subpoenas.”

One describes a check. The other says what the check produced.

• The whole point in one sentence: the carrier wants to know what came out of the work. • Both of these are attorney work. Both start with the same two words. The difference is that one says what the review produced and the other does not. • Left describes a check. A lawyer looked at the file. Nothing came out of it on paper, so the carrier called it clerical and paid nothing. • Right names what the review turned up, and what it was for. Paid without adjustment. • Say the clerical part slowly. That is a lawyer's judgment about whether subpoenas can go out, and it got labelled secretarial work. [!] These two lines are illustrative, tightened from real deduction reports. Do not quote a dollar figure against them. • Your strongest fact if anyone challenges it: the real cut entry was longer than this and it named the third parties, and it was still zeroed. So detail is not what saved the other one. Naming the output is. • If somebody asks whether the verb matters, it does not. Review/analyze entries were cut about 21 percent of the time and draft/revise about 26 percent. Almost identical. • What separates them is naming a result. Among review/analyze entries, the ones naming an output or purpose were cut about 7 percent of the time. The ones describing only a check, about half. [!] Those figures come from the reports I could read, not all 125. Offer to re-run it on the full set rather than defending the number. [!] This is the slide for the half of the room who write these entries or supervise the people who do. Say that out loud. • Nobody was padding anything. They were being brief because they were busy, and brief reads as clerical. [!] Name the ethics point before anyone else does. There is a version of this that is just dressing up an entry and I will not do that. The tool asks what the time covered. It never invents the work and it never tells you whether to bill.

The writing-standards file

We read 33 of his coverage letters and wrote down how he writes

Three days a letter became six and a half hours. And he stopped discounting his own bills.

[STORY] A senior coverage partner. We read thirty-three of his own letters going back four years and wrote down how he actually writes. • The architecture was invariant across all thirty-three: conclusion first, always, in paragraph one. Policy language quoted verbatim before it is ever paraphrased. A three-part close every single time. • We captured the phrases he always reaches for and the ones he refuses. • An escalation vocabulary with real triggers. Untenable means a bad legal position. Bad faith means legally unsupportable and causing harm. Fraudulent only when the carrier actually misquoted the policy. He never jumps a level without the facts. • Best detail, and it tells you how specific this gets: three letters were flagged as not his, because they closed Sincerely yours instead of his invariant Very truly yours. They turned out to be a different attorney's. • Outcome: three days a letter, and he routinely discounted the hours before billing because he could not bring himself to charge a client three days for one letter. About six and a half hours now. He billed more last month than in any month of his career, at the same rate. He simply stopped discounting.

A self-insured client’s compliance portal

When a system has no way in, you build the bridge yourself

What the portal allows

Every narrative typed by hand. Nothing in, nothing out.

What we built

One shared file. The portal gets filled from it.

Nobody will build a connector for one client’s portal. You do not have to wait for one.

• The purest version of the argument I opened with. Spend a minute here. [STORY] A very large self-insured client, the kind several of you defend, runs a compliance portal the firm has to report into. • Every narrative typed by hand, screen by screen, cut and paste across tabs. No way to load data in, no way to get data out. • Their deduction report is not even a report. You click through four or five screens and the only thing you can keep is a screenshot with no case number on it. • That team was thirty to forty-five days behind on billing, and the reporting burden was the reason. • Nobody is going to build a connector for one client's portal. The market for that is one company. So we did the other thing. • The team keeps one shared file current, the same file the whole pod works from, and the portal entry gets driven off that rather than out of six people's heads. [!] Be straight about where it stands. The reading side works, we pulled the deduction patterns straight off those screenshots. The entry side is still being built, and the technique choice carries different confidentiality postures. Do not sell a finished answer.

What it refuses to do

A chronology in your format, and proof it read the file

Proof of reading

Quotes three pages back, verbatim, before writing a row.

Reconciled

Every page accounted for, or it rebuilds.

Contained

Four people, read-only, logged.

