Legal AI cost calculator: what a seat licence really costs your firm.
Per-seat pricing looks small next to a partner's billing rate and large next to almost nothing else. Enter the quote on your desk and see the three-year number, the seats you are paying for and not using, and the month where building your own overtakes renting.
Your firm and your quote
Nothing is sent anywhere. The numbers stay in your browser.
Over 36 months
How sensitive is that break-even to the build estimate?
A single break-even month reads more precise than it is. Same licence terms, build cost moved up and down by half:
| Build cost | Amount | Break-even |
|---|---|---|
| Half | $90,000 | month 2 |
| As entered | $180,000 | month 5 |
| Half again | $270,000 | month 8 |
Estimates for planning, not a quote. Time-recovery and adoption inputs are assumptions you set, not benchmarks we assert. Vendor prices are whatever you enter from your own quote.
The short answer
Legal AI is sold per seat per month, usually on a twelve month minimum, often with a seat floor and a one-off implementation fee. For a firm of any size the three-year number is considerably larger than the monthly quote suggests, and at the end of it you own nothing. Building costs more on day one and less by a date you can calculate. That date is the only number that matters, and it depends entirely on your headcount, your quote and your term.
Four things a seat quote leaves out
The quoted per-seat figure is rarely what a firm ends up paying. Four line items account for most of the difference, and the first two are the ones that move the answer.
- Seat minimums. If the floor is twenty seats and you have twelve fee earners, you are buying eight licences for people who do not exist. The calculator makes this explicit rather than folding it into a total.
- Shadow seats. Licensed is not the same as used. A seat that nobody opens still bills. Your real cost per active user is the number worth defending to a partner committee.
- Implementation. A one-off onboarding, integration or training fee, paid in year one and easy to leave out of a monthly comparison.
- Renewal uplift. Year one pricing is an acquisition price. Ask what the renewal looks like before you sign a term you cannot exit.
Why we do not publish vendor prices here
Most vendors in this category do not disclose per-seat pricing publicly. Any table we published would be assembled from secondary reports, would be wrong for firms quoted differently, and would be stale within a quarter. We would rather the tool be exactly right for your quote than approximately right for a hypothetical firm. Enter what you were quoted and the model is true for you.
When licensing is plainly the right call
Building is not always correct and this tool will tell you so. If the workflow is generic legal research, if the vendor already does it better than a bespoke build would, if your user count is small, or if you need something running next month rather than next quarter, buy the seats. The arithmetic turns when the user base grows, when the term lengthens, or when the workflow is specific to how your firm actually works.
Ownership is the part to get in writing
If you do build, ownership needs to be explicit before anyone writes code: the pipelines, the prompts, the model configuration, the evaluation set and the right to take all of it elsewhere. A build you cannot leave with is a licence wearing a different hat. Gaper builds production AI systems that the client owns outright, which is the only reason the build side of this calculator is worth modelling at all.
Questions firms ask before they sign
How much does legal AI cost per lawyer?
It depends on the tier you are quoted, and most vendors in this category do not publish per-seat pricing at all. Reported mid-market pricing for the research-grade tools sits in the low thousands per seat per year, and drafting-only tools sit well below that. Because the numbers move and are rarely public, this calculator asks you to enter the quote you were actually given rather than assuming a figure on your behalf.
Why does this calculator not show me Harvey or CoCounsel prices?
Because publishing a competitor price we cannot verify would be guesswork presented as fact, and it would go stale the moment a vendor changed a rate card. Your quote is the only number that is true for your firm. Enter it and the model is exact for your situation instead of approximately right for a hypothetical one.
What does a per-seat legal AI quote usually leave out?
Four things account for most of the gap between the quoted number and the real one: a one-off implementation or onboarding fee, a seat minimum that bills you for people you do not have, annual uplift at renewal, and the training time before anyone is productive. The first two are inputs in this calculator because they are the ones that change the answer most.
What is a shadow seat?
A licensed seat nobody uses. Firms routinely buy a seat per fee earner and see a fraction of them active after the first month. You pay for the licence either way, so the honest cost per active user is higher, often much higher, than the cost per licensed user. The calculator shows this as a separate annual figure.
At what firm size does building cost less than licensing?
There is no universal crossover, which is the point of modelling it. Per-seat cost scales with headcount while a build is mostly fixed, so the larger the user base and the longer the term, the more the arithmetic favours building. Enter your own numbers and the calculator returns the month where cumulative build cost falls below cumulative licence cost, plus a sensitivity range around that month.
Should my law firm build or buy AI?
Buy when the workflow is generic, the vendor already does it well, and your user count is small enough that seats stay cheap. Build when the workflow is specific to how your firm works, when the data cannot leave your control, or when you have enough users that per-seat pricing compounds. Many firms end up doing both, licensing a research tool while building the workflow that is actually their differentiator.
Who owns the system if we pay someone to build it?
You should. That means the code, the prompts, the pipelines, the model configuration and the evaluation set. Ask any implementation partner to put ownership and exit rights in writing before work starts. A build you cannot take with you is a licence with extra steps.
What the evidence actually says about legal AI spend
Two things are well documented and worth separating from vendor marketing. The first is that adoption is now mainstream rather than experimental: the American Bar Association's annual technology survey has tracked a steep rise in firms reporting AI use, and the profession-wide question has moved from whether to adopt to how to pay for it. The second is that the failure mode is not the technology, it is verification. Courts have sanctioned lawyers for filings containing fabricated citations produced by general-purpose chatbots, starting with Mata v. Avianca in 2023, and the count of documented cases has grown steadily since. Both facts point the same way: the spend is real, and so is the cost of skipping the review step.
That matters for this calculator because it is the argument against the cheapest option. Handing lawyers a consumer chatbot is nearly free and carries the highest tail risk. A seat licence buys verification infrastructure. A build buys the same thing plus ownership. The question is not which is cheapest, it is which is cheapest at your size once you account for the control you need either way.
- ABA Legal Technology Survey Report, on adoption rates across firm sizes.
- Mata v. Avianca, Inc., the sanctions order that made verification a professional-responsibility issue.
- AI Hallucination Cases database, a running tally of filings containing fabricated citations.
How to run this with a partner committee
The output of this calculator is an argument, not a decision. Three moves make it land. Bring the three-year number rather than the monthly one, because the monthly figure is designed to feel small and the committee is approving the three-year commitment. Bring the shadow-seat figure, because it reframes the discussion from price per seat to price per person who actually uses it, and that is the number nobody has calculated. And bring the sensitivity range rather than the single break-even month, because a partner who has ever commissioned software knows the first estimate moves, and a range you volunteered is far more credible than a point estimate you have to defend.
Expect one objection reliably: that a build carries delivery risk a licence does not. That is true and worth conceding. The counter is that the licence carries renewal risk, seat creep and zero terminal value, and that those are certainties rather than risks. Both paths have exposure; only one of them ends with an asset.
Method and limits
Every figure in the model is one you entered. Licence cost is billed seats, respecting the minimum, times your quoted rate over the term, plus the implementation fee. Build cost is your one-off estimate plus annual run cost amortised across the same months. Break-even is the first month where cumulative build cost is at or below cumulative licence cost, and the sensitivity table re-runs it with the build estimate halved and increased by half, because a single month reads more precise than this exercise can honestly be. Nothing is transmitted anywhere; the calculation runs in your browser.
Want agents like these in your stack?
Book a free assessment, we'll map where an AI agent creates real leverage in your workflows and scope the first one to ship.