The state of agentic AI in the legal domain

The first wave of legal AI made lawyers faster at asking questions. Tools like Harvey, CoCounsel, and dozens of document-review startups gave attorneys a chat interface where they could summarize a contract, draft a clause, or search case law in natural language. This was genuinely useful. It reduced the time for specific tasks from hours to minutes.

But it left the actual work unchanged. A lawyer still had to open the tool, type the question, read the answer, decide what to do with it, copy it somewhere, and move on to the next task. AI was a faster way to handle individual steps. The workflow — the sequence of decisions, actions, handoffs, and follow-ups that defines how legal work actually moves — was still entirely manual.

The next wave is different. Agentic AI systems do not wait for a prompt. They execute multi-step workflows autonomously: making phone calls, sending emails, processing documents, updating CRMs, following up with insurance carriers, and escalating to humans only when legal judgment is required. The AI is not answering questions. It is doing work.

Plaintiff law is uniquely suited for agentic AI because the work is high-volume, process-driven, and repetitive across cases. A personal injury firm handling 200 active cases is running roughly the same sequence for each one: intake, qualification, letter of representation, police report, medical records, treatment tracking, demand preparation, negotiation, settlement, and closure.

Each stage involves predictable tasks: calls to make, emails to send, documents to request and process, CRM fields to update, deadlines to track, and follow-ups to manage. The legal judgment — case evaluation, litigation strategy, settlement decisions — is concentrated at a few critical points. The rest is operational discipline.

This is exactly the pattern agentic AI handles well. The system needs enough structure to know what to do next, enough intelligence to handle variation across cases, and enough governance to escalate when it reaches the boundary of its competence. Plaintiff firms provide all three: clear workflows, rich case data, and well-defined escalation criteria.

Not all agentic systems are created equal. The architecture determines whether the system is a demo or production infrastructure. The key requirements are orchestration, durability, tool integration, governance, and observability.

Orchestration means a central agent that can break a workflow step into tool calls, execute them in sequence, handle errors, and determine the outcome. AutoCounsel uses a Moderator Agent built on the OpenAI Agents SDK that can execute up to 20 tool calls per step, selecting from 17 specialized tools for email, voice, documents, CRM, web automation, and more.

Durability means the system survives crashes, retries failed operations, and resumes where it left off. AutoCounsel uses Temporal.io for durable workflow execution — every background operation runs as a Temporal workflow with automatic retries, timeouts, and crash recovery. This is not a background thread that dies if the server restarts.

Tool integration means the agents actually do things in the real world. They send emails through Microsoft Graph. They make phone calls through Twilio. They update records in Litify and Filevine. They purchase police reports through state portals. They are not generating text for a human to act on — they are acting.

Governance means the firm stays in control. Workflows include approval gates where attorneys review and approve before the system proceeds. Confidence thresholds trigger human escalation. Audit trails record every tool call, every decision, and every outcome. Role-based access controls limit who can design, execute, or modify workflows.

Across the legal AI landscape, several patterns are emerging. Companies that will define the next era of legal operations share common characteristics, even as they differ in approach and focus.

Case-aware context is becoming table stakes. The agent needs to know the parties, the insurance data, the documents on file, the communication history, and the execution state before it acts. Without context, agents hallucinate or take inappropriate actions. The best systems build a rich context window before every decision.

Source-grounded outputs separate production tools from toys. When an agent generates a demand letter, it should cite specific medical records, reference exact treatment dates, and calculate damages from actual bill amounts — not fabricate plausible-sounding numbers. When it writes an email, it should reference prior correspondence in the thread.

Voice AI is the new frontier. The most immediate ROI for plaintiff firms is answering calls — the one moment where speed directly converts to revenue. AI voice agents that can conduct structured intake interviews, qualify leads, schedule consultations, and capture clean data are replacing call centers and answering services.

Workflow execution separates platforms from point solutions. A tool that summarizes a document is useful. A platform that takes the summary, updates the CRM, sends a follow-up email, flags missing records, and schedules the next review task is transformative. The value is in the chain, not the individual step.

The biggest barrier to adoption is not capability — it is trust. Law firms handle sensitive client data, operate under ethical obligations, and face malpractice liability for errors. Deploying autonomous AI agents in this environment requires a level of governance that most AI companies have not built.

Firms need audit trails that record exactly what the AI did, when, and why. They need approval gates that pause execution for attorney review at critical decision points. They need workspace isolation that prevents data leakage between clients. They need role-based access controls. They need confidence thresholds that trigger human escalation when the AI is uncertain. They need compliance-grade language in their security documentation — SOC 2, HIPAA alignment, data encryption, access logging.

The right model is not full autonomy everywhere. It is governed autonomy: routine work moves automatically with high reliability, while legal judgment stays with the humans who are licensed to exercise it. The AI handles the operational discipline — the calls, the emails, the document processing, the CRM updates, the follow-ups — so attorneys can focus on the work that actually requires a law degree.

AutoCounsel was built around a specific thesis: let firms design how work should move, then let agents execute with visibility and control. The product architecture reflects this — a Moderator Agent coordinates specialized sub-agents for voice, email, documents, CRM, calendar, web automation, notes, research, and demand workflows. Each agent has clear capabilities and boundaries. The Moderator decides which to use, when, and in what order.

The roadmap extends this model. Demand package drafting — the most complex and highest-value task in pre-suit PI — is moving from manual assembly to agent-assisted orchestration: medical record discovery, gap detection, photo selection, cover letter generation, and attorney review flags. Settlement negotiation monitoring, treatment tracking, and provider communication are all becoming agent-driven.

The firms that adopt governed agentic systems will not just be more efficient. They will operate at a fundamentally different scale — handling more cases with the same team, responding faster to new leads, maintaining better documentation, and reducing the administrative burden that drives attorney burnout. The firms that wait will find themselves competing against operations that never sleep, never forget a follow-up, and never lose a lead because the phone went to voicemail.