In April 2026, only four months into the year, Uber's CTO announced the company had burned through its entire AI budget for 2026.
The COO went further, conceding the productivity case still hadn’t been made and that nobody could draw a clean line between what Uber was spending on AI and what it was getting back.
That story made headlines in tech circles but it deserves far more attention in legal, because the thing that broke Uber's budget, runaway token consumption with no clear link to output, is a problem legal is unusually exposed to and this is not just having an impact on law firms, but also in-house legal teams chasing efficiency gains.
Legal is one of the most token-intensive functions on earth
Tokens are the currency of every AI interaction, and legal spends more of them than almost any other function. Every clause reviewed, every contract risk-flagged, every precedent searched draws on the same meter.
Legal runs on documents more than almost any other function. Long contracts, multi-party correspondence and due diligence packs that run to thousands of pages all add up to a volume of text that would overwhelm most other parts of a business.
Agentic AI widens the gap and Legora just announced in June 2026 that it is moving to consumption-based billing for its new Agent Pro product. These tools don't draft a clause and stop, they run a whole review workflow on their own, reading and re-reading documents as they go, and the token count climbs fast. That is landing on an industry already spending at record levels; law firm technology spending rose 9.7% in 2025, the fastest real growth the sector has likely ever seen, driven by the race to deploy generative AI, according to Thomson Reuters and Georgetown's 2026 State of the US Legal Market report.
The subsidy era is ending
Most legal teams that signed enterprise AI contracts in 2024 or early 2025 are sitting on deals that won't survive renewal. Providers set prices before they fully understood how lawyers actually use these tools. The early pricing was effectively a subsidy to win the market, and it was never going to last.
Per-seat pricing is already disappearing, as Shawn Curran, CEO of Jylo, puts it: "Tokenmaxxing is a thing, per seat pricing is gone, if Anthropic, Microsoft and OpenAI have moved away from it, no-one is going to subsidise legal tech vendors on all you can eat. Therefore firms are going to have to spend more."
Firms that signed aspirational AI agreements are now confronting actual usage data and the numbers are rarely what the initial business case assumed.
BigLaw's answer: own the infrastructure
A handful of firms big enough to absorb the cost are taking a different route and building their own. Kirkland & Ellis is spending $500m over the next couple of years on its own AI platform, to own the technology rather than rent it from providers whose prices keep climbing. Simmons & Simmons built Percy, which reached 87% adoption among fee earners in a year, and Thomson Reuters is moving to open-source models, partly for better answers and partly for better margins.
But a $500m, multi-year build only works if you are a firm the size of Kirkland, able to fund it from your own revenue, hire your own engineers and spread the cost across an enormous fee base. It isn't an industry fix, and it isn't something most firms or in-house teams can realistically copy. For everyone else, the problem looks rather different.
In-house teams are caught in the squeeze
GCs are navigating pressure from both sides at once.
On one hand, the pressure to adopt: a firm without a credible answer to "how are you using AI?" is out of the running on competitive pitches, so the money goes out the door to test and roll out new tools. On the other side, the awkward fact that none of this has been evidenced as lower bills. .
If anything, the bills are going the other way. Industry-wide law firm revenue increased 12.6% in 2025, matching the pace set in 2024, according to Wells Fargo's Legal Specialty Group year-end survey. Average standard billing rates increased 9.6% year over year, up from 9.1% in 2024, and were the primary driver of revenue growth.
Increasingly, clients tell us that some of the cost of AI is beginning to make its way into what their firms charge, a cost those same teams are, in many cases, already carrying on their own deployments.
So the spend can end up landing on both sides of the relationship: once on the tools in-house teams adopt themselves, often under pressure from their own organisations to show AI savings, and again as firms start to reflect their AI costs in fees that were already on an upward path.
AI was meant to ease this. The promise was quicker turnarounds, greater efficiency, and more room for lawyers to focus on the genuinely complex work, and in some areas it may be delivering that. For some in-house teams, though, the bills haven't yet caught up with the promise.
Additionally, proving what that AI spend returns is still hard. Only 36% of CFOs say they are confident their organisation can achieve meaningful AI outcomes, which is part of why 77% of GCs report some form of tension with their CFO over spend. (According to the 2025 In-House Legal Budgeting Report from Wakefield Research)
All of which leaves an obvious question hanging: : if AI is delivering all these efficiencies, where are the savings going?
The big question for most in-house legal teams
Most in-house teams aren't going to build their own platform like Kirkland. What they need is a delivery model that captures AI's efficiency without leaving them exposed to uncapped, unpredictable token spend, or to a supplier whose incentives run in the opposite direction to their own.
That incentive problem is the part that usually goes unexamined.
Under an hourly or day-rate model, efficiency impacts profitability. Every hour AI saves is an hour the provider can no longer bill. Every rising token cost is something to mark up and pass on. The structure keeps the meter running.
Fixed pricing inverts that.
Radiant has charged on a fixed-fee basis since 2011, a flat monthly retainer or a price per contract, never hourly or day rates. That isn't a billing preference. It's the thing that aligns our interests with our clients'. When the price is agreed up front, the only way we improve our margin is by getting more efficient at the work itself..
So when AI can do something faster or more cheaply, we are motivated to find that efficiency and bank it: routing tasks to the right model rather than the most expensive one by default, cutting waste, and improving the process every time we run it. The client gets a predictable cost and the benefit of the efficiency. We carry the complexity, and the risk, of the underlying AI spend.
That's the structural difference. AI under hourly billing is a cost to be passed on. AI under fixed pricing is a reason to get genuinely better.
The questions worth asking before your next renewal
Before the next AI contract cycle comes round, these are the questions pressing:
Legal teams asking these questions now are the ones that will be well-positioned when the next round of renewals arrives.
The ones that aren’t will be the ones explaining the numbers to their CFO instead.
You might be wondering what your next steps should be. Let us guide you with three easy options:
We use cookies to monitor the performance of our website and to improve user experience. By continuing to browse our site, you agree to our use of cookies.
