Agentic AI vs Generative AI in Law Firms: Where Each Fits

By Jude Lee · · Comparison

Attorneys and a paralegal reviewing case documents on a laptop in a small law firm conference room

Thomson Reuters publishes a useful primer on agentic AI vs. generative AI, and the distinction is worth pinning down — because vendors are now selling both under one label, and the risk profile is not remotely the same.

The one difference that changes your risk profile

Generative AI answers. You ask; it produces a draft, a summary, a comparison; you read it, fix it, and decide what to do. The AI’s output is inert until a human moves it.

Agentic AI acts. Given a goal, it plans a sequence of steps, calls tools, checks results, and continues — creating a contact, opening a matter, writing a file to your document management system, sending a calendar invite. The model is doing the same underlying reasoning; the difference is that it now has hands.

That is the entire distinction, and it’s not academic. A hallucinated sentence in a draft costs you two minutes of proofreading. A hallucinated action — the agent opens a matter under the wrong client, or saves privileged work product to the wrong folder — costs you an afternoon and possibly a difficult phone call.

Generative AI (assistive)
Drafting, summarizing, comparing, brainstorming. Output lands in front of a human before anything happens. Setup: minutes. Failure mode: a bad draft you catch on review. Best for judgment-heavy, low-volume, high-variance work.
Agentic AI (takes actions)
Multi-step tasks across your systems: read → decide → write → notify. Setup: integration, permissions, logging, testing. Failure mode: a wrong action already recorded in a system of record. Best for repetitive, well-defined, high-volume work with a clear undo.

What the two buckets look like in practice

What follows is our own observation from firm deployments as of 2026, not survey data — treat it as a starting hypothesis to test against your own practice.

The generative side is the part most firms reach first, and it’s mostly unremarkable: first-draft correspondence, plain-language client explanations, summarizing a long transcript, comparing two versions of an agreement, reorganizing a messy fact pattern into a chronology. This is where value tends to show up in week one, and where the practical wins and hype for law firms separate quickly.

The agentic side, from what we see, is younger and narrower in production. Where it genuinely works today, it tends to look like: intake agents that capture a lead, run a preliminary conflicts search, and stage a matter for attorney approval; discovery triage agents that classify and route large document sets; abstraction agents that pull key terms out of PDFs and write them into structured fields. What these share is volume, repeatability, and a reviewable output — not autonomy over judgment calls. Our discovery triage playbook goes deeper on that shape of work.

Generative AI fails in front of you. Agentic AI fails behind you, in a system of record. Design for that difference or don’t deploy it.

A sorting rule for deciding what to hand to an agent

People search for the “80/20 rule for lawyers,” usually meaning the Pareto heuristic that a minority of clients, matters, or tasks drive the majority of the value. It’s a rule of thumb, not a finding — but it’s the right lens here. Applied to automation: don’t ask what AI could do, ask which handful of recurring tasks consume the most non-billable staff time in your firm this quarter, then check each against four filters. The thresholds below are my own working heuristics, not established criteria — tighten or loosen them to fit your practice.

Volume. Does it happen at least weekly? One-offs are generative-AI territory; there’s rarely a return on integrating them.

Definition. Could you write the steps down for a new paralegal on one page? If the answer is “it depends on the matter,” an agent will guess — and guessing is what you’re trying to eliminate.

Reversibility. If the agent gets it wrong, can you undo it in a few minutes? Drafting into a review queue: reversible. Sending a client communication or transmitting a court filing: not really.

Verification. Can a human check the output faster than doing the work? If review takes as long as the task, you’ve moved labor, not removed it.

Anything failing two or more filters stays generative — a human in the loop with an assistant, not an agent with credentials.

Fabricated citations remain the headline risk; courts have sanctioned lawyers for filing briefs containing non-existent cases, the best-known early example being Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y.), where the court imposed sanctions in June 2023. Verify every authority against Westlaw, Lexis, or the primary court record — a point we cover in our comparison of AI legal research tools.

Beyond hallucination: confidentiality exposure when client data flows to a tool you haven’t vetted; over-reliance by junior staff who can’t yet spot a plausible-sounding error; and silent drift, where an agent quietly starts mishandling an edge case and nobody notices for a month because nobody’s reading the logs.

