For law firms
What a small law firm should automate first — and what it must never automate
Once the policy exists and somebody owns it, a different question arrives, usually from a partner who has just read something: fine, so what do we actually use this for.
The honest answer is narrower than the market suggests: automate the administration around the matter, never the judgment inside it. Records requests, intake capture, billing narrative cleanup and document drafting are rule-bound, high volume and safe to hand to an agent with a named reviewer. Conflicts clearance, fee quotes, deadline commitments and legal judgment are not, and no amount of model improvement changes that — they are refusals, not roadmap items.
The test has four parts. Is the work rule-bound — could you write down what a correct outcome looks like in advance, without knowing the matter? Is it high volume — dozens of times a week, so a small saving compounds and a small error surfaces early? What is the escalation rule — the written list of conditions under which the agent stops and hands to a person? And where does a person sign? Anything failing the first two is not worth building. Anything without clear answers to the second two is not safe to build.
Records requests are the best first build and the least glamorous
Almost entirely rule-bound. Medical records, employment files, police reports: the request letter is a template, the tracking is a spreadsheet, the follow-up is a calendar, the receiving is a filing task. In personal injury, employment and insurance defense the volume is enormous.
The agent generates requests, tracks what is outstanding, sends follow-ups on schedule, files what arrives to the right matter and reports what is missing. It escalates when a custodian refuses, when a response is incomplete, when a fee is demanded, or when something arrives that looks privileged or out of scope. A paralegal or attorney reviews received records before they enter the working file, and nothing is summarized into a demand package without a person having looked.
No part of this is close to the practice of law, which is exactly why it works as a first build.
Intake is where response time decides outcomes, but the agent never clears a conflict
A new-matter inquiry has to be classified, a record created, parties extracted, required fields populated. Those are rules, and inquiries arrive overnight, over weekends and during trial.
The agent extracts party names, adverse parties and related entities, runs them against the conflicts database and produces a flag report. It never clears anything. Any hit, any near-match, any name it cannot resolve goes to the person who owns conflicts, with the reasoning attached. Conflicts clearance is where a person signs, always.
Clio's 2025 Legal Trends Report found firms using integrated intake technology saw 53% higher revenue among solos and 28% higher among small firms — Clio's correlation on Clio's own data, reported by a vendor with a product in the category, so hold it loosely.
Document production is the highest-value item and the one with a citation problem
Partly rule-bound, and the boundary is the whole design. The structure of an engagement letter, a standard demand, a records-request cover letter, a deposition summary, a routine motion shell follows patterns your firm already has. The judgment inside them does not.
The agent drafts from your templates, your prior work product and the matter file. It never sends. Anything containing legal authority routes into the verification checklist first — every citation pulled, read for the proposition offered, quotations matched character for character, the verifier named on the record. Anything the agent could not source from the file is flagged as an open question rather than filled in plausibly. That is the single most important design decision in the build.
The supervising attorney signs before it leaves the firm, and on anything with a citation in it, after the verification record is complete.
Billing narratives are rule-bound, and the fee rule is where firms trip
"Reviewed docs" is not a billable narrative and everyone knows it. What a client will accept is written down, often literally, in their outside counsel guidelines. Every timekeeper, every day — the volume test passes easily.
The agent rewrites narratives for clarity and guideline compliance from what the timekeeper actually wrote. It never changes time. It never combines or splits entries. It never invents a task not in the original, and where the original is too thin to rewrite honestly it flags the entry back to the timekeeper rather than embellishing it. The timekeeper approves their own entries; the billing attorney approves the bill.
One caution that matters more here than anywhere else: ABA Formal Opinion 512, July 29, 2024, requires disclosure of AI use where it bears on the reasonableness of a fee, and it bars billing a client for the lawyer's own learning curve on a tool. Decide your position before an agent touches a bill, not after.
The refusal list is a section of the specification, not a caveat
This gets written down before anything is built, and the firm approves it.
No legal judgment. The agent does not advise, assess the merits of a matter, choose a strategy or decide what a document should argue. If a task requires a lawyer's judgment, the agent's only correct output is a draft handed to a lawyer.
No fee quotes. Not to a prospect, not to an existing client, not "approximately."
No deadline commitments. The agent does not agree to an extension, propose a date or accept a scheduling request that binds the firm. It reports dates from the file and escalates anything that would create one.
No conflicts clearance — flagging only. Clearing a conflict is a professional judgment with disqualification and malpractice consequences behind it. We will not build a system that produces the word "clear."
Above all four sits the refusal that applies everywhere we work: we do not build anything that makes a decision about a person — employment, credit, medical, housing.
What to do Monday
- Count how many records requests the firm sent last month. If it is over forty, that is your first build.
- Write the escalation rule for that one workflow. One page, in your words.
- Name the reviewer. A role is not enough; write a person's name.
- Check whether your system of record holds the data the workflow needs. If it lives in Outlook folders, fix that first.
- Decide your fee-disclosure position before touching billing.
What we will not do
We do not build without a Blueprint, however well a firm thinks it knows its own process. Agent Blueprint is $4,900 — two weeks, one workflow you name, the specification, a test of whether your data can support it, and the escalation and refusal lists for your approval. Sometimes the answer is that you should not build, and you stop having spent $4,900 to find that out rather than $14,900. That refusal costs us deals and we keep it. Build is $10,000, $14,900 for both. A second agent is $9,900 because the context pack already exists.
Signet, a division of Circle Square Consulting. AI that holds up. Radnor, Pennsylvania. /renewal-dry-run