On November 10, 2026, the AI-native law firm Crosby told Artificial Lawyer that it had put a $5m liability policy on its paralegal agents. The policy comes from Testudo and is underwritten by Lloyd's of London, and Crosby believes it is the first time a law firm has insured the outputs of its AI agents. The insurance got the headlines. The number I keep returning to sits a few paragraphs higher. Before launch, Crosby says its agents reached 94% accuracy on escalation, a score that counted a task as right if the agent either finished it correctly or correctly handed it to a lawyer.
Read that metric twice. Crosby graded the agent on two outcomes, and one of them is the agent stopping. In that grading, an agent that knows the edge of its permission and walks a file to a human scores exactly as well as one that does the task itself. A managing partner at a five-lawyer firm can learn more from that one design decision than from any vendor demo running this month.
Small firms keep asking what they should hand an AI agent and what they should keep human. They usually want a list of tasks. The list matters, and I give one below. The thing that decides whether the list survives contact with a real client is the knock: the exact point where the agent stops, packages what it knows, and puts it in front of a person with a license. Firms that design the knock first will be able to hand over more work, sooner, with less fear. Firms that buy a tool and hope will spend 2027 explaining themselves to a judge, a bar counsel or an underwriter.
The month the rules caught up
Three things landed within about three weeks of each other, and together they settle most of the argument about where the lines go.
On September 30, Governor Newsom signed California's SB 574, which WilmerHale summarized on October 20. It takes effect in January 2027. It requires every citation in a court filing to be personally verified by the attorney responsible for submitting it, whether or not AI produced the citation. It requires disclosure of generative AI use in documents submitted to the court. It also says lawyers may not "delegate the practice of law" to generative AI, without defining where that line falls. The State Bar of California's 2026 practical guidance fills part of the gap. Per WilmerHale's reading, it says attorneys must not deploy agentic systems to make substantive legal determinations, communicate legal advice, or file pleadings without meaningful attorney supervision and review. The guidance puts it bluntly: "Critically, any use of AI must not diminish or abdicate professional judgment."
On Saturday, October 31, the New York State Bar Association's House of Delegates adopted its Legal Ethics Guidelines for Lawyers' Use of Artificial Intelligence. NYSBA framed the guidelines as a way to "ensure that legal decisions are in human hands." The report says lawyers may bill for time spent verifying an AI tool's output and may not bill for time spent learning the tool. It treats the cost of buying a tool as overhead. It extends the duty to supervise non-lawyers to AI systems, and it addresses reviewing the output of agentic systems directly.
On October 26, Clio, the practice management vendor many small firms already run on, unveiled a unified platform with agentic workflows, with the new interface due in November. Clio's own example is instructive. The platform can assess a prospective client's inquiry, identify missing information, prepare a consultation, and, following approval, open the matter and organize the initial work. Legal IT Insider reported that Clio says the new experience shifts responsibility for coordinating the work from the individual to the platform.
So within one month the regulators named the lines, and the vendor most small firms use put agents a click away from every matter file. The vendor's own example already contains a knock, in the words "following approval". The managing partner's job is to decide where every other knock goes.
What Crosby actually built
Crosby's description is worth reading closely because it is a small, specific architecture, and you can copy its shape without copying its stack.
Every new client message passes through an agent first. The agent does end-to-end work inside a defined capability set, using context Crosby stores on each client. Anything outside that set goes to a human lawyer, with the documents and context attached, and the agent reports updates back to the client. Each agent runs in a sandbox and sees only the relevant client's data, with full isolation between clients. In Crosby's screenshot the agent is labeled as an agent while it chats with the client.
The payoff Crosby reports is concrete. Human review time has fallen to zero for 20% of its preexisting legal tasks, including summaries of negotiations, pulling prior business context, and routing to the right people for approval. Crosby puts typical human review at about 180 hours a day and says the agents save 36 hours of it every day.
Emily Zhang, the Crosby engineer who leads the client agents, described the split this way: "Our lawyers are amazing at making high-stakes decisions." She went on to say that when a client wants to know the last time it agreed to a 3x liability cap, the agent answers in under five seconds. Memory and retrieval went to the machine. Decisions stayed with the lawyers.
The tasks Crosby moved to zero review have something in common. Every one of them is retrieval, summary or routing. None of them asks the agent to say what the law means for this client, whether a position is strong, or where money should go. Crosby drew its capability set around work where a wrong answer is visible and cheap to fix, and it scored the agent on knowing when it had reached the edge.
Three hard lines for a small firm
Every small firm needs three lines the agent never crosses, whatever the vendor says the product can do. I hold them as firm policy, and I think the rules issued this autumn point to the same three places.
No legal advice
The agent can tell a client that their documents arrived, that the hearing is on the calendar, that the lawyer will call by Thursday. It cannot tell a client what they should do, what their rights are, or what a clause means for them. California's guidance names communicating legal advice as something agentic systems must not do without meaningful attorney supervision and review. NYSBA lists the unauthorized practice of law among the risks its guidelines exist to prevent.