It records what the chart says. No causation. No liability. No opinion.

• By hand: fifty to a hundred record sets a month through nurse paralegals, roughly a minute a page, then the expert package, then all of it again when supplementals land. [!] Insert a measured throughput number only if you have one. Otherwise skip it and go straight to the controls. • Point them at the controls, not the speed. The speed is worthless if you cannot defend the output. • Proof of reading: before it writes a single row it quotes the first eight words of the first page, the middle page and the last page, verbatim. You can check that against the file in ten seconds. If it cannot read a page it stops and gives you nothing. • Reconciliation: every page either starts an entry or folds into one, and if the count does not balance it rebuilds rather than hand you a tidy number that is wrong. • Your format. Your header, your columns, Word or Excel, so nobody retypes it. • Containment is a folder, not a promise in a prompt. Four named people, read-only, full logging, and the agent runs with exactly those permissions and can reach nothing else. • The line to land: your file permissions are your AI security model. Most firms have never had to think that way, because until now nothing could read the whole share at once. [STORY] I have a healthcare client that prohibits AI review of its records outright. That is a completely legitimate answer, and the tool loses to the client's instruction every time. [!] Do not say Iowa law is stricter. Unverified. Talk about client and carrier requirements instead.

Three other firms

What else it has paid for

Captured time

3.7 hours recovered from three days’ work.

Intake

Six questions answered up front. Turnaround halved.

Novel theories

Seven to eight, from the record and the law.

Different practices. One context layer underneath all of it.

• Three more, quickly. • Captured time. A firm surveyed its own people and half said they were probably losing three to five hours a week to email they never wrote down. One attorney ran the tool and pulled 3.7 hours out of eight matters from his previous three days. Work already done, about to be lost. [STORY] Intake, and this is the one I would steal if I were you. At a multi-practice firm we read months of intake email to find what partners actually ask before approving a new matter. It was the same six things every time: conflicts, amount in controversy, equitable relief or money, urgency, capacity, and what actually happened. • The intake report answered none of them, so every new matter took at least one round trip and often three, purely to gather things knowable on day one. We rebuilt it to answer all six before it reaches a partner, and a non-lawyer runs it. Turnaround down about half. • Novel theories. A business litigation team certain their client was right on fairness grounds and unable to find the path to argue it. From the record and the governing law we generated seven or eight theories to evaluate. Some were bad. That is what evaluation is for.

The whole argument

Your firm already has the data. The work is reaching it.

01Visibility across every matter.

02A summary a partner can act on.

03Time captured where it was earned.

04Every entry checked against that carrier’s rules.

Measure before you buy.

Law Tech AI · IDCA 2026

• If they photograph one slide, this is the one. Pause on it. • Your firm already has the data. Every matter file, every brief you have filed, every bill you have sent, every deduction report sitting in a folder nobody opens. • The problem has never been that you lack information. Nothing could reach it, so it may as well not exist. • Building the infrastructure is what makes it reachable, and it is plain files in the document system you already own. • Then the four. Visibility, so you can see across every matter without calling six people. Summary, so a file nobody has time to read becomes something a partner can act on before lunch. Accurate billing records, so time gets captured where it was earned instead of evaporating at six in the evening. Carrier compliance, so every entry is read against that carrier's rules and your own history before it goes out. • The order matters, and this is where firms lose money. Measure first: interview each group, follow one matter end to end, and count every place a human retypes something a machine already has. That is where the money hides. • Skip that and start buying licenses and you will solve a problem that was never costing you anything. • Hand off to the four questions.

Step one, the version you can run yourself

Four questions, before you buy anything

What admin work are you doing that you are not getting paid for?

What is the thing you personally dread doing every single week?

Where are you entering the same information in more than one place?

What could you learn from your own data that nobody has ever analyzed?

Notice none of them is “can AI write this email.” Measure before you build, and before you buy.

