What AI Do Law Firms Use — and What Does It Cost?

By Jude Lee · · Comparison

Attorneys and a paralegal reviewing an AI-assisted document workflow on a laptop in a small law firm conference room

The three layers of AI most firms end up running

When people ask what AI most law firms use, they expect a single product name. In practice, a working small-firm stack looks like layers that do different jobs:

Layer 1 — the general assistant. Claude, ChatGPT, or Microsoft Copilot. Used for drafting, summarizing, rewriting, turning messy notes into a clean memo. Cheap per seat, immediately useful, and the layer most firms adopt first — often informally, before anyone approves it.

Layer 2 — AI inside the software you already own. Practice-management and document platforms have bolted generative features onto their existing products: as of early 2026, that includes Clio Duo in Clio and AI features marketed in MyCase and Smokeball, plus drafting and summarizing features in document systems. This paragraph is the fastest-rotting one in the article — check each vendor’s current product documentation before assuming a feature exists in your plan tier, because what’s bundled this year may have been an add-on last year.

Layer 3 — specialist legal AI. CoCounsel (Thomson Reuters, built on the Casetext acquisition), Lexis+ AI, and Harvey are the names that come up most. These are productized around legal research, deposition and document analysis, and firm-wide deployment. Harvey in particular is positioned toward large firms and legal departments; pricing is quote-based rather than published.

Layer 4 — the connective layer. Agents and integrations that let an assistant read and act on your matter data. This is the layer almost nobody sells off the shelf in a form that fits a 12-lawyer firm, and it’s where the interesting gains are.

Task wins plateau; workflow wins compound

Summarizing one document faster is a task win, and task wins flatten out quickly. The durable gains come from redesigning a whole workflow — intake to matter open, discovery receipt to review protocol, deposition to summary to issue chart.

That’s the practical difference between a chatbot and an agent. A chatbot answers when asked. An agent carries out a multi-step job: fetch the file, apply your review standard, write the output, put it back where it belongs, flag exceptions for a human. Our take on what actually works versus hype in 2026 goes deeper on that line.

Worth saying early: the redesign sometimes contains no AI at all. A document-assembly template, a one-page checklist, or a deterministic rules engine that fires the same way every time is cheaper to run, easier to audit, and doesn’t hallucinate. Reach for AI where the input is genuinely unstructured — not where a form field would do.

A tool that saves ten minutes per document is nice. A workflow that guarantees every document gets reviewed the same way is worth more, and it’s harder to buy.

Pricing models you’ll actually encounter, as of early 2026:

Questions to put in writing before you sign: Is client data used to train models? Where is it stored and for how long? Which seats does the license cover? What happens to your prompts, skills, and configuration if you leave? Is the AI feature included in your current plan tier or a separate SKU?

seats × monthly price × 12
Recurring software cost — plug in your own quote
Formula — supply your own inputs
setup hours × loaded cost
Implementation cost most firms forget to budget
Formula — supply your own inputs
hours recovered × your realized rate
Upside, only if the hours get reallocated to billable or business-development work
Formula — supply your own inputs

That last line is the one people skip. Recovered hours are not revenue until someone bills them or uses them to sell work. If a tool saves your paralegal four hours a week and those hours go into inbox drift, the ROI is zero. Our payback model walks through the arithmetic with your own inputs.

What an “AI lawyer” actually costs — and why the phrase is wrong

There is no product you can hire that practices law. Consumer-facing services that generate legal documents are software; jurisdictions regulate unauthorized practice of law, and responsibility for advice stays with a licensed lawyer.

For a firm, the honest cost question is: what does it cost to add AI capability to my existing team? Software subscription, plus implementation, plus supervision time, minus whatever it displaces. The supervision time is real and permanent — ABA Formal Opinion 512 (2024) addresses generative AI and touches competence, confidentiality, client communication, supervisory duties, and fees. Read it directly, and check your own state bar’s guidance, which may go further. If you’re billing clients for AI-assisted work, confirm how you charge for it with your state bar and, where the stakes are real, with ethics counsel.

Does AI change what you can charge?

