The bottleneck moved.

On July 15, Anthropic, Blackstone and Hellman & Friedman introduced Ode, a standalone enterprise AI services firm. The venture was first announced in May, backed by roughly $1.5 billion in committed capital from a consortium that includes Goldman Sachs, General Atlantic, Apollo and GIC. Its model is forward deployment: engineers embedded inside client companies, building and shipping AI systems from within, on a reported team of about a hundred where more than half are former founders. The stated principle is Claude-first, not Claude-only.

Read that announcement as a price signal. A frontier lab does not help stand up a billion-dollar services company to sell what its API already sells. It does so because something between the model and the customer is stuck, and the stuck thing is valuable enough to build a firm around. The models are good and getting cheaper. What has not gotten cheaper is the step where a company’s actual operations meet the machine: which approvals matter, where the real data lives, what the support team knows that the wiki does not, why invoice matching works the way it does. Palantir proved years ago that this step does not ship itself; it named the job forward-deployed engineering, and now the AI industry has adopted both the job and the name.

It is worth being precise about what the gap is made of. It is not a skills shortage, and it is mostly not an integration problem in the plumbing sense. It is that the knowledge an agent needs in order to act inside a company was never written down as structure. It sits in people’s heads, in old threads, in the habits of whoever has been there longest. Before an agent can do real work there, someone has to excavate all of that and turn it into something a machine can follow. That excavation is the product services firms sell. Billed by the engagement, it is expensive enough to support a $1.5 billion capital commitment.

Call it what it is: a retrofit tax. Companies that adopted AI after they had already calcified pay someone to dig their own operating knowledge out of the walls. The tax is real, the market for paying it is enormous, and the firms collecting it are doing genuinely hard work. But a tax on retrofitting is only owed by the retrofitted.

There is another way to not have the problem, which is to never create it. A company built AI-native from day one writes its operating knowledge as structure from the first week: decisions land in a repository with provenance, workflows are files an agent can read, the map of how the company works is the same map the humans use. Nothing needs excavating because nothing was ever buried. We have run this way since June: one shared brain, agents with jobs, humans making the calls. When a new capability ships in a model, it lands here the same afternoon, because there is no engagement to scope and no context to reconstruct. The context was never lost.

The honest caveat is that being AI-native is cheap only if you start that way. We started with a repository and nothing to migrate, which is exactly when starting that way costs nothing. An enterprise with thirty years of undocumented process cannot copy this, and that asymmetry is the whole point. The largest companies in the world are about to spend a decade and a trillion dollars paying down the difference between how they operate and what their software knows. New companies get to skip the debt entirely. The bottleneck moved from the model to the map, and the cheapest map is the one you drew as you walked.