On June 23, 2026, Legora moved its most capable product, Agent Pro, from seat-based licensing to consumption-based pricing. Customers now pay for the work the agent performs rather than for the number of people who can open it, and every run can be attributed to the matter that prompted it. Legora paired the change with a real-time dashboard, per-matter cost tracking, and spending thresholds that warn administrators before a limit is reached. The company frames this as the natural evolution from software you license to outcomes you buy, and other vendors will follow, because agentic usage varies too much from one matter to the next for a seat to absorb.
Most of the early commentary has treated this as a pricing-page story: enterprise costs will rise, budgets will be harder to forecast, and firms will need new controls. All of that is true. But consumption pricing also changes who controls the bill. When the cost of a tool tracks how it is used, the skill of the person using it stops being invisible. A lawyer who scopes a task well, grounds it, and verifies it in one pass will spend less than a lawyer who runs the same matter five times to get a usable answer. Under a seat license, the difference between those two lawyers showed up only in their work product. Under consumption pricing, it will show up on the invoice.
The meter is not new
Lawyers who trained before flat-rate research subscriptions became standard will recognize this immediately, because legal research used to work exactly this way. Westlaw and Lexis built their businesses on transactional pricing: a charge per search and a further charge for each document opened, with higher charges for material outside the firm’s plan. The model survives today in the transactional options both services still offer, where a single search may run on the order of $60 to $100 and each out-of-plan document adds to the total. For most of the history of computer-assisted research, running a query cost money, and running a sloppy query cost more.
That pricing produced a professional competence around controlling it. Law schools and firms taught cost-effective research as a discipline, and law librarians built guides devoted to it: start broad on a cheaper source, narrow with filters before opening documents, recognize the warning that a result sits outside your plan, and decline to open it unless you need it. The skill was not legal analysis in the ordinary sense. It was knowing how to get the same answer for less money, and it was valued precisely because the alternative appeared on a bill someone had to pay or absorb. Flat-rate subscriptions later buried that skill, because once research was all-you-can-eat, efficiency stopped having a price. Consumption pricing for AI brings the price back, and with it the relevance of knowing how to spend less.
What efficient use looks like with an agent
The content of the skill is different for an agent than for a Boolean search, but the logic is the same. An agent priced by consumption charges for the work it does, and the work it does depends heavily on how the task is put to it. A vague instruction produces a broad, expensive exploration; a poorly scoped task sends the agent down branches the lawyer doesn’t need; a prompt riddled with contradictions makes the model spend effort reconciling instructions that cannot be reconciled. I described that last failure in an earlier post on prompting: OpenAI reports that newer reasoning models expend reasoning tokens searching for a way to reconcile contradictory instructions rather than picking one at random. Under a seat license, those wasted tokens were the vendor’s problem, but under consumption pricing, they are a line item.
Efficient use, then, looks like the practices that already produce better output: state the outcome and the constraints rather than scripting every step, ground the task in the specific documents the agent should rely on, ask for the work in a form you can verify, and check it once at the end rather than re-running the whole matter because the first pass drifted. A lawyer who works this way spends fewer agent runs to reach a defensible answer. A lawyer who treats the agent as a slot machine, re-rolling until something looks right, spends more and often ends up with less, because each re-run is a fresh opportunity for a fabricated citation or an unverified inference to slip through. The expensive way to use these tools and the unreliable way to use them turn out to be the same way.
This is the part firms tend to underweight when they evaluate a consumption model. The variable in the cost equation is not only the agent’s capability or the vendor’s rate; it is the competence of the person directing it. A firm that rolls out Agent Pro without training its lawyers to use it economically has bought not so much a productivity tool as a metered cost, handed to people not yet able to control it.
Why efficiency benefits the client and the firm
Whether efficient use helps the client or the firm turns on how the firm bills for the tool. Suppose a firm passes Agent Pro consumption through to clients as a disclosed disbursement, the way it once passed through Westlaw charges. The lawyer who runs a matter efficiently produces a smaller charge, and the client pays less for the same result. The firm has given the client a reason to prefer it over a competitor whose lawyers burn three times the consumption to reach the same place, and the per-matter dashboard makes that comparison legible. Efficiency is no longer a virtue the firm quietly absorbs, but a number the client can see.
Now suppose the firm absorbs the consumption cost as overhead instead, recovering it through rates. Efficient use still pays, because every wasted run comes straight out of realization on a fixed fee or a rate the client will not let rise indefinitely. In the pass-through case, competence lowers the client’s cost; in the absorption case, competence protects the firm’s margin. Which way the benefit flows depends on the billing arrangement; whether there is any benefit at all depends on whether the firm’s lawyers know how to use the tool economically.
A third reason to build AI competence
The case for training lawyers to use these tools well is older than consumption pricing. Comment 8 to Model Rule 1.1 has, since 2012, required lawyers to keep abreast of the benefits and risks of relevant technology, and a lawyer who cannot describe a tool’s failure modes cannot supervise its output or vouch for the filing that results. That duty holds whatever the tool costs.
Competence also pays. A lawyer who uses these tools well produces sounder work faster, and clients have noticed: many now ask in their outside-counsel guidelines whether a firm’s lawyers are trained to use AI at all.
Consumption pricing adds a third incentive that may dwarf the first two. Agentic work burns tokens at every step, and the sums are not trivial: Harvey’s co-founder, Gabe Pereyra, has put the compute behind a single ‘draft a document’ query at roughly $20, and a contract review across 100,000 documents at about $20,000. A seat license hides those costs inside the vendor’s margin; consumption pricing, which Legora has now adopted and others may follow, delivers those costs to the firm, which pays for actual use. If Harvey has been heavily subsidizing its customers’ token costs, as many suspect, a shift to consumption pricing could cause a skyrocketing in technology costs. An incentive that began as an ethical duty, then grew into a competitive edge, has become a question of which firms can afford their own inefficiency.
This post draws on Legora’s announcement of consumption-based pricing and Law360 Pulse’s report on the switch; and a law-library guide to cost-effective electronic research on transactional Westlaw and Lexis pricing; and ABA Model Rule 1.1, Comment 8 on the duty of technological competence; and Gabe Pereyra’s per-task compute figures from the Sourcery podcast, as reported by BigGo Finance, alongside Artificial Lawyer on token costs and law-firm AI spend. It builds on earlier posts on how lawyers should prompt and the institutional floor.