The forward-deployed engineer is the most honest product the AI labs have shipped

For two years, the pitch has been that AI models would absorb the messy middle of software work. The premise is simple: you describe the outcome, the system produces it, and the layer of people who translate business problems into working software gets thin. The agency, the systems integrator, the internal platform team, all of that was supposed to compress.

Two weeks ago, Anthropic and three Wall Street firms put around $1.5 billion behind doing exactly that work by hand.

The venture pairs Anthropic with Blackstone, Hellman & Friedman, and Goldman Sachs, and embeds Anthropic's own applied AI engineers job is to sit next to a customer's staff and wire the model into the business. A few hours before that announcement, Bloomberg reported OpenAI was raising for almost the same setup, a venture called The Development Company (an unusually plain name), around $4 billion across nineteen investors. Goldman's Marc Nachmann described the goal as democratizing access to forward deployed engineers, a phrase worth sitting with, because it names exactly what the scarce resource turned out to be.

Spoiler: It was not the model.

The AI confession

A forward deployed engineer is a Palantir invention. You take your best engineer, the one who knows everything about how the system works, and you put them physically inside the customer, next to the person who knows everything about the customer. Neither of them can finish the job alone. The whole model exists because we're finally realizing the hard part of enterprise software is the collision between what the tool can do and what this specific, regulated, legacy-shaped company actually needs.

Spend a decade shipping software into other companies, and you learn this in your bones long before a press release confirms it. Anthropic's own description of a typical engagement has its engineers sitting down with a customer's staff to build tools that fit the workflows people already use. That is consulting with a model baked into it.

The deployment gap is the part worth being precise about, because no model release closes it. The distance between a demo that stuns a boardroom and a system that survives its sixth month in production is structural.

There is a cynical version of this story, and it deserves air. The AI labs sold autonomy, found that autonomy does not deploy itself, and are now selling the humans to repair the thing they sold. The Palantir model both ventures are copying is sticky on purpose. A custom system that an embedded team wires into a company's data, workflows, and compliance posture over six months does not get ripped out when the team leaves. It becomes load bearing. The company depends on whoever built it for as long as it runs. For the companies signing these engagements, that dependency is the actual product, and it should be priced into the decision as carefully as the integration itself.

The honest read is the more useful one, though. Strip the press release and the move is simple. The labs found where the money actually sits and are walking (running) toward it with a billion dollars in each hand. For every dollar a company spends on software it spends several more on services, and that ratio did not appear because buyers are irrational. It exists because turning capability into outcome inside a specific company is genuinely hard work that does not generalize. It is the first time in a while that what the labs say and what is actually true have lined up.

So the thing to watch is no longer the next model release, because the people closest to a company's weirdest problems are still the bottleneck, and they've realized no release cadence change will change that. The thing worth investing in is the small set of engineers who understand both the model and a given domain well enough to be very dangerous with both, because that exact pairing is what a billion and a half dollars just got priced at.

The question that now decides a build-versus-buy call is who owns the integration after the embedded team choppers back out, and how much of the operation ends up dependent on a vendor's roadmap. The model will be fine. The model is almost never the part that fails.

The unglamorous middle, the stretch between the impressive prototype and the thing that actually runs the business, has been the whole job in this industry for a long time. It is a little strange to watch it get a military title and a Goldman Sachs valuation. It is also, finally, the right thing to be charging for.

Has a slick demo ever sold a leadership team on something that fell apart by month six? If so, the model was the easy part, and the hard part now has a $1.5 billion price tag.