AI agents for law firms: screen intake in minutes and answer status calls without billing for them
A small firm loses more time to unqualified intake, incomplete conflict checks, and "any update?" calls than to legal research. An agent can absorb the first pass of all three — as long as it never crosses into giving advice.
A law firm's first agent should be an intake and communication assistant, not a legal one. It collects the facts needed to qualify a matter, assembles conflict-check data, drafts routine status updates from the matter file, and summarizes long documents with page references. Attorneys make every legal judgment, clear every conflict, and review every outbound message.
Key takeaways
- Intake is the highest-return target because unqualified calls consume attorney time and after-hours inquiries go to whichever firm answers.
- An agent can prepare a conflict check; only an attorney can clear one. Build the workflow so no matter opens without a documented review.
- Document summaries are useful only with citations to the source page. Without them, a summary is a liability with a confident tone.
- Confidentiality and unauthorized practice of law are the two constraints that shape every script, tool permission, and vendor contract.
Where the hours actually go
In a five-attorney firm, the phone is a filter that nobody owns. A paralegal or office manager answers 30 to 60 inquiries a week. Many are outside the firm's practice areas, in the wrong jurisdiction, or already past a deadline. Screening one call properly — practice area, facts, parties, timing, fee expectations — takes 15 to 25 minutes if it is done well, and most firms do it in five because the alternative is a backlog.
After-hours inquiries are worse. Someone calls at 8 p.m. after a car accident or a termination. They reach voicemail, leave a vague message, and call the next firm in the morning. The firm that answers first has an advantage that has nothing to do with legal skill.
Conflict checks are the second queue. A proper check needs the full legal names of every party, every related entity, and often the names of key witnesses and insurers. Staff type those into a database, scan results, and then guess whether a near-match is the same person. Incomplete intake data produces incomplete checks, which means the attorney either redoes the intake or accepts a risk they cannot see.
Status calls are the third. Clients call to ask whether a filing went out, whether opposing counsel responded, or when the next hearing is. The answer is usually in the case management system, but retrieving it, writing a careful reply, and documenting the call takes ten minutes that cannot be billed. Multiply that by a handful of clients per matter and it becomes a real drag on paralegal capacity.
Document review is the fourth, and it is the one most likely to be over-automated. A 300-page deposition, a set of medical records, or a contract production needs a chronology, a list of exhibits, and a summary that a partner can trust. Reading it takes hours. Summarizing it badly takes minutes and creates a malpractice exposure that surfaces months later.
Four use cases that pay for themselves
1. Intake screening and qualification
What the agent does: answers the inquiry by phone, chat, or web form; identifies the practice area; captures jurisdiction, key dates, and a short factual narrative; flags deadline-sensitive matters; collects the names needed for a conflict check; and books a consultation with the right attorney.
What stays human: the attorney decides whether to accept the matter, discusses fees, and gives any legal opinion. The agent must not characterize the strength of a case or estimate an outcome.
Metric to watch: intake-to-consultation rate, and the percentage of consultations that convert to an engagement. Pair them, or the agent will book a lot of consultations that should have been declined.
2. Conflict-check preparation
What the agent does: normalizes names and entity structures, runs the search across the firm's conflict database, groups near-matches, and produces a one-page report with the source of each hit. It also records when the check was run and against which names.
What stays human: the attorney reviews the report and documents the clearance decision. The agent never marks a matter as conflict-free.
Metric to watch: the share of conflict checks completed before the first substantive meeting, and the number that require a second pass for missing names.
3. Document summary and chronology
What the agent does: summarizes depositions, discovery responses, and records into a dated chronology with page citations, extracts party names and exhibit references, and flags inconsistencies between two documents.
What stays human: the attorney reads the underlying pages for anything contested, decides what is privileged, and verifies every citation before it goes into a brief. The summary is a map, not the evidence.
Metric to watch: hours per document set, and the correction rate when an associate checks the summary against the source. If corrections are frequent, the prompt or the model is wrong for the task.
4. Client status updates
What the agent does: reads the matter status, pulls the next scheduled event, drafts a short update in the firm's voice, and sends it for attorney approval. It can also answer logistics questions about scheduling and documents received.
What stays human: anything that touches strategy, expectations, or an assessment of the case. Every outbound message gets a named reviewer.
Metric to watch: inbound status calls per active matter, and the time from status change to client notification.
What not to automate
The line is clearer in law than in most industries, and it is enforced by bar regulators rather than by customer preference.
- Legal advice or opinions. No assessment of merits, no prediction, no "you have a strong case." That is the practice of law, and a disclaimer does not convert it into something else.
- Forming the attorney-client relationship. The agent can book a consultation; it cannot accept a client or imply that representation has begun.
- Conflict clearance. Software can search; a lawyer must decide. An auto-clear is a defect, not a time-saver.
- Deadline calculation. Statute of limitations and filing deadlines belong to a human docketing process with a second check. Never let an agent compute a legal deadline from a narrative.
- Privilege determinations and production decisions. What is responsive, what is privileged, and what is produced are legal judgments.
