The Legal AI Compute Cost Shift Report 2026: How Consumption-Based Pricing Is Hitting Legal Department Budgets — and Who Is Managing It Well

Research briefing — a review of public sources. The Legal Stack did not conduct a survey, benchmark or interviews for this piece.

What this briefing covers

The question behind this report is whether consumption-based pricing for legal AI is hurting in-house legal budgets and which departments are managing it well. The public evidence answers only part of that. No published survey we could find measures year-over-year legal AI spend after a pricing change, cost per matter with and without AI, or how budget predictability varies by governance model. An earlier version of this article contained figures of that kind. They were not drawn from any real dataset, and we have removed them.

What can be documented is this. Legal AI pricing is in flux and buyers find it confusing. Analysts warn that AI costs, which are driven by inference, are easy to underestimate. In-house leaders report that AI work often arrives without its own budget. A cost-management discipline for token spend is starting to take shape outside legal. The rest of this briefing sets out that evidence and then gives our analysis, labelled as such.

Pricing models are unsettled, not uniformly "consumption-based"

The original framing said the major vendors had all moved from seats to tokens. The public record does not support that. Bloomberg Law reported in June 2026 that AI legal software "flails on pricing models". Its sources described the following:

per-seat licensing as still the dominant software model
Agiloft testing consumption pricing for its Astra platform
Ivo charging according to the number of contracts loaded into its platform, and moving toward consumption pricing alongside seats

Lucy Bassli, a consultant to in-house teams, told Bloomberg Law that pricing is "all over the place," even within the same vendor. Nicole Bradick of Factor warned: "Token consumption is going to be a huge problem if you're not pricing for that."

The large research and drafting platforms do not publish list prices. Third-party trackers place Harvey's enterprise contracts at roughly $50,000 to $200,000+ a year. Thomson Reuters sells CoCounsel on custom or bundled quotes, and no published base tier has been confirmed. In practice, then, many departments face opaque, negotiated pricing. Some of it includes usage components and some does not. Clean per-token bills are not the norm.

That matters for planning. Consumption risk can sit inside the contract (overage clauses, credit pools, tier ceilings), in the vendor's own cost base and so in the next renewal price, or in direct API use by internal teams. Each one needs a different control.

Why AI costs are hard to forecast

Analyst commentary consistently says that AI cost estimates miss, sometimes by a wide margin.

Gartner VP analyst Mary Mesaglio warned at the Gartner IT Symposium on Australia's Gold Coast that AI cost miscalculations of 500 to 1,000 percent are possible, comparable to early cloud overruns, as heise reported.
In a 2026 report on optimising generative and agentic AI costs, Gartner predicted that through 2028 at least 50% of generative AI projects will overrun their budgeted costs, according to Campus Technology. Gartner expects inference to make up at least 70% of a model's lifetime costs. It cautioned that a production system "can be orders of magnitude more expensive" than a pilot. It also noted that agents, which make several model calls and run multi-step workflows, can push costs up significantly.

These are cross-industry figures, not legal-specific ones. Still, the mechanism carries over directly to legal work. High-volume, repetitive tasks multiply inference calls. Examples include first-pass document review, bulk contract extraction, always-on monitoring and multi-step agentic workflows. A pilot on a few hundred documents says little about the cost of a 200,000-document matter.

The budget side: AI work often arrives unfunded

The best recent legal-specific data concerns where the money comes from, not what the tools cost.

According to Law360 Pulse, Axiom released a survey in September 2026 of about 500 in-house leaders across six countries, conducted by InsightDynamo. It found the following:

92% of the time, requests to implement AI came without a dedicated budget and were paid for by reallocating other funds.
Only 8% said AI initiatives came with a dedicated budget, compared with 18% reporting separate AI budgets in 2025.
59% saw outside firms raise fees while promoting AI as adding value, up from 30% the previous year.
Only 4% saw fewer billed hours that they linked to AI.

Other published data points in the same direction:

CLOC's 2026 State of the Industry coverage reported that 85% of departments have a dedicated resource or committee to manage AI use. It also found that 47% expect internal budget increases, down from 65% the year before.
The ACC/Everlaw 2025 survey of in-house counsel found 59% had seen no generative-AI-driven savings from outside counsel.
Gartner predicted in November 2024 that legal, risk and compliance leaders would double their technology spending by 2027.

Taken together, departments are being asked to absorb AI costs out of existing budgets while the expected savings, at least from law firms, have not yet appeared. That is the real budget pressure, whatever pricing model a given vendor uses.