Two questions come up constantly: whether a $900 hourly rate is high, and how a lawyer gets to $500,000 a year. Both are really questions about rate, utilization, and realization.

On rate: rates at that level are generally associated with senior, specialized practice rather than general small-firm work, but the spread by geography, practice area, and seniority is enormous. Rather than trusting anecdotes, look at a published benchmark for your market — the Clio Legal Trends Report and your state bar’s economics survey, where one exists, are the usual starting points. AI doesn’t raise your rate; positioning, expertise, and client mix do.

On income: the math is roughly billable hours × realization × effective rate − overhead. AI moves those levers unevenly. It can raise billable hours by removing non-billable drag (intake admin, file organization, status emails). It can lower per-matter cost on fixed-fee work, which is where the effect is usually largest. Run the formula with your own numbers; anyone quoting you a dollar figure for “what AI adds” is guessing.

Where agents and MCP fit

The layer that off-the-shelf products serve worst is your own data. MCP — the Model Context Protocol, an open standard — is the plumbing that lets an assistant like Claude read from and act in the systems you already run: practice management, document storage, calendar, billing. It isn’t a product you buy; it’s a way of connecting things, with permissions you define.

Two concrete shapes for a small firm:

  1. Connect the assistant to your existing platform. If you’re on Clio, connecting an assistant so it can pull matter context and draft in place is a contained first project — see our walkthrough on connecting Claude to Clio.
  2. Build a small custom MCP server over the data no vendor exposes well — your closed-matter archive, your form bank, your conflicts history — so an assistant can answer questions against it under your access rules.

On top of either, skills (packaged, reusable instructions) make output consistent: one skill for your deposition summary format, one for your first-pass review standard, one for your matter-opening memo. Same job, same way, every time.

Build vs. buy, read fairly

Buy off the shelf
Fits when your need is common (research, summarization, first-pass review), you want support and someone else’s security posture, and you have no in-house technical capacity. The cost: vendor lock-in and roadmap dependence — your workflow moves when the vendor decides it moves, and export options may be thin.
Build custom
Fits when the work depends on your firm’s own data and conventions, you’ve hit a repeated integration wall, and you can name the process precisely enough to write it down. The cost: ongoing maintenance as APIs and models change, plus key-person risk if one person understands the build.

Most firms should buy first and build narrowly. A custom agent that automates an undocumented process will faithfully automate the chaos. If you can’t write your intake or review process on one page, that’s the first project — not the AI. We covered the fuller decision in when a firm outgrows off-the-shelf software.

A 30-day way to decide

  1. Inventory what you're already paying for

    List the AI features included in your current practice-management and document tiers. Firms routinely buy a third tool that duplicates something bundled.

  2. Pick one workflow, not one tool

    Choose a repeated, annoying, non-strategic process — deposition summaries, discovery intake, matter-opening packets. Write down how it works today, step by step.

  3. Run a two-week bake-off on real work

    Same three matters, same reviewer, two candidate approaches. Score accuracy, rework required, and how long the human check took — not how impressive the demo felt.

  4. Set the human checkpoint before rollout

    Decide explicitly what a lawyer verifies every time: citations, dates, privilege calls, anything going to a client or a court.

  5. Price the full cost and the reallocation

    Subscription + implementation + supervision time, against hours recovered and where those hours will actually go. If you can’t name the destination for the recovered hours, delay the purchase.

What stays human

Agentic tools are good at retrieval, first passes, consistency, and volume. They fail quietly — producing a confident, wrong output that looks identical to a right one. Two failure patterns to watch for specifically: a citation that is formatted perfectly and reads plausibly but points to a case that doesn’t exist (courts have sanctioned filers over exactly this), and a contract summary that captures the main obligation while silently dropping a carve-out or exception that changes the answer. Neither looks wrong on the page; both are caught only by someone who checks the source.

That’s why privilege calls, legal conclusions, court filings, and client communications need a named human owner — and why a plain rule-based automation or a well-configured piece of SaaS is often the smarter, cheaper call.

The firms that get the most out of this aren’t the ones with the longest tool list. They’re the ones that picked one workflow, wrote it down, automated the repeatable middle, and kept a lawyer squarely on the judgment calls.

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