# How to become an AI-native personal injury law firm in 2026 > How to re-architect a contingency PI firm around AI: demand generation, practice-area pods, case data, and the cited medical-record layer. 15 chapters. Canonical page: https://medrecords.ai/guides/ai-native-personal-injury-law-firm/ --- New — **Missing Records Detection:** flags every visit, provider, and date missing from the file. [See how →](https://medrecords.ai/product/missing-records-identification/) Operating guide, September 2026. 15 chapters. ## How to become an AI-native personal injury law firm in 2026 An operating manual for personal injury firms. It takes the 3-person law firm, breaks it into the parts AI changes, and rebuilds it around outcomes: demand generation, practice-area pods, case data, legal knowledge, and the medical-record layer every injury case runs on. By [Ahmed Jemaa](https://medrecords.ai/authors/ahmed-jemaa/), Co-Founder & CEO of Medrecords AI · Published 7 Sep 2026 By [Ahmed Jemaa](https://medrecords.ai/authors/ahmed-jemaa/), Medrecords AI. Updated 7 September 2026. How to re-architect a contingency firm so the book of business, the case, and the data belong to the firm instead of to the rainmaker. The architecture is adapted from Chia Jeng Yang's Foremark Legal essay, *Architecture for the Outcome-Based, AI-Native Law Firm*. The personal injury specifics, the record layer, the adoption data, and the build sequence are ours. Industry figures are attributed to the survey or vendor that published them, and no number is presented as our own research unless it is. 10 numbers ### The adoption gap in 10 numbers The plaintiff bar is not short on AI curiosity. It is short on AI it can verify. These figures are from a 2026 industry study of 207 U.S. personal injury legal professionals. 8 in 10 firms have engaged with AI in some form n = 207, March 2026 3 in 10 use AI as a regular part of how they work the adoption gap, in 1 line 99% of attorneys will not use AI output they cannot verify the hard constraint 79% reject fully autonomous AI and insist on human review human in the loop 75% of firms are at or beyond operational capacity why the gap costs money 57% need a week or more to produce a completed demand letter a data assembly problem 62% of pre-litigation attorneys already use AI by role 47% of pre-litigation paralegals already use AI by role 71% want accurate medical record summaries and traceable sources tied as the top asks 73% are optimistic or excited about AI fewer than 1 in 20 are skeptical **The same 10 figures, on one scale** — n = 207 — Industry study of U.S. personal injury legal professionals, March 2026. "8 in 10" and "3 in 10" are plotted as 80% and 30%. The 2 highlighted bars are the constraint. Verification is not a preference in this market; it is the price of entry. 8 of 10 — 80% Firms that have engaged with AI in some form. 3 of 10 — 30% Firms that use AI as a regular part of how they work. The gap is not awareness. It is trust, and it is also a revenue problem. A demand letter that takes weeks is not a drafting problem. It is a data assembly problem. 9 moves ### The operating system in 9 moves FromFinder ToA firm-owned demand engine FromMinder ToA practice-area pod with its own P&L FromGrinder ToThe strategic thought partner FromHourly billing ToOutcome-based pricing FromManual record work ToA cited record layer FromOne rainmaker's book ToFirm-owned intake FromApprenticeship by luck ToA living knowledge base FromGut settlement calls ToCase-data underwriting FromFirm management as a tax ToThe highest-credited work **In this guide.** Deciding whether to restructure? Read chapters 1 and 2. Running a firm and want the build order? Start at chapter 11. Building the record layer? Chapters 8 through 10. Choosing a vendor? Chapter 12. **Chapter tags.** Deciding if you are still weighing it. Building a practice if you already run one. Everyone for the rest. Chapter 1 Deciding ### The 3-person firm is the problem For a century, the law firm has been built on 3 personas. The finder wins the business, the minder manages the client and runs the matter, and the grinder produces the underlying legal work. Most partnership fights, compensation disputes, and failed mergers are downstream of the tension between these 3 functions. Personal injury makes that tension worse. The work is contingency, so the firm is already betting its cash on the outcome. The files are document-heavy, so the grinder's time is consumed by records. The clients arrive through marketing and intake, so the finder is often the person who controls the advertising budget. When the 3 roles fight, the case file is what suffers. #### Why the clock is tightening now 100,000+ — lawsuits filed in the weeks before Florida's H.B. 837 took effect 90 days — the safe harbor the law gave insurers 5 — things it changed at once: deadlines, evidence standards, one-way fee exposure, fault limits, and the safe harbor **Florida's H.B. 837 is the clearest warning.