On the ethics side, ABA Formal Opinion 512 (2024) addresses generative AI use by lawyers, covering competence, confidentiality, communication with clients, and fees. Model Rule 1.1, Comment 8 has long framed technology competence as part of the duty of competence. Your state bar may have its own opinion that differs in detail — check your jurisdiction’s guidance directly rather than relying on a summary, and confirm anything with real client-confidentiality stakes with counsel who knows your state’s rules.

How an agent actually reaches your systems

The missing piece for most firms isn’t the model — it’s the wiring. MCP (the Model Context Protocol, an open standard for giving an AI governed access to specific data and tools) is one way to solve it. Instead of pasting matter details into a chat window, you connect an assistant like Claude to a defined set of operations: look up a matter, list documents, create a draft. You choose which operations exist; anything you don’t expose, the agent cannot do. MCP is not the only path — vendor-native AI integrations, plain API scripts, iPaaS tools like Zapier or Make, and even RPA can connect the same systems; MCP’s advantage is that it’s model-agnostic and scoped per tool, which matters more as you add assistants.

Two routes if you go that way. Use an existing connector where your vendor offers one — see our walkthrough on connecting Claude to Clio via MCP. Or build a custom MCP server over your own data when your workflow spans systems no vendor connects. Building is real software: scoping, auth, testing, maintenance. It’s justified when the workflow is core to how you make money and no product fits — not because it sounds modern.

Alongside that, define skills: packaged instructions that teach the assistant to do one job the same way every time — your deposition summary format, your intake script, your document review rubric. Skills are often the cheapest reliability upgrade available, since they require no integration at all.

The “best AI program” question, answered honestly

There isn’t one. The right stack depends on what you already run. If your practice management platform ships competent AI features, use them first — that’s the lowest-friction path, and the vendor agents vs. custom build tradeoff rarely favors building until you’ve exhausted what you’re already paying for. For general drafting and analysis, a frontier assistant with a proper business agreement covering data handling is usually enough. Reserve custom agentic builds for the one or two workflows where an off-the-shelf tool genuinely doesn’t fit.

Sizing the payback with your own numbers

Searches like “do lawyers make $500,000 a year” show up next to AI-automation queries because the underlying question is really about rates and leverage. Lawyer compensation varies enormously by practice area, geography, and ownership structure — if you want a defensible wage baseline, the U.S. Bureau of Labor Statistics Occupational Outlook Handbook publishes lawyer earnings data. For automation math, don’t use a national figure at all. Use your own.

hours saved/week × working weeks × your realized rate
Annual recovered-capacity value — we use 46 working weeks after holidays and vacation; substitute your own
build + license + maintenance
Total first-year cost, not just subscription
1 workflow, 90 days
A pilot scope small enough to actually finish

Three buckets belong in the model: recovered hours (only counted if they’re genuinely reallocated to billable or business-development work — otherwise they’re comfort, not revenue), captured revenue (leads answered faster, time entries that stop evaporating), and errors avoided (harder to quantify; estimate conservatively or leave at zero). Our automation ROI walkthrough sets up the arithmetic; the inputs must be yours.

  1. Log where the week actually goes

    For two weeks, have staff tag recurring non-billable tasks. You’re looking for frequency, not drama.

  2. Split the list into assist vs. act

    Judgment-heavy and variable work goes to a generative assistant with a defined skill. Repetitive, well-bounded work becomes an agent candidate.

  3. Run the agent read-only first

    Let it propose the actions it would take, for two weeks, without taking them. Compare its proposals against what your team actually did.

  4. Grant one write action, with logging

    Draft into a review queue, not into a filing. Expand only after the proposal log is boring.

  5. Review the exceptions monthly

    The value is in the cases the agent flagged or fumbled. If nobody owns that review, the deployment decays quietly.

Where is your firm losing billable hours?

Get a free automation audit: we map your intake-to-invoice workflow and show you exactly what's worth automating — before you spend a dollar.

Get a free automation audit

Related guides