Law Firm AI Pricing: Per-Seat vs Usage vs Custom Build

By Jude Lee · · Comparison

Law firm team reviewing AI software quotes and a cost spreadsheet in a conference room

The pricing experiment nobody announced, but everyone’s running

Open three legal AI quotes and you will likely get three different billing shapes. Some vendors still sell seats, the way practice management software always has. Others — particularly the document-heavy ones — price per matter, per document, or per credit. Check each vendor’s own pricing documentation rather than trusting a summary, including this one; this part of the market moves fast as of 2026.

The reason the old default is wobbling is structural. A seat license prices access. Agentic software produces work — a drafted discovery response, an abstracted lease, a conflict check run at 11pm against your practice management system. Those two things don’t scale together. Output varies enormously between a paralegal who routes a dozen tasks a day through a well-built agent and a lawyer down the hall who logs in twice a month. Charging both the same is either a bargain or a waste, and firms usually find out which after the annual renewal.

A seat you bought for a lawyer who opens the tool twice a month isn’t an AI strategy. It’s a subscription.

The three shapes you’ll be quoted

Per seat, per month. The default for most law firm AI software and for the AI features bolted onto practice management platforms. Predictable, easy to approve, easy to compare across vendors. Its failure mode is utilization: you pay for the headcount you licensed, not the work you got. It also quietly discourages the behaviour you want — nobody delegates a borderline task to a tool when they’re not sure the firm is “using its seats.”

Per unit of work. Per matter, per document, per page, or per credit/token. Structurally honest: a quiet month costs less. The failure modes are forecasting (your CFO hates a bill that swings) and unit definitions (does a 400-page medical record set count as one document or four hundred?). Ask. Get it in writing.

By outcome. Priced against a completed deliverable rather than access or volume. The hard part is who decides the deliverable is done. Say an agent returns a first-pass discovery response that is complete, on-format, and technically responsive — and the supervising partner throws it out because the strategy on that matter changed. Under outcome pricing, is that a billable deliverable or not? A vendor can warrant that the work was produced; it cannot warrant that it survived legal judgment. Push that exact scenario into the negotiation and see how the vendor answers. Treat outcome pricing as a question worth asking, not a category you can shop today.

Seats × rate × 12
Per-seat annual cost
Illustrative formula — use your own figures
Units/yr × cost per unit
Usage-based annual cost
Illustrative formula — use your own figures
Build + model usage + upkeep
Custom build cost shape
Illustrative formula — use your own figures

What a custom build actually costs

Here is the honest version, because this is where buyers get oversold in both directions.

A custom agent — say, Claude connected to your document and practice management systems through MCP, with a packaged skill that runs your firm’s discovery triage the same way every time — has three cost lines, not one. There’s a build cost, which is lumpy and front-loaded (scoping, the connector or MCP server, the skill, testing against real matters). There’s metered model usage, which behaves exactly like usage-based vendor pricing: heavy months cost more. And there’s maintenance — the line most firms underestimate. APIs change, court rules change, your templates change, and a skill that isn’t re-tested drifts out of usefulness.

Relative magnitude matters more than exact figures. The build and upkeep lines are largely fixed; the model usage line scales with volume. That gives you a usable threshold without inventing numbers: if the target task runs only a few dozen or a few hundred times a year, fixed build and maintenance costs rarely get amortized down far enough to beat a per-seat quote. Custom starts winning when the same bounded job repeats thousands of times, or when no vendor’s workflow matches how your firm actually drafts.

What you gain is marginal cost that doesn’t scale with headcount. Adding your sixth associate costs roughly what the fifth did in model usage and nothing in license. What you give up is a vendor’s roadmap, support desk, and the comfort of someone else’s security review.