** It shortened deadlines, tightened evidence standards, removed one-way fee exposure, added stricter fault limits, and gave insurers a 90-day safe harbor. More than 100,000 lawsuits were filed in the weeks before it took effect. Compressed timelines turn a slow manual record review from a cost into a case-killer. An AI-native firm is not a traditional firm that bought AI tools. It is a firm designed from the ground up around new capabilities, the way Amazon was not just a bookstore with a website. #### The 3 personas, and what breaks at volume ##### Finder What it does in a PI firmOriginates cases through advertising, referrals, and personal networksWhat breaks at volumeThe book of business belongs to one person, and leaves when that person leaves ##### Minder What it does in a PI firmRuns deadlines, status calls, discovery, and settlement timingWhat breaks at volumeQuality depends on the one lawyer or case manager minding the file ##### Grinder What it does in a PI firmReviews records, writes demands, drafts pleadings, prepares expertsWhat breaks at volumeThe hours scale with the pages, not with the firm If you remember one thing An AI-native PI firm is not a traditional firm plus software. It is re-architected so the firm owns the acquisition engine, the case workflow, and the outcome data. Chapter 2 Deciding ### What AI-native actually means Buying a chatbot for demand letters is not AI-native. It is the old firm with a faster typewriter. AI-native changes the form of the 3 functions, and because you cannot bill an agent's hour, it changes the economics with them. #### The 3 functions, rebuilt Old form Finder — A rainmaker with a personal book Minder — One lawyer carrying the matter in their head Grinder — Manual production work AI-native form Finder — A firm-owned demand-generation engine Minder — A practice-area pod with its own P&L, engineering, and business operations Grinder — A strategic thought partner whose production work is absorbed by machines The firm shifts from selling time to underwriting outcomes. It evaluates a case up front the way an insurer evaluates a policy and ties its fee to the result instead of the clock. Personal injury is the natural first home for this shift. The fee is already contingent. The missing piece has always been the firm's own outcome data, because nobody kept it well enough to underwrite with it. #### The loop itself changes Most lawyers calibrate on a chatbot they tried once: type a question, read an answer, decide the next step yourself. That loop makes one person faster and leaves the firm exactly as it was. The newer loop is different. You give a system a goal, the context, the tools, and the constraints, and it plans, works, checks its own output, recovers from its own mistakes, and keeps going without a human turn between every step. ##### The assistant loop Who decides the next stepThe lawyer, every timeWhat starts the workA human promptUnit of work1 answerWho checks the outputThe lawyer, after the factWhat it changesHow fast one person works ##### The agent loop Who decides the next stepThe system, inside constraints the firm setsWhat starts the workAn event in the file: records land, a deadline moves, a gap opensUnit of workA task run end to endWho checks the outputThe system checks itself, then a lawyer signsWhat it changesHow the firm is organized That is the gap sitting behind the adoption numbers: 3 in 10 lawyers use AI as a regular part of how they work, and a good share of the other 7 are judging a category they last looked at 2 model generations ago. The distance between typing a question into a chat box and handing an agent a matter is not a difference of degree. #### Where the change shows up first 1 — ##### Intake becomes a product, not a phone call A claimant's first interaction is a form, a chat, a callback, and a status page, not a receptionist taking a message. 2 — ##### Each practice area runs as its own unit Its own economics, its own intake rules, its own experts and settlement benchmarks. 3 — ##### The medical record is where AI takes work off the lawyer's desk first It is the largest, most mechanical, most deadline-bound pile in the file. 4 — ##### The firm keeps a commercial record and an intellectual record Both compound with volume. Chapter 7 is how they differ and why you build both. If you remember one thing AI-native is a re-architecture, not an add-on. The firm changes what it sells, who does the work, and what it owns. Chapter 3 Everyone ### The finder becomes a demand-generation engine In the traditional firm, business arrives through the charisma and networks of individual rainmakers, which makes the finder the most powerful and expensive person in a firm. In the AI-native firm, the finder is a consumer growth operator or a B2B sales talent. Client acquisition becomes a repeatable operating function, and the book of business belongs to the platform rather than to any partner. For a PI firm, that means the intake engine is the product. A claimant's first interaction is a form, a chat, a callback, and a transparent status page. The firm qualifies cases before an attorney spends a billable hour on them. #### The demand engine in personal injury | Channel | Traditional motion | AI-native motion | | --- | --- | --- | | Search and local services | Buy leads and hope | Own landing pages, review velocity, and a tracked intake form | | Referrals | Wait for past clients and providers | Systematized referral nurture with a known source and follow-up cadence | | Intake | A receptionist takes a message | Product-led intake that screens, schedules, and routes by case type | | Case status | A client calls to ask what happened | Proactive status updates and self-service | Ryan Walker, ex-CTO of Casetext and CEO of General Legal, puts the constraint on the commercial rather than the technical side: Traditional law firms generally have no GTM function, and have the expectation that business arrives inbound on reputation, with none of the machinery every tech startup built to