In practice the hard part is the client who asks. Clients will type "so should I sign it?" into whatever box you give them. The agent needs a written rule for that exact moment, and the rule is a knock. It acknowledges the question, tells the client a lawyer will answer it, and opens a task for the responsible attorney with the thread attached. An agent that tries to be helpful at that moment is the most dangerous thing in your office.
No case-merit decisions
Whether to take the matter. Whether to settle. Whether a claim is worth filing, whether a defense holds, whether to recommend the plea. These are judgments about merits, and they stay with a licensed human whose name goes on the advice. NYSBA's phrase about keeping legal decisions in human hands covers this, and so does California's ban on agents making substantive legal determinations.
The trap is subtle. An agent that scores incoming leads by "likely case value" is making a merits call wearing a marketing costume. An agent that ranks a stack of documents as "helpful" and "unhelpful" to your client has made a merits call before any lawyer opens the folder. Let the agent sort by date, by party, by document type, by whatever a paralegal could sort by without knowing the law. Keep the adjectives about strength and value for the lawyer.
No trust or IOLTA accounting
This is the line I hold hardest, and it is my position rather than a quotation from either bar report. Money in a client trust account or an IOLTA account belongs to clients. The safekeeping duty (Rule 1.15 in the ABA Model Rules) sits personally on the lawyer, and trust account problems are among the fastest routes to discipline that a small firm can find. A mistake there is a breach of a fiduciary duty, and it surfaces in a bar audit long after the agent that made it has been updated twice.
So the agent never writes to the trust ledger, never moves funds between trust and operating, never approves a disbursement and never performs a reconciliation. It can read. It can flag that a retainer has arrived, or that an invoice is ready to be paid from trust once a human approves the transfer. It can draft the reminder that this month's reconciliation is due. The pen that touches client money stays in a human hand, and the agent's credentials should make it technically impossible for that to change.
What to hand an AI agent in a small law practice
With the lines drawn, the useful list is long. Everything on it shares Crosby's profile: retrieval, summary, routing and first drafts, where a mistake is visible and a human signs before anything leaves the building.
Intake, up to the door
Let the agent take the inquiry, collect the facts, run the conflict search, flag missing information and prepare the consultation file. That is close to Clio's own example. The knock comes before the engagement decision. A lawyer decides whether to take the matter, and the agent opens it only after that approval.
The record of every conversation
Clio Phones records, transcribes and summarizes calls and attaches them to the right matter, and Clio Mobile can do the same for in-person meetings with permission. This is excellent agent work, because nobody at a small firm writes complete call notes. NYSBA's guidelines say to notify all parties when using AI for recording and summaries, so the notice belongs in the agent's script, every time, before it starts listening.
Chronologies, dockets and deadlines
Clio says a new filing can prompt its platform to assess implications, update the matter chronology and prepare a response for the responsible lawyer. I would split that in two. Updating the chronology and calendaring the deadline go to the agent. Assessing implications and preparing the response are drafts for a lawyer, and the knock sits between them.
First drafts that a lawyer signs
Engagement letters from firm-approved templates, court forms, discovery requests, routine correspondence and research memos are all fair work for an agent, provided a lawyer reads every line before it goes out. SB 574 makes the citation step explicit in California from January 2027, since every citation in a filing must be personally verified by the responsible attorney. Build that into the workflow as a separate, logged step, so the verification is something your system can prove happened.
Time entries and billing prep
Clio says its phone system prepares time entries for review. Let the agent draft them. Then apply NYSBA's billing rules: verification time can be billed, learning time cannot, and the tool itself is overhead. Say an agent cuts a two-hour research task to twenty minutes of generation and thirty minutes of checking. Under NYSBA's guidelines the thirty minutes of checking is billable work, and the two hours you no longer spend never were. Firms on hourly billing should run that arithmetic on their own matters before a client runs it for them.
Client status, labeled as a machine
Crosby's agent identifies itself as an agent in the client chat. Copy that. A client should always know whether they are talking to a person, and NYSBA's guidelines push in the same direction on disclosure. The agent can answer "where is my file" all day. The minute a status question turns into a "what should I do" question, it knocks.
The login is a line too
The most instructive sanctions story this month involves an older case. On November 10, 2026, Greenberg Traurig's Kathryn Cole wrote up the LiveVideo sanctions in the Southern District of New York for the National Law Review. A filing had kept a leftover line from a chatbot, "Here's a revised and more coherent version of the text, with duplicate ideas removed and missing ideas added to lists." The magistrate judge did not rest her Rule 11 recommendation on the AI drafting itself. Part of what she cited was counsel's abdication of his nondelegable duties, including letting others use his ECF credentials to file papers he did not write, review or authorize. The AI text served as evidence of that abdication.
Every small firm about to install agentic workflows should take that point personally. Most agent failures in law will look like this one: a filing system, an email account or a payment portal that the agent could reach because someone handed it a lawyer's keys. The Ninth Circuit made a related point in Malkeet Lnu v. Blanche, as WilmerHale describes it. The court said its sanctions were imposed for failing to comply with procedural and ethical duties that apply regardless of how legal papers are prepared, and it did not sanction the use of AI as such.
