How will AI change the job market? I won't go into that territory. If you want to sound smart while knowing nothing at all, you go with one of the common phrases: "AI won't replace jobs, but people using AI will replace people who don't." Or: "AI won't replace work, but it will change the way work is done." Both are crafted so that nobody can disagree, which is exactly why they tell you nothing.
What I will go into is a pattern that has emerged over the past few years of actual AI work, one I have a personal stake in. Back when I practiced law, I was the go-to guy for everything: litigation, corporate and M&A, insolvency, real estate, taxes. I've done it all. When I got fed up with it, I taught myself Python, badly at first. Now I spend my days implementing AI in legal work, which means writing code, arguing about data models, and explaining to lawyers why the thing they saw in a demo won't survive contact with a real matter.
From inside a law firm, that trajectory reads as a failure to commit. Legal careers reward depth in one lane: you become the person for German fund structures, or for cross-border insolvency, or for merger control, and the market pays you for the narrowness. Breadth reads as dilettantism. Every time I picked up something outside the lane, the honest internal question was whether I was building a career or just avoiding one.
But the calculus has changed. Not in the abstract, not as a prediction: generalist skills are suddenly in high enough demand that the market has created job titles for them. AI Engineer and Forward Deployed Engineer are real titles with real salaries at Anthropic, OpenAI, Google, or Cursor. Both reward exactly the instinct I spent fifteen years treating as a liability — being competent across domains rather than excellent in one. The tide is turning, and the reason is economic, not sentimental.
Integration Is the Moat
Here's the economics. Every enterprise can now buy the same frontier models. Same API, same weights, same benchmark scores. I've argued before that the model is the least durable layer; what matters here is what follows from that. When the model is a commodity, it stops being the moat. The moat is integrating it into a business that actually exists.
Integration is where the value moved. It means deciding which workflows are worth touching at all, where an LLM belongs versus deterministic code versus a human in the loop, how it plugs into existing systems, and what happens when it fails. That work sits on the seam between business mess and technical systems. It is generalist territory by construction.
The two new roles are that territory, formalized. The AI Engineer, a term coined by swyx in his 2023 essay The Rise of the AI Engineer, builds with foundation models rather than doing deep ML research. Look at what the job actually spans: system design, full-stack coding, prototyping, evaluation and observability, databases, information retrieval, OCR and document pipelines, and model selection with cost and latency tradeoffs. Half of them are the data-quality work I keep arguing is the real bottleneck in legal AI; several weren't named disciplines three years ago. The list's length is the point: it's why the role is scarce and why it's paid like it is.
The Forward Deployed Engineer originated at Palantir and has since spread across the frontier labs. An FDE embeds with a strategic customer and ships production systems inside that customer's environment. The job description is consulting plus engineering. Done well, the job is the best of both rather than the average of both. That distinction matters. An average of consulting and engineering is a slide deck with a prototype attached. The best of both is a tech person who can also sit in a room with a partner, understand what they actually do all day, and build for that.
What consulting decks never capture: the real process usually lives in one person's head. The documented workflow says a deal-closing checklist is maintained in the DMS. What actually happens is that one senior associate keeps the live version, three partners email her fragments, and the "system of record" is reconciled on Thursday afternoons. Automate the documented version and you've built something nobody uses. The generalist's advantage is seeing the gap and fitting the system to where the work really happens.
The second thing decks miss: there's one way for a workflow to go right and roughly a thousand ways for it to go wrong. Building for the happy path is what a demo is. Engineering the unhappy paths (the counterparty who sends a photo of a signature page, the clause the model confidently misreads) is what production is. I've written about that gap before, and it's the clearest line between someone who can demo AI and someone who can deploy it.
A third thing the decks miss — and this one lowers the bar instead of raising it: you have to speak the domain's language, not be an expert in it. Build for a bank without being able to talk about payment rails and risk, or for a law firm without speaking the language of the practice sitting across the table (the fund lawyers will expect you to know what an LP side letter is), and you've lost the room in the first ten minutes. But "gets the gist and can hold a conversation" is a much lower threshold than "qualified practitioner." That's precisely what makes the generalist viable.
