The improvements in productivity with legal AI so far are disappointing. The models are more than good enough. Like, seriously... Have you ever tried to build a video game using Fable?
Success with AI today has very little to do with buying the right tool or choosing the right model. In our engineering team we use both Codex and Claude Code. Some folk prefer Codex. Others Claude. Most switch depending on token limits. Models are becoming a commodity because we don't need to pay for the frontier to get work done anymore, and tools are becoming commodity because anyone can build them on a rainy weekend.
And yet, law firms and legal departments spend huge amounts of time and money on evaluating legal AI tools, then, oddly, celebrate wildly once they have made their final decision. What exactly is being celebrated here? What has been achieved in merely selecting the tool?
Recently, I searched LinkedIn for law firm posts announcing the AI platform they had decided to roll out “firm wide”. I looked for posts that were at least 12 months old, and then I combed through the firm’s LinkedIn feeds for any sign of a follow-up post, anything that amounted to “we have saved our clients $12M this quarter” or “our firm’s revenue has meaningfully increased without a corresponding increase in headcount”.
But there was nothing.
Instead, confusingly, I found contradictory announcements of an entirely different vendor being selected. “We are proud to announce our decision to partner with Harvey on a firm-wide roll out,” began one post from a law firm who 12 months earlier had posted an almost identical announcement about their firm-wide roll out of Legora.
Something isn’t working
Clearly, something isn’t working. AI platforms are being bought and vendors are being funded at ridiculous scale, yet the corresponding impact on productivity isn’t there. In the Thomson Reuters Institute’s 2026 AI in Professional Services Report, 85% of law firms and 75% of corporate legal departments said they either do not collect ROI data on their AI use or are not sure whether they do. Most of the market cannot tell you if any of it is working.
This is worrying on a number of levels. The entire AI supply chain is dependent upon sufficient demand for compute, and a great deal of this demand comes from VC-backed startups and scaleups. Demand requires ROI, eventually, or that demand collapses.
Nobody in that chain is short of capability. The models are good enough. But in the majority of cases the approach to improving productivity with AI is entirely wrong.
What is missing has a name: diffusion. It is everything that has to happen between signing a licence and seeing a return.
Shyam Sankar, Palantir’s CTO, put it at the centre of their third-quarter earnings call last week. Today’s models, he argued, are already capable of solving a huge proportion of economically valuable problems. The challenge is no longer making them more intelligent. It is the hard work of diffusion, “which sounds passive, but is anything but”.
The wrong unit
The fundamental unit of the enterprise is the workflow, not the worker.
That is Sankar’s explanation for why so little AI diffuses. His charge against the AI labs is that they treat the individual knowledge worker as the atomic unit of the enterprise. They are wrong. “You can make any set of workers more productive and generate no actual ROI,” he said. At Flank, we agree.
So much of the legal AI market is obsessed with the idea of the individual lawyer as the unit of productivity. They work on the assumption that getting this individual lawyer to increase their throughput will unlock some new paradigm of productivity. Evidence for this is everywhere, from per-seat licences to MS Word add-ins to ROI metrics that include adoption rates, and beyond. All of it is centred around optimising for the wrong thing, whether one person got through their own tasks a little faster.
The result is that the individual lawyer feels more productive, and the business sees no difference.
Lili, our CEO, made the same point recently with an ordinary commercial contract. The work starts with someone in the business. It involves gathering information, understanding company policies, drafting or reviewing documents, communicating with counterparties, obtaining approvals, and eventually getting something signed and stored in the right place.
The lawyer is involved in that process, but the lawyer is not the process. You could make the lawyer twice as productive and still have a workflow that is slow, fragmented, and expensive to operate.
A licence reaches the lawyer. Diffusion has to reach the workflow.
The diffusion layer
Palantir’s revenue grew 96% year on year last quarter, the fastest it has grown as a public company. That is insane growth for a business that made $1.1B in net income in the same three months. At that size, you would think they would start to slow down. And yet, over the last twelve months, Palantir’s largest twenty customers by revenue paid an average of $146M each to Palantir, an increase of 76%.
And I’m throwing these numbers at you to make a single point: Palantir is not just growing, they are providing such a significant return on investment to their customers that their customers are throwing hundreds of millions of dollars at them for more. Net dollar retention is 165%.
So what exactly is Palantir doing differently from the endless stream of AI application layer businesses who fail to generate meaningful ROI for their customers?
They do the diffusion themselves. Sankar calls Palantir “the diffusion layer”. They do not sell a tool to a worker and hope. They bring a platform configurable enough to hold the customer’s real workflow, and people who go in and build that workflow alongside the customer, with AI, software and humans put together inside it.
His example was complex commercial underwriting at AIG, where the whole process runs on Palantir’s platform and was built by the two companies together. Alex Karp, Palantir’s CEO, calls the combination “product and artistry”.
Diffusion means getting intelligence into the messy reality of enterprise processes, systems, rules and human oversight. This is difficult work. It is also where most of the value will be created.
Palantir has also noticed who wins. The customers getting the most from AI are not trying to trim their operating costs; Sankar called that goal lame. The ones who struggle are narrowly focused on reducing cost. The winners go after the work the business always knew it should be doing and never had the capacity to take on.
Between licence and ROI
The real opportunity in legal AI is for a system that bridges the gap between licence and ROI. It combines a highly configurable platform with human experts who are fluent in both the technology and the workflows.
Neither works alone. A platform without the experts is a licence, and we have seen where licences lead. Experts without the platform are a consultancy, and their work does not compound.
That system is what we built Flank to be. Flank deploys agents that do legal work.
The platform is configured to each client’s own systems, documents and rules. A forward deployed team, fluent in both the technology and the legal work, builds each agent from an Agent Design Document, and the client watches a prototype do their work on their own paper before they commit. The agent then does the work end to end. It raises the few points it is unsure of or that carry risk, and the client’s own lawyers sign those off.
For multiple global hospitality groups, Flank agents draft entire franchise agreement packs from a heads of terms, to the group’s own playbook, supervised by the group’s own lawyers before release. Our forward deployed team, when reviewing success of these accounts, only gives a shit about the work that's actually getting done. There are no prizes for the number of users.
Like Palantir’s best customers, ours are not the ones chasing a cost line. They ask about the work that is not getting done: supplier paper signed without review, renewals that lapse, requests that never reach legal. An agent makes that work viable for the first time.
Of course there will always be legal work that benefits enormously from a skilled lawyer working with an intelligent assistant. But for the huge volume of recurring work passing through enterprise legal teams, we will look back and wonder why we spent so long trying to make the person processing the work faster.
We have been building Flank around this conviction for a long time, and every year it feels a little less contrarian. Capability is commodity; diffusion is rare.
How AI was used: Claude helped me order the piece, checked the Palantir figures against the earnings call, found the Thomson Reuters statistic, and drafted four linking sentences from my notes and our own existing material, which I edited.
