How to make a law firm AI-native: a playbook

In May 2025 two firms, Ellis George and K&L Gates, were ordered to pay $31,100 because a brief built with CoCounsel, Westlaw Precision and Google Gemini cited authority that does not exist[28]. These were legal research products sold to lawyers, not a public chatbot.

In April 2026 a federal magistrate judge in the Northern District of California sanctioned the owner of a small firm personally, $1,001 that the firm was not allowed to pay for him, after a junior lawyer he supervised filed a brief with an invented case. "At minimum," the order says, "a supervising lawyer should read and understand the content of all pleadings and check citations to ensure their accuracy"[13]. In mid-May the State Bar of California approved guidance that tells lawyers not to let an AI system file documents, communicate with the court or send client information out of the firm on its own[6,7].

Meanwhile the economics have not moved. In Clio's 2025 benchmarks a lawyer records 3.0 billable hours of an 8-hour day and the firm collects 2.4 of them (vendor data)[32]. About 90% of legal spending still flows through hourly billing (vendor research)[35], and the ABA says hourly matters must bill the time actually spent, not the time an AI tool saved[1]. Midsize firms are under more pressure than the largest: profit growth in the first quarter of 2026 was about half the pace of the Am Law 100, with productivity per lawyer falling (vendor research)[38].

Agents can now draft much of a matter. The question is whether a firm's walls, review and fee model hold when they do. (In-house legal departments share the matter workflow but not conflicts across many clients or billing, so they are not covered here.)

Five things set the AI-native firm apart:

1. 1. Every agent acts for a named lawyer and sees no more than that lawyer may. A screened lawyer's agent is screened too. Opinion 512 names the risk: a tool can disclose information to people in the firm who are barred from it "because of an ethical wall"[1].
2. 2. Nothing reaches a court, a client or opposing counsel without a lawyer's review. California says no document goes to the court "without lawyer review and approval"[6].
3. 3. Supervision leaves a record. Courts now ask supervisors what they did[13,14].
4. 4. Consent and AI instructions are tracked per client. One client's outside counsel guidelines may forbid putting its documents into a model that another client expects the firm to use.
5. 5. Pricing changes on purpose. Hourly matters bill actual time; agent-heavy work moves to fixed or outcome fees.

No firm will replace its billing system, conflicts database, DMS and docketing engine to become AI-native. Agents will work across them, from several vendors and some the firm builds itself. Each product can control its own AI, but none of them can make a lateral's screen hold for an agent that runs in another product, or say for one matter which agents touched it and who reviewed what went out.

OrchKernel is a control layer that sits between agents and those systems. Agents ask it before they act. It runs each agent as a named person with that person's walls, holds what needs a lawyer or analyst, denies what no agent may do (file, send, move trust money), and keeps a tamper-evident record by matter. It does not decide conflicts, define walls, compute deadlines or replace any system of record.

Where Intapp already covers this. Intapp, whose clients include 97 of the Am Law 100, now sells agents that clear conflicts and vet laterals, "Walls for AI" and a log of what each agent saw, inside its own suite[24]. A firm that runs most of its intake and conflicts work there may get much of this from Intapp. OrchKernel's place is across systems: one point where every agent, whoever built it, acts as a named person under the same rules, approvals and log. The blueprint near the end maps it to the firm's 20 control points.

Clio's 2025 benchmarks, drawn from its customers' data, show lawyers recording 38% of an 8-hour day as billable, firms invoicing 88% of that and collecting 93% of what they invoice[32]. Median lockup is 93 days of revenue: 43 days of unbilled work and 32 of unpaid invoices, reported separately. These are vendor figures, weighted toward solo and small firms.

Partners also write down about 300 hours a year on average (vendor research)[39], mostly routine work that took longer than a client will pay for: the work agents now draft fastest.

Worked rates grew 7.4% in 2025 against 2.8% inflation[36], and firms in Thomson Reuters' index ended 2025 with margins above 40%[37]. Both figures come from Thomson Reuters, which sells CoCounsel and Westlaw. Some volume practices have already moved: an insurance defense firm on KPI-based retainers calls the billable hour "a gigantic waste of time for everyone involved"[40].

ABA Opinion 512 says a lawyer billing hourly "must bill for their actual time". A lawyer who spends 15 minutes prompting a draft bills those minutes plus review, not the hours the draft would have taken by hand. A tool that works like overhead is not billed at all[1]. Florida and Texas say the same[8,9]. On an hourly matter, every hour an agent saves comes off the bill unless the fee model changes.

US legal services employed about 1,249,000 people in September 2026, up 1.9% on a year earlier (preliminary)[21], and there were about 1.37 million resident active lawyers in 2025[25]. The government projects 5% growth for lawyer jobs from 2025 to 2035[22] and none for paralegals and legal assistants, naming AI as the reason demand "is expected to be limited"[23].

A new matter from a partner's email to an opened matter, in a firm at Stage 2. The marked lines are where agents stop because they have no tool to go further.

In Thomson Reuters' survey of 1,514 professionals at the end of 2025, 41% of law firms said their organization uses generative AI; across professions 15% use agentic AI and only 18% collect any measure of return[34]. Respondents were screened for familiarity with AI and the publisher sells AI tools, so treat these as upper bounds. Clio reports that 86% of mid-sized firms use AI and 60% have formal policies (vendor survey)[33]. The gap is mostly definition: Clio counts any use, Thomson Reuters asks about the organization.

Only 57% of mid-sized firms run cloud practice management, against 71% to 74% of solo and small firms[33]. Agents need connected systems, and midsize firms have the fewest. Clients are not pushing yet: 54% of corporate legal respondents think outside firms should use AI, and 67% do not know whether theirs do[34].