• Every engagement starts with a much longer version of this across the whole practice. Four they can ask themselves this afternoon. • What admin work are you doing that you are not even getting paid for? The work that never reaches an invoice at all, which is a bigger category than the work you discount. • What is the thing you personally dread every week? Dread is a reliable signal that something is repetitive and badly supported. • Where are you entering the same information in more than one place? Matter number into the practice management system, the billing system, the calendar, the file name, then the client portal. Four systems, one fact, four chances to get it wrong. • What could you learn from your own data that nobody has ever analyzed? Sent mail carries your writing style. Deduction reports carry what a client keeps cutting. Billing entries carry how your people describe their work, which is what you train a time-capture tool on. • All of it already exists, and nobody has read it in bulk because until recently nobody could. • Land it: notice none of the four is can AI write this email.

None of this requires buying anything

Four things you can start Monday

01Draft the knowledge base on one matter. Correct it.

02Five documents of one type. Draft the playbook.

03Open your matter drive. See who can reach what.

04Give it data you already have. Ask for the pattern.

Law Tech AI · IDCA 2026

• Four things, none of which require buying anything. • One note first, because it decides how they do these. If your tool reaches your files directly, Cowork or Codex on your own machine, or Copilot and Gemini inside your document system, you point it at a folder. If all you have is a chat window, you upload the documents. All four work either way. • One. Take the matter folder you know best and have it draft the case knowledge base. Posture, parties and counsel, what is in dispute, what is open. It will get maybe eighty percent right, and correcting eighty percent takes twenty minutes where a blank page takes an afternoon you do not have. • Two. Pull five documents of one type you have already filed and are happy with, and ask for the rules that produced them. Correct that too. • Three. Open your matter drive and look at who can actually reach what, going by the permissions rather than the policy. • Four. Give it a pile of your own data and ask what the pattern is. Deduction reports, a year of sent letters, your time entries. Pick whichever pile somebody already complains about.

Back to the question in the title

AI makes you a better lawyer

01You see every matter and what is due.

02The newest person works to your standard from day one.

03Whoever picks up a matter starts current.

04Your time goes to judgment.

It is here, your clients are using it, and so are your competitors. Choose the right tool and it makes you a better lawyer.

• Back to the question in the title one last time. Nothing I have shown you replaces a lawyer. • Be concrete about how it makes the lawyer better, because that word gets thrown around. • You can answer what is happening across every open matter and what is due this week without calling four people first. • The newest person on the file works to the standard you set, rather than to whichever old brief they happened to find on the server. • Anybody who picks up a matter starts current, so two days of reading yourself in stops being the price of covering for a colleague. • Your own time goes to the part only you can do. Reading yourself into a file, hunting for the right form, working out what moved while you were in trial: none of that was ever the practice of law. It was the tax on it. • Leave them with this: every one of those came from a firm writing down how it works and putting it where the software can read it, not from a smarter model. • The models keep improving whether or not you do anything, and that part is free. The writing down is the part that is yours. • Close on this, in your own words: AI is here to stay. Your clients are using it. Your competitors are using it. Choose the right tool and there is no reason not to use it to be a better lawyer. • Callback to the terms-of-service slide. Right tool means the one whose contract you have actually read.

Thank you

Questions, and where to find me

Reach out through the code and I will send you these slides, along with an example knowledge base and an example playbook you can adapt to your own matters.

QR code to the Law Tech AI community page

Scan it, or type itmailchi.mp/law-tech.ai/webinar-qr

Jennifer Case, J.D. · Law Tech AI

LinkedIn: linkedin.com/in/jennifer-case-1b5486b

law-tech.ai · jennifer@law-tech.ai

• I am around after this and here through lunch, so come find me. • The code goes to one page: the weekly newsletter, the monthly AI mastermind for lawyers, and a link to book time to talk about their own firm. • The address is typed underneath because you cannot scan a screen with the phone you are already holding up to it. • Offer: reach out through the code and I will send these slides, plus an example knowledge base and an example playbook they can strip down and use on their own matters. • Thank them for having me, and for being awake at eight in the morning on day two.