What AI in a law firm really costs
The price per user is the smallest item. Infrastructure, integration, running the thing and the time spent checking the output cost more.
Ask what AI costs in a law firm and you usually get a figure per user per month. It is rarely wrong and rarely complete. Model usage itself is the smallest item and the most predictable one. Everything standing next to it is where the money goes.
What the usage itself costs
Billing runs one of two ways. Under usage-based billing the firm pays for the text processed, so per request and response. That item moves with the workload and stays modest for typical firm tasks, because extracting a case header, checking a deadline or routing incoming post are short operations.
Under licence models the firm pays per head per month, no matter how much work gets done. That makes the invoice predictable and the purchase tends to be more expensive, because you also pay for the people who barely touch the tool. Which route is cheaper is not decided by the list price but by how unevenly usage is spread across the firm.
Infrastructure costs money even when nobody uses it
The second block arrives before the first request runs and keeps running when no more come. It covers private network paths, log storage, customer-managed encryption keys and test environments.
An example from a law firm setup on AWS (Amazon Web Services, Amazon’s cloud platform): 15 private network endpoints had been set up there for model access (VPC endpoints via PrivateLink, Frankfurt region). They cost around 104 US dollars a month in provisioning fees alone, whether or not a single request ran. They were dismantled again because at that scale the benefit did not carry the base fee.
That is the typical case. It is not the requests that are expensive, it is the permanently provisioned infrastructure that gets set up once and never looked at again.
Setup, integration and running the thing
The third block is the setup, integration with the existing systems included. A model that cannot reach the practice management software, the document management system and the mailbox shifts the work into copy and paste. Integration is the part that costs time, whether the firm buys it in or does it itself.
The fourth block is operation: updates, managing permissions when staff change, reviewing the logs, briefing new colleagues. This block never ends, which is why calculations leave it out most often. What appears as a one-off setup in the first quote is a recurring line item in year two.
The item that appears in no calculation
AI output that has to be checked moves work around, it does not remove it. If you are calculating the benefit you have to net off the review effort, and for legal output that effort is rarely small. A text that looks thorough and is wrong in one place costs more review time than one that is obviously useless. Why the plausible output is the expensive one is a topic of its own.
For the calculation that gives a simple rule: the gain is the processing time saved minus the review time. Tasks where the review takes almost as long as the work itself do not pay for themselves, however impressive the output looks on first reading.
How to tell in advance whether it pays
Not from the tool, from the task. A task carries its own weight if it comes up often, is clearly bounded and is not billed today. Often, because setup effort only amortises through repetition. Clearly bounded, because otherwise there is no way to tell whether the output is correct. Not billed, because no revenue disappears there when the work gets faster. That is exactly why the entry point lies in non-billable work and not in the pleading.
The same calculation covers how firm knowledge is made accessible. A method that reads several documents one after another for every question triggers more model calls than one that hands over pre-selected excerpts; in exchange, running the search database drops away. The post on RAG and agentic search deals with that relationship.
That leaves the point that costs least and saves most. Cost alerts belong in place before the first access is set up, not after the first surprising invoice has arrived. A monthly threshold and a named person who follows up on a warning are enough. The same move is part of hardening a law firm AWS account anyway, there for security reasons. For costs it applies for the same reason: the cheapest time to learn about something going wrong is while it is still small.
The content of this blog was created partly with the support of AI.