Named products are examples from the sources, not recommendations, and their performance claims are their own.

A preregistered Stanford study found Lexis+ AI, Westlaw AI-Assisted Research and Ask Practical Law AI producing hallucinated answers 17% to 33% of the time[31]. Buying a legal-specific product is a reasonable choice and not a control. The independent AI Hallucination Cases database logged 2,145 court decisions worldwide involving AI-invented content by 4 October 2026, 854 involving lawyers and 1,473 in the US[26].

Platforms got the capital. Harvey was valued at $11 billion in March 2026 on $190 million of 2025 revenue[42,47]. Legora reached $5.55 billion in March 2026 and $100 million of annual recurring revenue in April[43,48]. Clio passed $500 million of recurring revenue after a $5 billion valuation and a $1 billion purchase of vLex[41].

The full rebuilds are new law firms. Norm Law uses its parent's agents, "employs human attorneys to supervise them" and charges by outcome; Norm raised $120 million at a $1.2 billion valuation in July 2026[44]. Crosby set up "an actual law firm" for startup contract review with under-an-hour turnaround[45]. Eudia Counsel, co-owned by a legal AI company, was approved as an Arizona alternative business structure in June 2025 and works on fixed fees[49]. In England and Wales the regulator authorized Garfield.Law, an AI-driven firm for small debt claims, on condition that named solicitors stay accountable for everything it produces[17].

Each started without an hourly book of business to protect. We found no established midsize US firm that has rebuilt around agents and published the results.

Six stages, ordered by how close agent work gets to a court, a client or the other side: drafts that stay inside the firm, then writes into the firm's own systems with approvals, then legal work under recorded supervision, then narrow autonomy inside the firm, then pricing. Two timing rules apply throughout. Many firms push to bill and collect in November and December, so billing workflows stay untouched through year-end. Court calendars never pause, so docketing changes always run in parallel with the old process first.

Written for a firm starting in October 2026: Stage 0, plus two Stage 1 workflows in one practice group. The plan leaves billing alone through the year-end push and starts docketing as a parallel run in the new year.

An agent can prepare each of these. A named lawyer, or a person with the authority, makes the decision and answers for it, however good the agent gets.

An ethical wall is an access rule the firm already keeps for people. Under Rule 1.10 a lateral's conflict can be screened only if the lawyer is timely screened from any participation in the matter[3]. California warns that a poorly set-up agent may disclose confidential information "across different matters"[6]. A research agent that searches the DMS under a service account sees every matter, so a screen recorded for the lawyer means nothing for her agent.

The fix is for the agent to act as the lawyer, so the DMS gives it exactly what it would give her.

Real cases first. Most AI failures in law so far are verification failures in filings, and the sanctions are moving up from the drafter to the signer and the supervisor. Each case is tied to the control point that addresses it.

Scenarios, not reported events. Each is what an agent with too much access could do, mapped to the control that stops it.

Almost all of this is existing conduct rules and court rules applied to new tools. None bans AI, and none moves a duty from the lawyer to the software. California's 2026 guidance, written for agentic AI at the state Supreme Court's request, is the closest thing to an agent control list[6].

Four requirements recur. The exact obligation differs by jurisdiction and use; these four are the base.

Research, review and summarizing are where legal users already apply generative AI most[34]. The junior task becomes checking every authority and fact against its source, the step every sanction case above skipped. How juniors learn judgment without writing first drafts is an open question; we found no source that answers it, so design work that still teaches, and track it.

The government projects no growth in paralegal jobs to 2035 and names AI as the reason[23]; the work moves toward running and checking agent workflows. Conflicts analysts move from searching to judging what the agent's list missed. Docketing clerks confirm dates instead of computing them, with the cited rule in front of them. Billing coordinators enforce each client's guidelines and handle the rejections agents cannot fix.

Rule 5.3 already makes managing lawyers responsible for measures giving "reasonable assurance" that nonlawyer help meets their duties[5], and Opinion 512 applies it to AI tools[1]. What changes is the evidence: the order in Hill v. Workday reads like a list of what a supervisor must be able to show[13].

Someone owns approved tools and models, agent playbooks, what agents may do without approval, and change control, including the "periodic reassessment" California asks for[6]. In England the SRA expects the compliance officer for legal practice to own new technology[18]. In a 30-lawyer firm this is part of a job; at 300 lawyers it is a role, usually in knowledge management. The general counsel's office owns consent, AI clauses and the response to outside counsel guidelines.

California suggests "building the time savings created by AI use into their competitive rates"[6]. The committee decides, matter type by matter type, what moves to fixed or outcome fees, and what the firm shows clients about agent work on their matters. The new AI-native firms started there[44,49].

OrchKernel is the layer between AI agents and the systems a firm runs on, as described in the missing layer. Agents ask it before they act; it checks the rules, holds what needs a person, and records what happened.

What it is not. OrchKernel does not replace practice management and billing, the conflicts database, the walls system, the DMS, the docketing engine or trust accounting; they stay the systems of record. It does not clear conflicts, decide screens, compute deadlines, file anything or sign anything.

The controls a law firm needs whatever tools it uses, who owns each, and what enforces it.

OrchKernel is source-available under the Business Source License 1.1 and runs on your own servers, so you can read the code that enforces these controls. Running it in production for your own firm is covered; the license page says what needs a commercial license.

Record the baseline before Stage 1, then track the same numbers at each stage. Public benchmarks barely exist for this work: the detailed law-firm data is paid or for members, and the public figures come from vendors. We have not quoted any number we could not source.

Sources were read in October 2026; dates are publication or data dates. Last reviewed October 2026.