A cost-management discipline is emerging, mostly outside legal

Legal ops does not need to invent token cost governance from scratch. The FinOps Foundation runs a working group on "Tokenomics: Managing AI Value in SaaS Model Token Costs." It calls managing token spend from model providers the top challenge FinOps practitioners now face, and states: "The token is the billing unit, not the value unit." Its framework covers:

Visibility: API key governance and proxy layers.
Unit metrics: cost per query, per user and per successful outcome.
Budgets and anomaly alerts.
Optimisation levers: model right-sizing, which it presents as the highest-impact lever, plus caching, batching and context management.
Maturity path: a crawl-walk-run sequence that ends in governance and chargeback.

Its list of reasons token spend is hard to manage also fits legal departments. Billing lacks native attribution, users buy access outside procurement, usage spikes unpredictably, and output tokens cost more than input tokens.

On the value side, Thomson Reuters' Future of Professionals 2025 report, based on 2,275 professionals, found that participating professionals expect AI users to save five hours a week within a year. It also found that organisations with a visible AI strategy were twice as likely to report AI-driven revenue growth, and only 22% had one. That is survey-reported expectation, not measured savings. It is still the right kind of figure to set against spend.

Analysis: what "managing it well" plausibly looks like

The following is The Legal Stack's analysis, not survey findings.

Know which pricing exposure you actually have. List every AI contract by pricing basis: seat, document volume, credit pool, usage tier, or direct API. Usage risk only applies where the contract or the architecture passes it through.
Negotiate ceilings before you need them. Where usage components exist, ask for hard caps, alerts at set thresholds, rollover of unused credits, and price protection at renewal. Bloomberg Law's reporting suggests pricing varies between peers, so terms are negotiable.
Model the big matters, not the pilot. Following Gartner's pilot-to-production warning, price a realistic high-volume scenario before rollout, such as a large review set or a full contract repository. Do this especially for agentic workflows.
Attribute cost to matters or business units. FinOps practice treats allocation and chargeback as the mature end state. For legal, matter-level attribution also makes it possible to compare AI cost against the outside-counsel spend it is meant to replace.
Treat AI as a line item. The Axiom finding that most AI work is funded by reallocation points to an obvious fix: give AI its own budget line with an owner.
Measure value in outcomes. Track cost per completed task and hours redeployed, not tokens consumed.

We do not have evidence that centralised governance "outperforms" other models by any particular margin, and we make no such claim. Centralised visibility is a precondition for most of the controls above, though, and that is a reasonable basis for preferring it.

Sources

Bloomberg Law, "AI Legal Software Flails on Pricing Models, Frustrating Buyers" (15 Jun 2026): https://news.bloombergtax.com/legal-ops-and-tech/ai-legal-software-flails-on-pricing-models-frustrating-buyers
Campus Technology, "Gartner: Half of Gen AI Projects Could Exceed Budget by 2028" (22 Jun 2026): https://campustechnology.com/articles/2026/06/22/gartner-half-of-gen-ai-projects-could-exceed-budget-by-2028.aspx
heise online, "Up to 1000 percent more: Gartner consultants warn of expensive AI bills" (2024): https://heise.de/-9864445
Law360 Pulse, "AI Costs Outpace Budget Gains For In-House Legal Teams" (8 Sep 2026): https://www.law360.com/pulse/articles/2522500
CLOC, "CLOC's 2026 State of the Industry Report: Benchmarking Data Is the Compass for Legal Operations To Navigate Change" (2 Mar 2026): https://cloc.org/?p=16187
FinOps Foundation, "Tokenomics: Managing AI Value in SaaS Model Token Costs" (accessed Oct 2026): https://www.finops.org/wg/token-economics-saas/
Thomson Reuters, "The AI adoption reality check: Firms with AI strategies are twice as likely to see AI-driven revenue growth" (26 Jun 2025): https://www.thomsonreuters.com/en/press-releases/2025/june/the-ai-adoption-reality-check-firms-with-ai-strategies-are-twice-as-likely-to-see-ai-driven-revenue-growth-those-without-risk-falling-behind
CostBench, "Harvey AI Pricing 2026: Enterprise Legal AI ($50K-$200K/Year)" (verified 1 Jul 2026): https://costbench.com/software/ai-legal-tools/harvey-ai/
CostBench, "CoCounsel pricing" tracker (accessed Oct 2026): https://costbench.com/software/ai-legal-tools/cocounsel
Everlaw, "64% of In-House Counsel Expect GenAI to Reduce Reliance on Outside Counsel, 50% Expect Lower Outside Counsel Costs, ACC/Everlaw Survey Finds" (14 Oct 2025): https://www.everlaw.com/press/release/acc-report-2025/
Gartner, "Gartner Predicts Legal, Risk and Compliance Functions to Double Technology Spend by 2027" (13 Nov 2024): https://www.gartner.com/en/newsroom/press-releases/2024-11-13-gartner-predicts-legal-risk-and-compliance-functions-to-double-technology-spend-by-2027