AI agents for law firms: the four worth building, and what each one actually does (2026)
What an AI agent does inside a law firm, the four we build first (intake, follow-up, pipeline, reporting), what each costs, and what never leaves the attorney's desk. From a team that took one firm's response time from 52 hours to under four minutes.
Your phone now has an agent on it. Meta's Muse went to the top of the app charts in September, and three weeks later Meta pointed it at small businesses, with connections to QuickBooks, Slack and Shopify. Another one, Instinct, will take a text message and go make the phone call for you.
So a managing partner asks the reasonable question. If my phone can do that, what is my firm waiting for?
An AI agent for a law firm is software that carries one job from start to finish: it reads what comes in, decides what to do under rules the firm wrote, takes the action in the firm's own systems, and hands off to a person at the points the firm chose. The useful ones are not general assistants. They are built around a single job, such as intake, lead follow-up, pipeline review or reporting, and wired into the CRM and calendar the firm already runs. We build these for law firms. One intake agent took a firm's response time from 52 hours to under four minutes and raised consultations 2.4 times, with no added staff.
That's the definition. The trouble starts with the word "agent," which is currently being used for three different things. Our guide to AI agents for business covers the same three outside the law.
Three things called an agent
The personal agent. Muse, Instinct and their cousins. Brilliant at errands. It knows your calendar and nothing about your conflict check, your fee agreement or which matters you decline.
The legal product. CoCounsel, Harvey, Spellbook, Clio Duo. These are what Google lists when you search this topic, and they are good at the work lawyers trained for: research, drafting, contract review. If that's your bottleneck, buy one. We'd tell you the same on a call.
The firm's own agent. Built for one job in your practice, using your intake criteria, your CRM, your calendar and your voice. This is the kind nobody sells in a box, because the box would have to contain your firm.
Most small and mid-sized firms don't lose money in research. They lose it between the phone ringing and the consultation being booked. That's the third category's territory.
An agent is a job description, not a product
Here is the test we use. If you can't write the agent's job the way you'd write a posting for a new hire, you aren't ready to build it.
What it handles. What it must never do. Who it reports to. How you'll know it's doing well.
Firms that start from a product end up with six tools and no adoption. Firms that start from the job description end up with something that works on Monday.
The four we build first
| Agent | The job | Hands off to a person when | Result we've measured |
|---|---|---|---|
| Intake agent | Answers new inquiries at any hour, qualifies against the firm's criteria, collects documents, books the consultation, creates the CRM record | The matter is outside criteria, the prospect asks a legal question, or anything looks urgent | Response time 52 hours to under 4 minutes; consultations up 2.4× |
| Follow-up agent | Reads where each lead came from and runs the follow-up sequence written for that source | A lead replies with a question or goes quiet past the set window | 4 source-specific tracks; every lead routed to a matched sequence |
| Pipeline agent | Reads every call, deal and lead in the CRM each morning and sends each team member their own action list before 8 a.m. | Always. It briefs people; it doesn't act for them | 100% of open leads with a scheduled next action, across a 5-person team |
| Reporting agent | Rebuilds attribution and marketing reports from the raw records on a schedule, and flags what changed | A number moves outside the expected range | One firm's paid-search revenue was being reported as $8,500. The real figure was $56,500 |
Every row is a build we shipped for a law firm or a professional-services firm. The case studies are public: intake, the daily sales briefing, source-based follow-up and attribution.
Look at the third column before the fourth. The handoff rule is the design. An agent with no handoff rule is a liability with a login.
What stays on the attorney's desk
Legal advice. Conflict decisions. Fee quotes. Anything a client would reasonably believe came from a lawyer.
The intake agent we built does not answer "do I have a case." It says an attorney will, and books the time. That line was written down before any code was.
Then there is confidentiality. Client data should only pass through tools whose terms and audit trail your firm has actually read. Consumer agents on a personal phone don't qualify. Even inside Claude, Anthropic's HIPAA agreement leaves some agent features out, including cloud sessions, computer use and the Chrome extension, which is why our guide for law firms spends more time on the compliance layer than on the model. The model is rarely the risk. The plumbing is.
What it costs to have one built
Search results for this topic contain no prices at all. Here are ours, from the same public rate card as everything else we do.
| Build | Typical engagement | Price |
|---|---|---|
| One agent on clean systems (follow-up, reporting) | Sprint, one to three weeks | $2,000 to $5,000 |
| One agent plus the cleanup and connections under it (intake, pipeline) | Project Build | $5,000 to $25,000 |
| Several agents, maintained and extended | Embedded Partnership | $3,000 to $8,000 per month |
The intake build took three weeks. Most of the first one went to discovering that the firm didn't need the chatbot it had asked for. The build log has every decision in order, including the dead ends.
You also pay for the platforms it runs on, in your own accounts, at cost.
Buy, wire, or build
Buy the legal product when the bottleneck is legal work.
Wire what you already own when the tools are fine and the gaps between them are the problem. Often that's automation with no agent in it at all, and it's the cheapest fix on this page.
Build an agent when the job needs judgment, runs every day, and touches more than one system. Intake is the usual first one because it's where the money leaks and where the handoff rule is easiest to write.
Where to start
Write one job description. If it takes more than a page, the job is too big; cut it in half.
If you'd like to see where your firm stands first, the AI visibility scorecard is free and shows how your firm appears when a prospective client asks an AI for a lawyer. If you already know which job you'd hand off, book a call and bring the description.
Your phone can already make a dinner reservation. The firm's agent should be able to book a consultation.
Frequently asked questions
What is an AI agent for a law firm?
An AI agent for a law firm is software that carries one defined job from start to finish: it reads incoming information, decides what to do under rules the firm set, takes action in the firm's systems, and hands off to a person at agreed points. Common first jobs are client intake, lead follow-up, pipeline review and reporting. It differs from a chatbot, which only answers questions.
What can AI agents do for law firms today?
In production today, agents answer new inquiries around the clock, qualify prospects against a firm's criteria, collect documents, book consultations, run follow-up sequences, brief the team on the pipeline each morning and rebuild marketing reports. Legal products such as CoCounsel and Harvey also handle research, drafting and contract review under attorney supervision.
How much does it cost to build an AI agent for a law firm?
At Automaton's published rates, a single agent on clean systems is a Sprint at $2,000 to $5,000. An agent that also needs data cleanup and system connections, such as intake, is a Project Build at $5,000 to $25,000. Ongoing maintenance and additional agents run $3,000 to $8,000 per month. Platform costs are separate and stay in the firm's own accounts.
Should a law firm buy a legal AI product or build a custom agent?
Buy a legal product when the bottleneck is legal work: research, drafting, review. Build a custom agent when the bottleneck is the firm's own process, such as intake or follow-up, because no product knows your criteria, your CRM or your calendar. Many firms do both.
Is it safe to use AI agents with client information?
It can be, if client data only passes through tools whose terms, data handling and audit trail the firm has reviewed, and if the agent hands off to an attorney for anything resembling legal advice. Consumer agents on personal devices generally do not meet that bar. The attorney remains accountable for the work.
Can Meta Muse or another personal agent run a law firm's intake?
Not well. Personal agents are built for one person's errands and know nothing about a firm's intake criteria, conflict rules or fee structure. A firm's intake needs an agent built around those rules and connected to the firm's own CRM and calendar, with a written handoff to an attorney.