scale a sales process. Replicating the sales motions tech startups already run, and applying them to a traditional industry like the legal industry, is the needle-mover. Rainmaker-only economics are deprioritized because the book of business belongs to the platform. Retention inverts for the same reason: a departing lawyer walks away from a client acquisition engine, case evaluation models, and accumulated outcome data that are difficult to replicate solo. If you remember one thing Stop renting the book of business from a rainmaker. Build a firm-owned intake engine that qualifies a case before an attorney touches it. Chapter 4 Everyone ### The minder becomes the pod The minder is the workhorse of a traditional firm. In personal injury, this is the lawyer or case manager keeping deadlines from slipping, deciding when to push and when to settle, and keeping the client from calling in panic. In the traditional firm, all of that lives in one person's head. In the AI-native firm, that craft leaves the lawyer's head and gets embedded in the machinery. Each practice area operates as its own decentralized pod with a practice lead, dedicated engineers, and a dedicated business operating officer, and each pod carries its own profit and loss on top of a shared platform. #### Traditional firm vs pod Traditional firm Owner of the case workflow — One lawyer or case manager Operations — Firm-level COO, overhead nobody owns Quality signal — Word of mouth after something goes wrong Client experience — Unmeasured AI-native pod Owner of the case workflow — A system the pod builds and improves Operations — Business ops lead inside the pod Quality signal — Observable per pod, per lawyer, per case type Client experience — NPS measured per matter Practice areas differ too much for a firm-level executive to do anything but firm-level ops. An auto-accident pod and a premises-liability pod need different intake rules, different experts, different settlement benchmarks, and different follow-up cadences. The AI-native firm measures client NPS per matter, like every other consumer industry. A client who wins a settlement but never knew what was happening can still leave a detractor score, and that score is now a management signal. If you remember one thing Each practice area becomes a pod with its own P&L, its own engineers, and its own business ops lead. Quality becomes observable per pod, per lawyer, and per case type. Chapter 5 Everyone ### The grinder becomes the strategic thought partner Nobody applies to law school dreaming of document review. They apply to become the person in the room when the case is being decided. The AI-native firm builds its technology around institutionalizing that thought leader, and calls it the 100x lawyer. The grinder's production work is absorbed by machines. The lawyer moves up to developing the legal theories that open new case types. In personal injury, a validated theory might be a new angle on rideshare liability, a premises duty argument, or a pattern of delayed surgery that turns a small injury into a damages case. #### The grinder, before and after ##### Old work Production. The hours scale with the pages. - Read thousands of medical pages. - Draft the demand from scratch. - Research a week's worth of case law alone. - Handle routine deadlines. ##### New work Judgment. The hours scale with the firm. - Verify a cited chronology and direct the inquiry. - Select the theory, negotiation posture, and expert strategy. - Test a hypothesis against the whole corpus at once. - Step into the moments that do not resolve to pattern matching. Research becomes quantitative. Agents sweep case law, dockets, verdicts, and the firm's own outcome data so a hypothesis is tested against the whole corpus at once rather than against whatever an associate could read in a week. The lawyer directs inquiry instead of performing it. That theory-development function has direct commercial consequence: a validated theory opens a new case type, and a case type at sufficient volume seeds a new pod. #### The instruction is where the judgment lives Most lawyers who try AI type "summarize these records," get back something mediocre, and conclude the category is a toy. The failure is the instruction. Templates were never the scarce thing. Every firm in a practice area has the same demand letter shell. What separates a good lawyer from an average one is what they do with it: which entry they pull forward, which gap they refuse to concede, how they frame damages for this venue and this adjuster. The vague ask Prompt — Summarize these medical records You get — A readable narrative you have to re-verify from page one Net time — Roughly zero, once you check it The specified ask Prompt — Build a chronology of the 90 days around the collision. Cite every entry to its source page. Flag treatment gaps over 14 days, any prior complaint to the same body part, and every provider named in the bills but absent from the records. Order by date, not by producing party. You get — A cited chart you verify by spot-check, plus a named list of what is missing Net time — The sorting goes; the deciding stays Write that second instruction down in enough detail and it stops being a prompt. It becomes one lawyer's judgment, encoded, running on every matter in the pod instead of only the ones that lawyer personally touches. This is the part that matters for firm management: knowledge that used to require years of sitting near the right partner becomes a first draft that