Light from the Dark Side

Management summary

• Kirkland & Ellis, the largest law firm in the world, has put $500 million into a system built with Palantir. It will be reported as an AI story. It is really a data story.

  
  

• The clever part is not a better chatbot. It is turning decades of messy fund paperwork into clean, structured data the firm can rely on without rechecking. Tools like Harvey and Legora, for all their strengths, cannot do this dependably, because they work by prediction and will sometimes get it wrong.

  
  

• The software behind this will eventually be available to buy. The clean data will not. That is the real barrier, and most firms cannot keep even a simple client database in order.

  
  

• So the firms that pull ahead will not be the ones with the deepest pockets. They will be the ones whose data is in order, and on that score almost everyone is starting behind.

  
  

• Preparing means two things: take data seriously enough to hire a senior person to own it, and accept that the firm will change shape, with data and AI specialists working beside lawyers in client-facing teams, some at partner level and paid accordingly.

---

I will admit something. Until a few weeks ago, whenever I heard the name Palantir, a particular picture formed: government contracts, intelligence agencies, surveillance programmes, software for tracking enemies abroad and, now and then, citizens at home. I never examined the impression. If I am honest, I never really looked into the company at all. Palantir simply sat, in my mind, somewhere on the Dark Side. It turns out I was wrong.

What changed my mind was a deal. This month Kirkland & Ellis, the largest law firm in the world by revenue, announced a multiyear partnership with Palantir, backed by a $500 million commitment to its own AI.

Most of the profession read the headline, filed it under “another big firm does AI,” and moved on. That filing is the mistake.

The genius of what Kirkland has done has almost nothing to do with drafting and almost everything to do with data. Once I understood why, I stopped thinking about the Dark Side and started thinking about how badly the rest of the profession has misread what this signals.

Why the Kirkland and Palantir deal is really about data.

The hard part of private equity fund formation was never writing the documents. It was keeping track of what they promise. A single fund carries hundreds of side letters, each granting one investor something specific: a fee discount, an excuse right, a transfer restriction, a most-favoured-nation clause that silently entitles them to any better terms a later investor receives. Across a dozen funds and several hundred investors, those promises form a web of live obligations that interact with one another, with every new fund term, and with every transaction the manager wants to make, for years. Checking a continuation vehicle against that web is not a reading task. It is a tracking task at a scale no human team can hold in its head.

It pays to be precise about what tools like Harvey and Legora actually do, because they do it well. Both are built on large language models. Ask Harvey whether a side letter’s MFN clause matches the template, and it will tell you. Feed Legora a thousand contracts and it will pull the key terms into a grid, with citations back to the source. For drafting, summarising and review, these are a genuine advance, and any firm not using something like them is already behind.

But they share three limits, and none is a bug the next release will fix. First, they guess. A language model predicts probable text; it estimates rather than knows. On Harvey’s own benchmark, its best model still invents roughly one claim in 500, and independent testing of legal research tools has found higher rates. The emerging consensus is that hallucination is inherent to these models, not a defect to be engineered away. One error in 500 is excellent for a drafting assistant. For a covenant check across a $50 billion raise, where a single missed clause is a multimillion-dollar breach, almost always right is the wrong standard.

Second, they work one question at a time, against the documents. Nothing durable sits underneath. Ask again next month, after three new side letters land, and the tool rereads the pile from scratch. There is no single source of truth holding, at all times, the current state of every obligation. Third, they struggle with identity at scale. “Texas Teachers,” “TRS” and “Teacher Retirement System of Texas” are one investor; a model that infers this across decades of inconsistent drafting accumulates quiet errors.

What Palantir adds is the thing the announcements bury under the word “ontology,” and it is simpler than it sounds. Palantir does not point a model at the documents and ask it to be careful. It uses the model once, as a labourer, to lift each obligation out of the prose and turn it into a structured fact: an MFN clause, owned by this investor, in this fund, triggered by these conditions, expiring on this date. That fact then lives inside a governed model of the whole business, where every fund, investor, clause and transaction is an object linked to every other. And here is what matters. Once the obligation is structured data, the compliance check no longer runs on the language model at all. It runs on fixed logic, hard rules applied to hard data. The probabilistic engine reads. The deterministic engine judges. That division of labour is the whole game.

So the genius is not that Kirkland bought cleverer AI. It is that Kirkland refused to let the answer to “does this breach a covenant” come from a model’s best guess, and built the layer where it does not have to. Harvey and Legora answer questions about documents. Kirkland is building a system that answers questions from a model of its business. One is a brilliant reader. The other is a system of record that uses AI to fill and query itself.

The advantage is real, and it is worth being exact about its source. It is not the software, which others will eventually license. It is three things a rival cannot simply buy: the capital; the proprietary corpus, the nearly $500 billion of fundraising Kirkland touched in a single year, which is the reference data the system learns from; and ownership of the application layer, where the firm’s own judgement is encoded. This is a moat, not a monopoly. Every fund still has investors who need their own counsel, and there will always be a firm across the table. But that firm is now doing a fundamentally different job from one still billing hours to reread the side letters.

