What stays scarce

The two most-read documents about near-term AI, Situational Awareness¹ and AI 2027,² both end in an intelligence explosion: models automate AI research itself, and a decade of progress compresses into a year. We are skeptical of that last step. But buried under the explosion is a quieter premise that deserves to be taken seriously on its own, because it is already visible in enterprise budgets. Cognitive work is collapsing in price.

The premise you can keep

Strip away the recursion and both documents make the same observation: models are becoming drop-in workers for a large class of white-collar tasks. Reading documents, watching how work happens, interviewing people, synthesizing findings, drafting reports. You can reject every 2027 timeline and still notice that this class of work is precisely what enterprises currently buy by the hour, at enormous prices.

Consider the arithmetic of a typical consulting diagnostic, the phase where a firm figures out how a client's operation actually works before changing it. Ten consultants for twenty weeks, fifty hours a week, at a blended $300 an hour is $3 million, and three to six months of calendar time. In our experience most of those hours, we estimate around 70 percent, go to gathering and synthesis: interviews, shadowing, reading process documentation, reconciling what people say with what they do, and writing it all down.

That 70 percent is the back office of a services firm. It is the analyst pyramid: the people who collect, structure, and summarize so that a partner can decide. And it maps almost exactly onto what models are good at today, no intelligence explosion required. Steady, boring diffusion of current capability is enough to reprice it.

The bottleneck moves

Here is the part both documents skip past. Speeding up a share of the work does something unintuitive to the whole. Amdahl's law, borrowed from computer architecture: if a fraction p of a project can be accelerated by a factor s, the whole project speeds up by

Make the automatable 70 percent of a transformation 10 times faster and the project completes 2.7 times faster. Make it infinitely fast and you hit a ceiling of 3.3.

The curve flattens because the remaining 30 percent starts to dominate. And that remainder is stubbornly human: deciding which of a hundred workflow variants to standardize on, getting a data owner to grant access, convincing a department head to change a process her team has run for a decade, carrying accountability when the change lands badly. When execution gets cheap, the binding constraint becomes trust. Scarcity moves, and value moves with it. Relationships appreciate.

The hybrid firm

This is why we think the firms that win enterprise AI will look strange to both software people and services people. Pure software stalls at the ceiling: it accelerates the 70 percent and then waits on decisions it has no standing to make. Pure services can make the decisions and hold the relationships, and then burns millions of dollars a project doing by hand the work software now does well.

The durable shape is a hybrid: software as the back office, experienced people as the front. The pyramid inverts. Instead of a partner leveraging thirty analysts, a small senior team leverages a platform that watches the work, reads the documents, runs the interviews, and drafts the maps and specs, with every claim cited back to evidence. The people spend their time where the ceiling is: in the rooms where conviction gets built and decisions get made.

That is what we are building at Cobalt. The platform does the gathering and the synthesis. The people we pair it with carry the relationships, the judgment, and the accountability. If the essays are even half right about the price of cognition, the back office becomes software everywhere. What stays scarce is the trust to act on what it finds.