starts from a much higher baseline. The review is still an attorney's. If you remember one thing The grinder's highest-leverage work is developing a validated legal theory that opens a new case type. Production work moves to machines; judgment moves up. Chapter 6 Building a practice ### The people model that results Partnership still comes through apprenticeship, because the old pipeline is dead. The associate model trained lawyers by burying them in production work, and that has not diminished significantly as laterals are prioritized in the AI age. The AI-native apprenticeship runs on a living knowledge base. Agents capture the firm's emails and case correspondence and structure them into institutional memory. A young lawyer queries the knowledge base the way a prior generation asked a senior attorney, and gets answers drawn from how the firm's best lawyers actually handled the situation. #### Intelligence replaces hierarchy A law firm pyramid is an information-routing structure. It exists because legal judgment was scarce and had to be rationed across matters, so the base of the pyramid carried throughput: organizing records, first-pass research, drafting from a form, chasing status. Information moved through people, and the layers were how it moved. Delegate the throughput layer and the shape stops paying for itself. Every lawyer is then valued for judgment rather than output, which changes staffing, apprenticeship, pricing, and the way the firm works with clients at the same time. FromLeverage counted in bodies per partner ToLeverage counted in judgment calls per lawyer FromJuniors hired for throughput ToJuniors hired for judgment, client relationships, and supervising machine output FromThe constraint is production ToThe constraint is review, prioritization, and deciding what to work on That has 2 consequences. The bar for when an associate hire makes economic sense has moved, and the job you are hiring for has changed with it. And the bottleneck moves rather than disappearing: a firm running this way can generate more finished work than its old coordination structure can absorb, which is a real problem and a better one to have. #### 3 stages, as Eddie Nasser frames them Eddie Nasser, Head of AI at Keller Postman and ex-Head of AI at Crosby, frames the progression in 3 stages. Stage 1 — **Single-player querying** — A lawyer asks an AI system for help. Stage 2 — **Multiplayer querying** — The system draws on the collective context of many lawyers across the firm. Stage 3 — **Multiplayer proactive suggestions** — Agents understand what the team is doing and surface the next action themselves. #### Compensation follows the platform Traditional firm Pay — Hours-based, with some origination Billing — Hourly, tracked Equity — Annual profit distribution, little retained earnings Management — A tax the best lawyers avoid AI-native firm Pay — Flat salary plus bonus plus equity Billing — No hourly billing Equity — Every lawyer, junior and senior, holds equity in the management services organization Management — The highest-credited work in the firm Most industries, except manufacturing floors and lawyers, do not track hours worked internally as a measure of productivity. Motivated people work hard because they are motivated and incentivized well, not because their hours are tracked. Traditional partnerships are annual profit-distribution machines with little retained earnings, so no partner has a reason to invest in anything that pays off after retirement. Equity makes long-term investment rational and justifies the management credit. If you remember one thing Make building the firm the best job, not a tax. Equity makes long-term investment rational; management credit makes the systems worth building. Chapter 7 Building a practice ### The 2 data infrastructures The shared platform rests on 2 data infrastructures, and they collect different data for different reasons. ##### Case Data Infrastructure Also calledThe firm's **Commercial Record** StoresSettlement values, case outcomes, counterparty behavior, and the negotiation patterns of insurersAnswersWhat is a case worth, and when to settleCompounding effectBetter underwriting and intakeHuman boundaryA lawyer signs the case evaluationThe questions it answers in PI - Based on our historical case data, if a client has this injury profile in this venue, what is the typical range of settlements offered? - When we deal with this insurer, which elements are they typically more sensitive to than others? - Which case types close fastest, and which produce the best return on the cost to acquire them? ##### Legal Knowledge Infrastructure Also calledThe firm's **Intellectual Record** StoresThe research, briefs, arguments, and case correspondence in which the firm's lawyers actually reasonAnswersHow to win, and how the firm reasonsCompounding effectBetter lawyers and reusable judgmentHuman boundaryA lawyer owns the argumentThe questions it answers in PI - What fact patterns and arguments survived a given insurer's motion practice? - For this injury type, what are the different ways we have tied the job or the mechanism to the damages? - Does this judge favor one damages framing over another? The firm collects the commercial record because underwriting requires it. An insurer cannot price risk without actuarial history, and a contingency firm cannot decide which cases to take, what they are worth, or when to settle without an outcome history of its own. AI can make negotiation memory institutional: lawyers can see which positions were taken, where counterparties moved, and what tradeoffs worked across prior deals. It collects the intellectual record because, in the Foremark framework, judgment has always been its scarcest asset, and the one it could never store. #### Why the record layer is the hard part Software teams reorganized around agents first, and not because engineers are early adopters. Their work already sits inside a closed, machine-readable loop: the instructions are digital, the tools are digital, the environment is digital, and another machine can test whether the output is correct. A personal injury file is the opposite. Fax artifacts, handwriting, duplicate productions, bills that do not reconcile to the treatment they claim, and no automatic way to check whether a summary is true. That is why the medical record is where an AI-native PI firm either compounds or stalls. Closing the loop means every generated line traces to a source page, so a person can verify it in seconds and the system can say when it cannot. Once the loop closes, the scarce resource stops being production and becomes context: how much of a matter, and of the firm's history with matters like it, you can put in front of the people and systems working on it. Both records exist to answer that. If you remember one thing Build both records on purpose. Case data tells you what a case is worth; legal knowledge tells you how to win it. Neither decides for you. Chapter 8 Building a practice ### Outcome-based economics and underwriting Personal injury already works on outcomes. The AI-native move is to stop treating that as a given and start underwriting it like a business. The firm evaluates a case up front the way an insurer evaluates a policy. It uses its own outcome history to decide which cases to take, what they are worth, and when to settle. It ties the fee to the result instead of the clock, and it prices flat-fee work where the client wins and the scope is clear. The point is not that every task gets shorter. Some do. The more durable change is that a given hour of attorney time carries more: wider issue spotting, a larger comparable set, a record that was actually read instead of skimmed. Price the outcome, because the hour has stopped being a stable unit of anything. #### Pricing models for an AI-native PI firm | Model | What it looks like | When it fits | | --- | --- | --- | | Contingency | A percentage of the recovery, with the firm carrying the cost of the work | The standard PI case, now underwritten with the firm's own data | | Flat fee by stage | A fixed price for intake, investigation, or demand preparation | Clear scope, or where the client wants certainty | | Hybrid | A lower contingency with a flat fee for a defined record or demand step | Cases where the medical work is the cost driver | | Portfolio underwriting | The firm treats a case type as a book, not 1 file at a time | Volume case types with enough outcome history | **The boundary matters.** AI can organize evidence, surface comparable outcomes, and draft the evaluation. A licensed lawyer signs it. No model decides merit, settlement value, or strategy on its own. The firm collects this data because underwriting requires it, and it is the same data that makes the acquisition engine smarter over time. If you remember one thing The fee is already outcome-based in PI. The AI-native move is to make intake, valuation, and settlement data-driven, with a human signing every call. Chapter 9 Building a practice ### The AI-native case file The case file is where a PI firm loses the most time and leaves the most value on the table. Split it into 2 columns: attention and judgment. ##### Attention The sorting tax. Software can take this column, but only where every line cites its source page. - Records out of order. - Duplicates. - Handwriting. - Missing pages. - Billing reconciliation. - The chronology a paralegal builds before anyone can argue the case. ##### Judgment This column stays human. It is what a lawyer is licensed to answer. - Liability. - Causation. - Damages. - Negotiation. - The settlement call. #### Where the hours go These are the manual tasks an industry buyer's guide lists as the quiet time sinks, with its published time estimates. **Manual hours per case, by task** — 6 tasks — Industry buyer's guide estimates. The faded bar is open-ended: the guide states 4 hours or more, with no upper bound. Nothing here is legal judgment. Every bar is the sorting tax, and it is the part of the file that scales with pages instead of with the firm. Manual task — Hours — Basis Building medical chronologies — 6 to 10 — stated, vendor estimate Reviewing scanned documents for key treatments — 3 to 5 — stated Rebuilding case knowledge after months of stagnation — 4 or more — stated, open-ended Drafting demands from scratch — 3 to 8 — stated Identifying gaps in treatment or records — 2 to 4 — stated Building deposition outlines and case strategy prep — 3 to 6 — stated 6 tasks, 1 case — 21 to 37 — our sum of the ranges above; row 3 has no upper bound, so the high end is a floor Source: Industry buyer's guide estimates. The total is arithmetic on those published ranges, not a separate measurement. #### The file as a live model of the case Treat a matter as a self-contained world: its records, bills, correspondence, research, deadlines, and the firm's history with cases like it. Once that world is machine-readable, work stops waiting for someone to remember to ask. The file changes, and the change is what starts the work. 1. **New records arrive.