This is where technical discipline earns its keep. Not every legal query needs the most powerful model on the market. In fact, most do not.
Concerns over the cost of AI have become a recurring theme in LegalTech and beyond. Conversations between vendors and buyers now centre on concrete use cases and measurable payoff, rather than potential and speculation.
Buyers have stopped taking AI’s value on faith, and that shift, while it might sound uncomfortable for vendors, could well be an extremely beneficial shift for the legaltech market at large.
Buyers are getting smarter about ROI
A slick demo used to be enough to win a deal in legal AI’s early days. That’s no longer true. Prospective clients now show up to sales conversations with their own files and their own test questions, ready to see whether a tool actually holds up under real conditions. I welcome that shift. Being pushed to prove ROI forces us to demonstrate genuine value. A client asking hard questions before signing is a very good sign: as they are far more likely to see the deployment succeed.
Part of what makes this possible is that AI tools are finally giving legal teams the visibility they never had. Many in-house departments have never systematically tracked factors like turnaround times, or how many matters they’re handling, but that is changing. LegalTech platforms can now surface exactly how time is spent and how quickly work gets resolved, which means results are no longer a matter of opinion.
That visibility is proving even more powerful outside the legal department itself. Other enterprise departments — whether this is compliance, operations, HR, marketing, sales — are increasingly using legal AI tools. The numbers speak for themselves. To use one of our clients as a real-life example: we’ve seen an insurance claims team go from manually processing about 150 claims per-person-per-month, to roughly 700 in the same period once agentic AI came into the workflow.
It’s time to be smart: Tokenmaxxing and reckless spending got us here
This new rigor didn’t come out of nowhere. Instead, it is a direct response to a period of genuinely reckless spending. Uber reportedly burned through its full-year AI budget by April and had to cap employee spending on tools like Claude Code and Cursor.
Elsewhere, the rise of “tokenmaxxing”, treating heavy token consumption as a proxy for skill or productivity, pushed costs even higher. Amazon is understood to have quietly retired an internal leaderboard that ranked token usage. By early August, Microsoft was reportedly telling staff internally that the goal should be results, not token volume, and it too introduced spending caps.
Seen in that light, the current focus on cost and return isn’t a sign that enthusiasm for AI is fading. It’s simply what happens once a technology matures past its hype phase. As models get more capable, and ROI is genuine, the bills get bigger. Justifiably, the conversation turns to whether the spend is justified.
The AI Cost pressure runs both ways
It would be a mistake to think only buyers feel this squeeze. Vendors are under just as much strain, since their margins depend heavily on the processing costs of the large language models underneath their products. For LegalTech startups, getting this balance right is central to whether the business survives at scale.
Most vertical AI companies still charge per seat, mainly because procurement teams want predictable costs they can plan around. That model, however, sits awkwardly alongside ever-rising token consumption, squeezing vendor margins as usage climbs. The alternative is usage-based pricing, which solves the margin problem. However, the downside risk is that it reignites that exact fear that buyers want to avoid: those of unpredictable, spiraling costs that got us here in the first place.
Smarter engineering will win out
This is where technical discipline earns its keep. Not every legal query needs the most powerful model on the market. In fact, most do not.
Vendors who succeed will be those who match model size to task complexity, instead of simply defaulting to the priciest option as default. Additionally, winners will build multiple agents across a workflow, with each agent running on the best suited model (again, rather than the priciest or most recent). There’s no need to reach for a bazooka when a penknife will do the job just as well!
None of this cost scrutiny is bad news. If anything, it’s forcing LegalTech companies to build on sturdier foundations, and design business models that can genuinely deliver returns. This has to be good news: for themselves, for their investors, and ultimately for the clients who depend on the technology working as promised.
Join the conversation!