- Fee quotes and settlement authority. Both create enforceable expectations. Keep them with the attorney.
If a task would require a license to do badly, it requires a license to do with software. The agent can prepare the work; it cannot own the judgment.
Guardrails and compliance
Unauthorized practice of law. The agent must identify itself as a firm assistant, not an attorney, and use a fixed refusal when asked for advice: it can explain that it will collect information for an attorney to review. Test this with direct prompts — "just tell me if I should sue" — and check that the refusal holds.
Confidentiality and privilege. Client information may only go to a vendor with a written confidentiality agreement, no training on firm data, encryption in transit and at rest, access controls, and a defined deletion schedule. Some state bars also require client notice or consent before using generative tools on client matters; check your jurisdiction and the firm's malpractice carrier.
Competence and supervision. Under the ABA Model Rules, a lawyer must understand the tool well enough to supervise it and remains responsible for its output. That means a named reviewer, a documented review step, and a policy on which tasks are permitted. "The software did it" is not a defense.
Communication. Status updates must be accurate and timely. An agent that sends a confident but stale update is worse than no update. Pull from the matter file at send time, and require approval before anything leaves the firm.
Jurisdiction limits. Intake should screen out matters in states where no attorney is licensed. A friendly agent that collects a detailed narrative from an out-of-state caller can create an expectation the firm cannot meet.
Advertising rules. Automated outreach and testimonials fall under the same rules as any other communication. No guarantees, no misleading claims, and required disclaimers where your state mandates them.
Logging and litigation hold. Keep a record of what the agent said and when. Build the ability to export and preserve those records if a matter goes to litigation. Deletion policies must yield to a hold.
What it costs
| Line item | Range | Notes |
|---|---|---|
| Intake and status agent | $100–$500 / month | Voice adds per-minute cost; web and chat sit at the lower end |
| Document processing | $0.02–$0.10 / page | Varies with model choice and whether OCR is required |
| Implementation and integration | $2,000–$12,000 one-time | Practice management and document system integration drives the top of the range |
| Security review and policy | $500–$3,000 one-time | Vendor review, retention policy, and staff training |
| Likely payback | 2–6 months | Depends on matter value; one retained matter often covers a year of software |
Document-heavy deployments should be modeled per page, not per seat, because the variable cost is where the surprises live. Run a 200-page sample through the vendor's pipeline before you sign and count the citation errors yourself. The cost guide covers the contract terms that turn a low subscription into an expensive year.
A 30-day rollout
Week 1 — scope and policy. Pick one practice area and one jurisdiction. Write the list of topics the agent may not discuss. Have the managing partner approve the review workflow and the vendor's confidentiality terms. Confirm with your malpractice carrier whether client notice is expected.
Week 2 — build the intake flow. Script the questions, the refusals, and the escalation rules. Connect the calendar and the conflict database. Test with the firm's own lawyers posing as difficult callers, including one who asks for a legal opinion three different ways.
Week 3 — shadow mode. Run the agent alongside the existing intake process. Every inquiry gets a human response and an agent transcript. Compare them. Fix the questions that produce vague answers and the refusals that sound robotic.
Week 4 — live with review. Turn on after-hours intake with a callback promise, and start drafting status updates for one attorney's matters. Review every outbound message before it sends. Measure consultation bookings and the time saved on status calls.
A document summary that reads well is not the same as a summary that is right. Ask the vendor to summarize a document you already know, with page cites, and check three specific facts. If the citations drift, the tool is not ready for client work regardless of how polished the prose is.
The three metrics that prove it worked
- Intake-to-consultation rate. Qualified inquiries that become booked consultations. Track it with the conversion from consultation to signed engagement, so volume gains do not hide a quality drop.
- Time to first response. Median minutes from inquiry to a human or agent reply, including nights and weekends. This is the number that wins matters against firms of similar size.
- Unbilled status-call minutes. Hours per week spent answering "any update?" questions. If it does not fall, the agent is drafting updates nobody reads or sending them too late.
Add a quality guardrail you review monthly: the number of agent-drafted messages an attorney substantially rewrote. Rising rewrites signal that the tone or the source data has drifted.
Frequently asked questions
Can an AI agent perform a conflict check for a law firm?
It can prepare one. The agent can collect the full legal names of parties and related entities, search the firm's conflict database, and produce a report that highlights potential matches. An attorney must review and clear the conflict; software cannot make that call, and an auto-clear is a malpractice risk rather than a feature.
Does using an AI agent violate attorney-client confidentiality?
Not automatically, but the arrangement matters. Client information should only go to a vendor with a written confidentiality agreement, no training on firm data, encryption, access controls, and a defined retention period. The lawyer remains responsible for competence and supervision, so the firm needs a policy on what may be submitted and who reviews the output.
How much does an AI agent cost for a small law firm?
Intake and status-update agents typically run $100 to $500 per month in software. Implementation ranges from about $2,000 for a configured intake flow to $12,000 for a build tied to the practice management and document systems. Document-heavy work adds usage fees that are usually billed per page or per document.