** They are sorted, deduplicated, cited, and merged into the existing chronology, and a human reads the difference rather than the whole thing again. 2. **A treatment gap opens.** It surfaces with its dates, before the defense builds an argument on it. 3. **Bills stop reconciling to the records.** The specials are flagged against the treatment they claim. 4. **A provider appears in the bills but not the records.** The missing production is named now, not discovered at the demand stage. 5. **A limitations or notice date approaches.** The file escalates itself instead of waiting for a calendar review. 6. **A comparable case settles.** The valuation range for open cases with that profile moves. None of those is a decision. Each one is a lawyer being handed a decision on the day it matters, rather than the week somebody happened to open the file. #### The same case, 2 ways | Step | Attention | Judgment | | --- | --- | --- | | Medical records arrive | Sort, deduplicate, OCR, flag handwriting | Decide what the record proves | | Chronology of the critical window | Draft every entry with a page cite | Verify the entries and choose the window | | Missing records | Rebuild what should exist from what was produced | Decide why the gap matters | | Demand package | Assemble the cited record and draft the facts | Write the demand, pick the number, set the posture | | Settlement | Surface comparable outcomes | Accept, reject, or negotiate | ##### Case economics - The money story — Billed charges, payments, liens, write-offs, and the specials that belong in the demand - What AI does — Assembles a billing ledger and flags inconsistencies - What stays human — A person owns the number and the negotiation ##### Case signals - What the record hides — An unfulfilled specialist referral, a treatment gap, a future-care recommendation, a credibility issue the defense will find - Why surface it — Found before the defense finds it, each one becomes preparation instead of a surprise The publisher's software handles the attention column: it sorts, deduplicates, reads handwriting and imaging, and returns a chronology in which every line cites its source page, under a signed business associate agreement. It does not decide liability, causation, damages, or settlement value. Those are the questions a lawyer is licensed to answer. **Disclosure.** This manual is published by [Medrecords AI](https://medrecords.ai/), which makes software for the attention column in a PI firm. It is priced per deduplicated page, with duplicates free, and you can watch it run on a case like yours in a 30-minute demo. Chapter 14 is the publisher's offer, labeled as such. Nothing else in this manual depends on using it. If you remember one thing Give the machine the attention column only where every line cites its source. Keep liability, causation, damages, and the settlement call human. Chapter 10 Building a practice ### The AI-native case flywheel The firm that runs the same loop on every case compounds faster than the firm that treats each case as a one-off. All 7 steps feed each other. **The AI-native case flywheel** — Step 7 feeds step 1: the reasoning captured on the case that just closed is what makes the next intake decision sharper. Hover or tap a node. 1. 1 — **Intake** — Qualifies the case and routes it by type 2. 2 — **Record layer** — Ingests, sorts, deduplicates, and cites every page 3. 3 — **Case signals** — Surface gaps, referrals, billing issues, and credibility risks 4. 4 — **Demand** — Drafted from a verified chronology and a reconciled ledger 5. 5 — **Settlement** — Negotiated with comparable outcomes from the firm's own history 6. 6 — **Outcome data** — Records what was offered, accepted, and rejected 7. 7 — **Knowledge** — Captures the reasoning, and feeds step 1 The loop is the point. Every settled case makes intake smarter, valuation sharper, and the next demand stronger. If you remember one thing The flywheel, not any single tool, is the defensible asset. Every case feeds the next one. Chapter 11 Building a practice ### The 90-day build Do not convert the whole firm at once. Convert one practice area into a pod, prove the economics, then copy the pod. **The 90-day build** — 3 phases — One practice area, not the firm. Each phase ends in an artefact you can put on a table, not a status update. The order matters. You cannot instrument a pod whose hours you never counted, and you cannot count the hours until you have picked one practice area to count. ##### Phase 1: days 1 to 30, map the current machine | Action | Owner | Success metric | | --- | --- | --- | | Pick one practice area with enough volume and a known cost driver | Managing partner | One named pod | | Time the last 10 files from intake to settlement | Practice lead | Attention vs judgment hours, in writing | | List the data the firm already has and the data it does not | Business ops lead | A 1-page data inventory | | Choose 1 record-review platform that signs a BAA and cites pages | Practice lead | A tested file, not a sales call | ##### Phase 2: days 31 to 60, build the record layer and intake | Action | Owner | Success metric | | --- | --- | --- | | Run the same 10 files through the record layer | Paralegal or case manager | Time to first merit call drops | | Write the pod's intake rules and case-type routing | Practice lead plus business ops | Intake qualifies before an attorney touches it | | Stand up the case log and the knowledge capture | Business ops lead | Every matter has 1 source of truth | ##### Phase 3: days 61 to 90, instrument the pod | Action | Owner | Success metric | | --- | --- | --- | | Record outcome data per case: injury type, venue, insurer, demand, settlement | Business ops lead | A queryable commercial record | | Record the firm's own reasoning: briefs, demands, expert choices, negotiation notes | Practice lead | A queryable intellectual record | | Measure NPS per matter | Business ops lead | A baseline score before you scale | #### The pod is ready when 3 things are true 1. The record layer returns **cited output a lawyer verifies** , instead of raw pages. 