The gap that is actually opening

Almost no firm outside the largest American practices will ever have either ingredient. They will not write a nine-figure cheque, and they do not sit on a proprietary record of half a trillion dollars in annual deal flow. Realistically, every firm outside the US is a national champion: large at home, small against Kirkland. The temptation is to call this someone else’s problem and wait. That is the wrong call, because the capability will arrive off the shelf. Every serious enterprise software vendor now sells a version of the same idea, and legal-specific versions will follow. Within a few years a national champion will license an obligation-tracking engine without rebuilding Palantir from nothing. The platform commoditises.

Which sounds like good news, until you remember the platform was never the hard part. The data is. An engine like this is only as good as what you feed it, and what you feed it is your own data, cleaned, reconciled and trustworthy. This is exactly where firms are weakest. The honest test is simple: most firms cannot keep a CRM clean. They cannot say with confidence who knows whom, which partner owns which relationship, or what the firm has done for a client over ten years, because that information is scattered across inboxes, spreadsheets, document systems and partners’ memories, and no one owns it. A firm that cannot maintain a contact database has no hope of maintaining a live model of every obligation across every fund. Feed a clean engine dirty data and it will industrialise the mess.

So the real divide is not between firms that can afford the software and firms that cannot. It is between firms whose data is in order and firms whose data is chaos. Almost everyone starts behind, which is the one genuinely good piece of news for the smaller firm: data discipline is not bought with scale. A determined firm that gets serious now can close the gap faster than its size would suggest.

This is also where the argument about outside capital stops being a matter of principle. Building real data capability costs money a partnership funds from its own distributions, which means partners feel it in their own pay this year for a benefit that lands three years out. Partnerships are structurally poor at exactly that kind of investment. Private equity money buys the runway to get the data house in order before the capability becomes the price of entry rather than an edge. The firms that took the investment will turn out to be the ones that could afford to act in time.

What a sane firm does now

Two things follow, one obvious and one structural. The obvious move is to treat data as a strategic asset rather than an IT cost. In practice that means giving data an owner with real standing: not a junior analyst parked in operations, but a senior data specialist with the authority to fix the plumbing and to change how lawyers record what they do, reporting high enough that partners have to listen. Most firms do not have this person. They have a help desk.

The structural move is the one that will be resisted, because it changes the shape of the firm. As AI absorbs the production work that fills the base of the pyramid, the base narrows. Fewer juniors are needed, and those who remain must be more capable from their first day, working on judgement and client contact rather than document volume. The pyramid becomes a diamond, and the swollen middle of that diamond has to become a destination in its own right, not a waiting room on the way to an equity most will never reach.

The diamond is not only a story about lawyers, and this is the part the profession is least ready for. It brings into the firm a class of senior professionals who are not lawyers at all: data scientists, legal engineers, pricing specialists, AI leads. In the old firm they sat in the back office and were paid as such. In the firm that is coming, they sit on the client-facing team. A data scientist who builds the model that shows a client which of its obligations are exposed, and presents it in the room, is not support staff. The work is economically indistinguishable from a partner’s.

Which forces the awkward question I devote a chapter to in my latest book, Law Firm Partner Compensation: how do you pay such a person? In most jurisdictions a non-lawyer cannot hold equity, so formal partnership is closed to them. Pay them like a senior associate, though, and you will not keep them, because the pool of people who can do this work is small and the competition fierce. The honest answer is to build a structure that behaves like partnership even where the title is forbidden: shadow equity that tracks the firm’s performance and falls when profits fall, bonuses tied to outcomes you can actually measure, and a real voice on the matters they own. Recruit people for partner-level work and pay them like assistants, and they walk. That is not a moral observation. It is arithmetic.

I began by confessing that I had filed Palantir under the Dark Side and left it there, unexamined. The irony is not lost on me that the company I associated with watching people from the shadows is the one that has shown my own profession something clarifying about its future. The light it casts is not flattering. It falls on an industry that has spent two years arguing about chatbots while the real contest moved to the data underneath, where almost no one was looking and almost no one is ready. Kirkland’s half a billion is the part no one else can copy. Taking data seriously, and letting non-lawyers into the room where the value is made, costs almost nothing. Which is why so few will do it. There is, it turns out, light from the Dark Side. The only question is who is willing to look at what it shows.

This article is part of a weekly series drawing on the themes of Law Firm Partner Compensation by Jaap Bosman and Jaime Fernández Madero. If you would like to know more about this topic, read the book.

Go to Law Firm Partner Compensation

Our book Law Firm Partner Compensation is available worldwide on Amazon, national online book sellers, and can be ordered at your favorite at your favorite bookstore