AI automation for law firms: what to build first.
Most legal AI content is written by people selling software they have never operated. I run a production AI system inside legal intake: a mass-tort reply engine that calls back people who asked a law firm for help, screens the claim, and moves qualified callers toward a signed retainer. It has logged 50,000 calls. This guide is what that operating experience says a firm should automate, in what order, and where the lines are.
Automate intake speed first: answering, screening, scheduling, callback chasing. Drafting and research second. Billing hygiene third. Never let a system anywhere near legal advice; route that to a person by architecture, not by hope. Judge every automation on one number a partner already cares about: signed retainers, collected hours, or cost per case.
The four automations, in the order that pays.
| # | Automation | Why this order | The number it moves |
|---|---|---|---|
| 1 | Intake response: AI answering, screening, scheduling, and callback chasing, 24/7 | A lead contacting you is contacting competitors in the same hour. Speed decides who gets the retainer. My field data: the lead who cannot answer at 2 p.m. often answers at 7 — coverage wins cases people never see you lose. | Signed retainers |
| 2 | Document drafting: first-pass demand letters, discovery responses, client updates | High volume, strong templates, human review already built into legal workflow. AI drafts, a lawyer signs. | Billable hours recovered |
| 3 | Case research & summarization: medical records, depositions, files | Real leverage, but quality control is harder; do it after your review discipline exists. | Hours per case |
| 4 | Back office: billing narratives, time capture, status reporting | Worth doing; just never the reason a firm falls behind competitors. | Realization rate |
Where the automation must stop.
In my mass-tort engine the line is bright: anything that smells like legal advice goes to a person. The agent can confirm an appointment, collect facts, schedule a callback. It is an intake assistant, not a lawyer, and the architecture remembers that even when the model forgets. Build every legal automation with the same discipline:
- No legal advice, by architecture. The handoff to a human is enforced in the system's structure, not requested in a prompt.
- Confidentiality mapped before launch. Vendor agreements that exclude training on your data, retention controls, and a written map of what never leaves the firm.
- Fallbacks that keep the call moving. When the model degrades, the agent drops to a scripted recovery path that still moves the call forward. At 50,000 calls, a failure that shows up once in a thousand calls has already happened to fifty real people.
- Replay before you ship. Every prompt change runs against transcripts of past calls, the strange ones especially, before it touches production.
The full architecture, latency budgets, fallback ladders and guardrails included, is published in Prompt architecture for production voice agents. The conversion side, what makes an AI call get answered and trusted, is in Voice AI that actually converts.
Where firms waste the money.
Skip these
- A chatbot on the website before the phone and form response problem is solved. The website is not where your leads are leaking.
- "AI transformation" retainers with no number attached. If the vendor cannot name the metric, the metric is their invoice.
- Automating research before review discipline exists. An unreviewed hallucinated citation is a sanctions motion, and the case law on that is no longer hypothetical.
- Custom builds where an off-the-shelf tool scores 8/10. Custom is for the workflow that differentiates the firm, usually intake.
Operator experience, not a reseller pitch.
I'm Micah Berkley, Fractional Chief of AI, based in Miami. Two decades in enterprise tech: SRE and cloud architecture at Google, machine learning at BMW of North America, big-data marketing science at Fashion Nova. The legal-intake system described on this page is mine; the 50,000 calls are the receipt. For firms, engagements run the same way as everything I do: one KPI, 90 days, your team trained to run what gets built. If you're evaluating the executive seat itself, the fractional Chief of AI buyer's guide covers costs and the questions to ask.
Direct rates are published at /consult ($150/hour individuals, $240/hour business). Every engagement starts with a free 30-minute diagnostic call.
The questions I get every week.
Straight answers, no hedging. If yours isn't here, ask me directly.
What should a law firm automate with AI first?
Intake response speed. A lead who contacts a firm about an injury or dispute is contacting other firms in the same hour, and the firm that responds first usually wins the retainer. An AI intake system that answers, screens, and schedules within minutes is the highest-leverage first automation for most consumer-facing practices.
Can AI give legal advice during intake?
No, and the system must be built so it cannot. In the production legal-intake system Micah Berkley runs, the line is explicit: the agent confirms appointments, collects facts, and schedules callbacks, and anything that resembles legal advice routes to a person. The architecture enforces that boundary even when the model would not.
Does AI intake actually work for law firms?
Micah Berkley operates a mass-tort reply engine, an AI voice system that calls back people who asked a law firm for help with an injury claim, screens the claim, and moves qualified callers toward a signed retainer. It has logged 50,000 production calls. The published field notes cover what makes calls get answered, trusted, and filed.
What does AI automation cost for a law firm?
Off-the-shelf intake and drafting tools run from tens to a few hundred dollars per user per month. Custom production systems are project-priced in the tens of thousands. Micah Berkley publishes direct rates at micahberkley.com/consult ($150/hour individuals, $240/hour business) and starts every engagement with a free 30-minute diagnostic call.
Is AI safe for confidential client information?
Only with deliberate architecture: vendor agreements that exclude training on your data, retention controls, and a clear map of what never leaves the firm. A competent implementer answers questions about where the data goes without checking notes. If a vendor cannot, that is the audit finding.
Will AI replace paralegals and intake staff?
In practice it re-tasks them. The systems worth building remove the repetitive volume — after-hours answering, callback chasing, first-pass screening — and hand humans the judgment calls. In legal intake the handoff is mandatory anyway, because the moment a conversation approaches legal advice it must reach a person.