2. The intake engine **routes by case type** , before an attorney spends an hour. 3. The pod's **economics are visible per case** . If you remember one thing Convert one pod first. Prove the margin, the NPS, and the data loop, then copy the pod to the next practice area. Chapter 12 Building a practice ### How to evaluate an AI platform Every vendor now claims to have AI. The diagnostic is whether the platform helps a team move a case forward, or just answers prompts. A buyer's checklist, adapted from the questions leading plaintiff firms are already asking: **0 of 17** checked. Ask for the cited output first; the rest is easier to verify once you have it. Capabilities - **Reads the medicine** *Does it understand medical records, bills, and specials, not just summarize text?* - **Cited to the page** *Does every generated fact link back to the source page?* - **Case and firm at once** *Does it work case by case and across the whole firm at the same time?* - **Compounding knowledge** *Does its knowledge compound across matters, drafting style, and handoffs?* - **Runs a task end to end** *Can it complete a task and check its own work, or does it only answer questions one at a time?* - **Inherits model progress** *When the underlying models improve, do you get it, or do you wait on someone else's roadmap?* - **Survives a big production** *Can it process a 10,000-page production accurately?* Verification - **Human review step** *Is there a human review step before output becomes work product?* - **Click through to the page** *Can a paralegal click from any line to the page it came from?* - **Shows its work** *Does the vendor show which files it read and how it built the answer?* - **Authoritative research** *Does legal research connect to an authoritative source rather than open web search?* Security and control - **BAA, SOC 2, encryption** *Is there a signed business associate agreement, SOC 2 reporting, and encryption in transit and at rest?* - **Isolated, no training** *Is client data isolated to your firm, with no model training on it?* - **Exit is possible** *Can you export or delete your data on exit?* Adoption and proof - **Results from firms like yours** *Can they show year-one changes in staff load and case output?* - **A named human** *Is there live onboarding, a named contact, and a recent release cadence?* - **Priced to iterate** *Does the pricing let you iterate without paying for every draft?* #### The approaches compared | Approach | What it is good for | Where it falls short | | --- | --- | --- | | Demand-only AI | Fast simple demands | No litigation support; treats every case the same | | Pure AI without a human loop | Fast and cheap | Hallucinations, no legal/medical QA layer | | Legacy case management with bolted-on AI | Familiar interface | Brittle, siloed, not built to reason across records | | General assistants | Brainstorming and internal drafting | No legal/medical context, no source linking, privacy risk | | Agentic operating system | 1 source of truth from intake to settlement | Requires rollout and change management | If you remember one thing Ask for a cited output on your own worst file, not a polished demo. Source-level traceability is the line between a tool you can trust and a tool you have to redo by hand. Chapter 13 Everyone ### Safe AI and the boundaries that protect the firm AI drafts and organizes. A licensed lawyer decides, signs, and stands behind the result. The boundaries are not technical; they are the practice of law. ##### Safe to automate Mechanical work, verifiable against a source page. - Sorting, deduplication, OCR, and handwriting extraction. - Drafting a cited chronology. - Surfacing comparable outcomes. - Drafting routine correspondence. - Routing intake by case type. ##### Human-only The practice of law. A licensed lawyer signs each one. - Deciding liability and causation. - Verifying the chronology and signing the report. - Deciding the settlement number. - Attorney-client privileged advice and strategy. - The final decision to accept or decline a case. #### A working governance loop 1. **Define** which tasks are AI-assisted and which are lawyer-led. 2. **Require a 2-tier review:** paralegal verifies facts, attorney approves final work. 3. **Require the system to check itself** before a human sees the draft: every asserted fact traced to its page, every low-confidence passage flagged, internal contradictions surfaced. 4. **Require a signed business associate agreement** before any protected health information moves. 5. **Train** attorneys on verification, paralegals on source checks, and support staff on escalation. 6. **Put an AI provision in the engagement letter** that names attorney supervision, ties data handling to the confidentiality duties you already carry, and secures client consent. 