Off-the-shelf legal AI (per seat)
Cost is flat, visible, and starts the month you sign. No engineering ownership. Fastest path to a result. Scales linearly with headcount — punishing if only a few people use it heavily, generous if everyone does. You inherit the vendor’s workflow, including the parts that don’t match how your firm actually drafts.
Custom agent build
Cost is front-loaded, then metered. Scales with work volume rather than seats. You own the workflow and the data path. Requires someone accountable for upkeep — internal or a vendor on retainer. Only pays back if the underlying task is high-volume, repetitive, and genuinely specific to your firm.

A cost model you can fill in yourself

Don’t compare list prices. Compare cost per unit of work delivered, then compare that against what the work costs you today.

  1. Pick one workflow, not 'AI'

    Choose a bounded job: deposition summaries, lease abstraction, first-pass discovery responses. Firm-wide AI spend is unbudgetable; one workflow isn’t.

  2. Count the units

    How many of these did you do last year? Pull the real number from your practice management system rather than estimating.

  3. Cost the current state

    Hours per unit × the loaded hourly cost of whoever does it now (associate, paralegal, or outsourced vendor). That’s your baseline — hours × rate, with your rates.

  4. Cost each option per unit

    Per-seat quote ÷ expected annual units. Usage quote × units. Custom build: (build ÷ years amortized) + (units × model usage) + annual upkeep.

  5. Add the review tax

    Every option still needs a human to check the output, and the specific failure modes are known: fabricated or mis-cited authority that reads fluently, and silent drift when your templates or the court’s rules change but the prompt or skill doesn’t. Neither announces itself. Estimate review minutes per unit and add them to all three columns. And price in the thing no billing model changes: responsibility for the work product stays with the supervising lawyer regardless of who or what drafted it.

  6. Decide what the recovered hours do

    Recovered paralegal hours only become money if they’re reallocated to billable or revenue-generating work — intake follow-up, collections, matter progress. Say so explicitly instead of assuming savings appear.

Our law firm automation ROI calculator walks the same arithmetic in more detail.

Questions that change the number

On commercials, get written answers to: What exactly is a unit, and who counts it? What happens at overage — hard stop or silent billing? Can unused credits roll over? Is there a minimum seat count, and can it go down at renewal? Does the price include the integrations we need, or is the connector to our document management system a separate line?

On data — which is a confidentiality question before it is a procurement one — ask: Does the vendor train models on our inputs or outputs, and can we opt out in the contract rather than a settings toggle? Where is the data stored and processed? Who at the vendor or its subprocessors can access matter content, and under what controls? Is there a confidentiality addendum (the analogue of a BAA in healthcare) covering client information? And on exit: what happens to our prompts, skills, and outputs? ABA Model Rule 1.6 and the duty of technology competence in Comment 8 to Rule 1.1 make these your questions to answer, not the vendor’s — and your state’s version of those rules governs, so check it.

That integration line is where a lot of small-firm deployments quietly die. A per-seat AI tool that can’t see your matter files is a very expensive chat window. Our comparison of Harvey, CoCounsel, and a custom agent goes deeper on where each sits for a firm under about fifty lawyers.

Billing clients for it is a separate question

How you buy AI and how you charge for AI-assisted work are different decisions with different rules. The American Bar Association’s Formal Opinion 512 (July 2024) addresses generative AI under the Model Rules, including fees and how costs may be passed to clients. Read the opinion itself, then check your own state bar’s guidance — states have issued their own ethics opinions and they don’t all say the same thing. Confirm with your bar’s ethics counsel before you change a fee agreement.

Where this lands for most small firms

An opinion, stated plainly: most firms under fifty lawyers end up hybrid, and that’s fine. Per-seat for the broad assistant everyone touches. Usage-based or custom for the one or two high-volume, firm-specific workflows where off-the-shelf output doesn’t match how you actually work. And — genuinely — nothing at all for the workflows where a template, a checklist, or a plain automation rule already does the job for less money and less risk.

The pricing model matters less than the discipline behind it. Price one workflow, measure it, and renew only what you can show is being used. As of 2026, that’s the only part of this market that isn’t changing every quarter.

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