And there is no canonical entry point into this work. Product managers are vibe-coding prototypes. Lawyers are, too. Backend developers are building frontends, and coders are taking on product work, because AI and AI-native project management tools let them. When all of that sits in one head, or in a small team, projects stop waiting on handoffs: AI has made code execution fast, so the goal now is to streamline everything around it. My own route in ran through Python, self-taught, badly at first. Some days it still is: I'm by no means an expert in all the fields this role spans. I'm good enough and eager enough to dive into each of them, spot where I fall short, and find ways to fill the gaps, sometimes with outside help, often with AI itself. What matters isn't where you start but the willingness to expand from there: the time, the grit, and the curiosity to take on unfamiliar technical material, plus hands-on fluency with the frontier AI coding tools. The door is open from any direction. Walking through it is still work.
The New Roles in the Wild
The titles have crossed into legal, too. As of this writing, Eudia, an AI-native law firm, is hiring AI Engineers, Forward Deployed Engineers, and, my favorite detail, Forward Deployed Lawyers. Crosby, another AI-native firm, has an open req for "Member of Technical Staff, Generalist": a law firm hiring a generalist by title.
New roles, new org structures, new procurement strategies: all of it is being invented in public, because building AI into your own systems demands a skill set firms didn't have two years ago. The most useful thing to look at isn't the winner but the spread.
Kirkland & Ellis committed $500 million over three to four years, starting with a Palantir-built "Fund Formation Engine" for private equity fundraising. But the story is the staffing: roughly 180 in-house AI engineers and data scientists, plus 250-plus attorneys (about a hundred of them equity partners) organized into "AI Pods" that work alongside the Palantir engagement. The co-build works because Kirkland is simultaneously building the capability to own the result.
Clifford Chance placed a different bet: an AI knowledge platform built with Microsoft and Epiq, 400,000-plus reclassified, permissioned documents living inside Microsoft 365 and Azure, launched in July 2026, plus a wexler.ai partnership on the dispute resolution side. Vendor partnerships layered onto infrastructure they already run, not a co-build.
White & Case went buy-and-build: a proprietary in-house platform called Atlas on Azure AI Foundry, alongside Legora rollouts. Their own framing: adopt the best available tools while developing their own capabilities.
Three firms at the same tier, three genuinely different shapes of the same goal: AI capability built into the firm. Every one of those shapes runs on the cross-domain people described above: engineers who can talk to lawyers, lawyers who can work alongside engineers. Notice what none of these bets is: just a Harvey or Legora seat rollout, which plenty of firms have done. SaaS adoption doesn't create these roles. Building does.
What Happens When They Leave
One risk never makes it into the proposal. Companies bring in outside AI experts because of a talent shortage or an internal political blockage. Then the engagement ends, and the shortage and the blockage are still there, now with a production system on top of them that nobody in-house fully understands. The organization is left dependent rather than equipped. It's the same wariness I brought to model and vendor lock-in, except this lock-in is people-shaped, which makes it harder to see and much harder to unwind.
The test of an engagement is what remains when it ends: you should be in a better position than before the consultants arrived. Ideally you won't need them anymore, because they've built your ability to fly on your own. In practice, that means the best engagements end by creating these roles inside the client: the in-house owners who absorb what gets built and keep it alive.
Someone Has to Own It
So the precondition is really a statement about people. Capability doesn't live in a platform; it lives in the people who can own it. That's why the generalist profile became the scarce asset of this era. It also means buy-versus-build was a false choice. Kirkland did both, and the 180 engineers aren't a hedge against the Palantir engagement. They're what makes it worth $500 million instead of an expensive pilot. Outside help compounds when someone in-house can absorb what gets built. Without that, you're chasing the headline without the thing that makes it work.
One last thing. If your first reaction to all of this is "but what about data regulation, particularly in the EU," then you're reading the wrong blog. That's not an objection; it's the problem you're supposed to be solving. The generalist stance is to go find the solution that makes AI workable in your company, not to hold up the regulation as the reason you can't. Kirkland, Clifford Chance, and White & Case didn't wait for the rulebook to make it comfortable. They built.
Was I building a career or avoiding one? Building one, it turns out. I was just early. And the route in is open from any direction now, for anyone willing to do the work.