7. **Review the policy quarterly** as tools and bar rules change. **The risk is real.** Courts have sanctioned lawyers for AI-hallucinated case citations in several documented matters. The failure is rarely the model; it is output that looks right but cannot be traced. That is why source-level citations and human review are not features. They are the permission slip to use AI in a case at all. Those sanctions were not caused by AI. They were caused by AI without a verification layer, filed by a lawyer who did not check. Most jurisdictions now read technology competence into the duty of competence, and that cuts both ways: a firm using these tools without controls has a problem, and a firm refusing to use them at all is defending a harder position every year. The opposite failure is quieter. A lawyer who leans on a system outside the range where it is reliable, and stops interrogating what comes back, ends up worse off than one who never used it. Experience is what makes this safe, which is why 20 years of pattern recognition is the asset AI makes more valuable, not less. If you remember one thing AI drafts and organizes; a licensed lawyer decides, signs, and stands behind the result. No source, no trust. Chapter 14 From the publisher ### The record layer, and an offer Everything above this chapter stands without it. This chapter is the publisher's, and you should read it as an advertisement with the numbers you would want from any vendor. [Medrecords AI](https://medrecords.ai/) makes software for the attention column in a PI firm. It sorts, deduplicates, reads handwriting and imaging, and returns a chronology in which every line cites its source page, under a signed business associate agreement, priced per deduplicated page with duplicates free. 10¢ — a page to start, duplicates free, no subscription Every line — cites its source page, or it is not said Signed BAA — before any protected health information moves #### What it does not do It does not decide liability, causation, damages, or settlement value. It does not retrieve records from providers, sign opinions, or automate the firm's risk decisions. Those are the lawyer's work. The offer #### See it run on a case like yours, in 30 minutes. A 30-minute demo on a live file: sorting, deduplication, handwriting, and a cited chronology you check against the source page while we watch. Bring your questions on price and the business associate agreement. From 10 cents a page, duplicates free, no subscription. [Book a 30-minute demo](https://medrecords.ai/demo/) [See the chronology](https://medrecords.ai/product/chronology/) Scheduling only. No records move from a public page. Chapter 15 Everyone ### Sources and method The 3-person framework, pod model, 2 data infrastructures, compensation structure, and named quotes are adapted from Chia Jeng Yang, *Architecture for the Outcome-Based, AI-Native Law Firm*, Foremark Legal, September 2026. The adoption and trust figures are from a 2026 industry study, n = 207, March 2026. The workflow time estimates are from an industry buyer's guide. The H.B. 837 summary is from the public record of the law. The personal injury application, the attention-versus-judgment split, the case flywheel, and the record layer are Medrecords AI's. Product facts are what Medrecords AI publishes on its own site. No numbers in this manual are fabricated, and none are presented as an independent study. ##### What each source carries - Foremark Legal — The 3-person framework, the pod model, the 2 data infrastructures, the compensation structure, and the named quotes - Industry study — The 10 adoption and trust figures. n = 207 U.S. personal injury legal professionals, March 2026 - Buyer's guide — The 6 manual-task time estimates in chapter 9 - Public record — The H.B. 837 summary in chapter 1 - Medrecords AI — The PI application, the attention-versus-judgment split, the flywheel, the 90-day build, and the record layer ### More guides - [**How to become an AI-native workers' comp claims adjuster in 2026** — What workers' comp adjusters earn, how caseloads and file review actually work, and where AI fits before a…](https://medrecords.ai/guides/ai-native-workers-comp-adjuster/) - [**How to create a verified medical chronology from thousands of scanned records** — How to build a verified medical chronology from thousands of scanned records: OCR, deduplication, date and…](https://medrecords.ai/guides/verified-medical-chronology-from-scanned-records/) - [**How to deduplicate and organize medical records for personal injury litigation** — How to deduplicate and organize medical records for personal injury litigation: exact vs near duplicates…](https://medrecords.ai/guides/deduplicate-medical-records-for-personal-injury/) - [**How to get medical records for a lawsuit: the complete field guide** — How to get medical records for a lawsuit: HIPAA access, authorizations, subpoenas, court orders, fees…](https://medrecords.ai/guides/how-to-get-medical-records-for-a-lawsuit/) - [**How to read a VA claims file (C-file): a field guide for representatives and veterans** — How to request a VA claims file (C-file) with VA Form 20-10206 or VBMS, what each section holds, how to read…](https://medrecords.ai/guides/how-to-read-a-va-c-file/) - [**How to read an autopsy report: findings, opinions and the file behind them** — Read an autopsy report section by section: external and internal exam, toxicology, cause, mechanism and manner…](https://medrecords.ai/guides/how-to-read-an-autopsy-report/) ### Ranked lists