So give agents their own identities with their own permissions. The agent gets an account that can read the matter, write drafts and create tasks. It never gets the credential that files with a court, sends from a partner's address without a human click, or moves money. When something goes wrong, your logs should show which identity acted, and the answer should never be "the partner" when the partner was asleep.
Measure the knock
Crosby's 94% is a number most small firms have never thought to collect, and it is the one I would build the whole program around.
The usual way to evaluate a legal AI tool is to ask whether its answers are right. That metric pushes vendors toward agents that answer everything. An escalation metric rewards a different behavior. It asks, for each request, whether the agent did the work it was allowed to do correctly and handed off the work it was not allowed to do. An agent that drafts a pleasant reply to "should I sign it?" fails this test even if the advice happens to be sound.
For illustration, here is how a five-lawyer firm could run it without a data team. Pull the last 200 client messages and intake inquiries from your inbox. Have a senior paralegal mark each one with the correct outcome: handle, or knock, and to whom. Run the agent against the same 200 in a sandbox. Count the disagreements, and read every one where the agent handled something that should have knocked. Those are the only errors that can hurt you badly. An agent that knocks too often costs you some lawyer minutes. An agent that answers when it should have knocked costs you a client, a grievance or a sanction.
Suppose the agent knocks on 40 messages a paralegal would have handled. That is a tuning problem, and you can widen its capability set next quarter. Suppose it handles three that should have gone to a lawyer, and one of them is a client asking whether to accept a settlement. That is a stop-ship, whatever the overall score says. The asymmetry is the policy. Widen the capability set slowly, one task family at a time, and re-run the test every time you do.
What your carrier will ask
The detail in the Crosby interview that should change how small firms plan is the underwriting process. Before writing the policy, Testudo ran a technical analysis of Crosby's systems. It covered the technology and models, governance controls and logs of human review steps, along with how the agents work, what data they access and how the models are trained. Testudo also reviewed internal contracts and policies. The policy covers agents doing legal tasks under attorney supervision and agents doing paralegal work on their own.
Crosby is an AI-native firm that went looking for coverage. Your firm will meet the same questions from the other direction, at renewal, from the carrier that already writes your malpractice policy. The questions will be close to Testudo's. Which tasks can the agent do alone? Where are the logs of human review? Can the agent reach client funds? Can it file? When a client asks for advice, what happens?
A firm that can answer with a written capability set, a list of knock points, an escalation test with a score attached and separate agent credentials has something an underwriter can price. A firm that answers "we use the AI features in our practice management system" has handed the underwriter a reason to exclude the risk or charge for the uncertainty.
Why the vendor cannot draw your lines
Clio's Jack Newton said at the launch that "The ability to take on work is only as valuable as the understanding behind it." He is right, and it cuts both ways. Clio can give your agents context. It cannot know that your family law practice treats every message mentioning a child as a knock, or that your estate practice never lets an agent confirm a distribution date, or that one of your partners promised a client that only she would speak to them. Those rules live in your firm's judgment, and the platform will run whatever rules you give it, including the default ones you never looked at.
This is the structural shift of the last thirty days. Until October, a small firm could treat AI as a feature it switched on matter by matter. With agentic workflows arriving inside the system of record, California writing "delegate the practice of law" into statute, and New York telling lawyers to supervise AI as they would supervise a non-lawyer, the firm now has a delegation policy whether it wrote one or not. If nobody writes it, the vendor's default configuration becomes the policy.
Draw the doors before you open them
Managing partners should stop shopping for features and start drawing doors. Write down the capability set: the task families the agent may finish alone. Write down the three hard lines, legal advice, case merits and client money, and enforce them with permissions as well as prompts. Give every agent its own identity, with no filing credentials and no ledger write access. Build an escalation test from your own messages and score it before launch, after every change and before every renewal. Then widen the set as the scores earn it.
This is an architecture problem, and it has to be solved before the tool arrives, because the tool arrives with its own answers already filled in. Agor AI Advisory designs these delegation maps for practices that want the 36 hours without the grievance: capability sets, knock points, credential separation, escalation tests and the evidence file your carrier and your bar will ask to see. The firms that get this right in the next two quarters will hand their agents more work in 2027 than their competitors dare to, because they will be able to prove where each agent stops.
Sources
- Artificial Lawyer, Crosby Insures Paralegal Agents, Interview, November 10, 2026
- WilmerHale, California Adopts Law to Impose New Guardrails on Attorneys and Arbitrators Use of AI, October 20, 2026
- New York State Bar Association, NYSBA Releases In-Depth Guidelines To Advise Attorneys on Ethical Use of AI, October 31, 2026
- Legal IT Insider, Clio unveils unified platform with agentic workflows, AI phone system and new DMS, October 26, 2026
- National Law Review, Three Strikes and You're Out, Sort of, November 10, 2026