AI use 25% — redactional
Last Thursday, Microsoft announced the Frontier Company: $2.5 billion and 6,000 engineers whose job is to go inside client organizations and make AI actually work. Not sell licenses. Not run webinars. Sit inside the client, build the system, run it.
Two days earlier, Amazon had committed $1 billion to the same idea. OpenAI and Anthropic both launched deployment ventures in May, backed by private equity money. Judson Althoff, who runs Microsoft’s commercial business, wrote that this “goes beyond what has been labeled as Forward-Deployed Engineering.” Maybe. But let’s be honest about what just happened: the companies that spent three years telling us the model is the product are now spending billions on people.
I think this is the most important signal of the year, and I think almost everyone is reading it wrong. The consensus reading is “the labs are going after the consultants’ lunch.” My reading is different: we are watching the industry rediscover, in real time, what IBM knew in 1988.
Let me explain, starting from the factory floor.
The 40% problem
At OSS Ventures, we build software companies for industrial operations. Since January, our portfolio and studio teams have deployed more than a hundred agentic systems inside factories and supply chains. The gains are real and they are large. One program at a plastics compounder was validated by the client’s finance team at $4.9 million in recurring annual gains. A sock factory in Turkey, of all places, is on track for eight figures. This is not pilot theater. This is money.
But here is what nobody puts in the press release: in a typical deployment, roughly 40% of the data the agent needs does not exist.
Not “exists but is messy.” Does not exist. And once you see why, you can’t unsee it. Every system of record in the enterprise, the ERP, the MES, the CRM, the QMS, was designed to record what happened, not to automate what is being done. They are accounting instruments. They capture the transaction after the fact, at the granularity finance or compliance required, and nothing more. The tacit routing decision the scheduler makes at 6 a.m., the reason the operator overrode the setpoint, the actual sequence of a changeover: none of it is in any system, because no system ever needed it. Humans carried that context in their heads and the record was written afterward.
An agent cannot work from an after-the-fact ledger. It needs the operational reality, and the operational reality was never digitized. So every serious deployment starts with an unglamorous phase of building the missing 40%: sensors, interfaces, ontologies, small data products that capture work as it happens rather than after it happened.
The market has already voted
If you think this is just a deployment anecdote, look at public market prices. Most of the listed system-of-record companies have lost around half their value in the last twelve months. The market has quietly repriced a belief that dominated enterprise software for two decades: “whoever has the data wins the AI era.”
It turns out the incumbents don’t have the data. They have a record of the data’s shadow. Their real moat, and it is a real moat, is organizational and brand-based: they own the budget line, the procurement relationship, the IT department’s trust, the muscle memory of ten thousand companies. That is worth a lot. It is not the same thing as owning the substrate AI needs.
Which leads to the conclusion I keep coming back to. We are not in a moment of synergetic system design, where AI gets layered elegantly on top of an existing digital estate. We are in a moment of first equipment. The estate has to be built.
The dinner
A few weeks ago I had dinner with a CTO I admire enormously, a woman who has been shipping enterprise technology for thirty years and has seen every cycle since client-server. We were comparing notes on the state of enterprise AI: enormous productivity gains available, near-zero installed base of the systems needed to capture them, and a deployment effort that dwarfs the software itself.
At some point she stopped and asked the question that produced this post: what is the historical equivalent? A disruptive technology that could drive enormous productivity gains, but that required heavy, hands-on deployment into organizations with essentially zero equipment rate?
We looked at each other and said it at the same time. IBM.
Nobody remembers the AS/400
There is a whole chapter of computing history that the current generation has simply never heard. Before the AS/400 and its ancestors, a mid-sized company had no computing equipment at all. Payroll was clerks. Inventory was cards. The productivity gains from automating that work were staggering, and everyone knew it, but the gains did not arrive by shipping a box. IBM sent people. Systems engineers and consultants went into the client, studied how the work was actually done, and built ad-hoc systems around the machine. The machine was almost the excuse; the deployment was the product.
And here is the part I find delicious: it was fashionable. Working at IBM in that era carried the prestige that working at a frontier lab carries today. Being one of those consultants was the hot job, exactly as being a forward-deployed engineer is the hot job in 2026. The results were huge, the demand was insatiable, and IBM built one of the great business machines of the century on it. Not on the hardware margin. On the deployment, and then on the annuity that followed: maintenance contracts, the recurring revenue attached to systems that IBM’s people had built and that nobody else understood.
I think we are walking into five years that will rhyme with that era almost beat for beat. Immense demand for FDEs. Ad-hoc systems built inside individual companies, with great economic value, because the missing 40% is different in every plant and every supply chain. And then, the interesting question: what is the modern equivalent of the maintenance contract?
My answer: the productization of ontologies. Once you have deployed into thirty factories in a vertical, you notice that the missing 40% is different everywhere but rhymes everywhere. The supply chain ontology, the R&D ontology, the quality ontology: these harden into modules. The ad-hoc work funds the discovery; the vertical modules become the annuity. That is where the recurring value of this era will sit, and it is exactly the position IBM’s maintenance business occupied.
The problem the labs have
So why am I not simply predicting that Microsoft, OpenAI and the other labs win this era outright? They have the capital, the talent and now the deployment arms. Two reasons, and the first one was said out loud on national television last week.
Alex Karp went on CNBC and accused the frontier labs of charging enterprises for tokens that create no value while absorbing their IP and their “alpha.” Discount the theatrics, and the fact that he was there to sell Palantir’s alternative. The structural point survives the messenger. When your deployment arm’s parent company gets paid per token, there is an obvious conflict of interest baked into every architecture decision. Will the embedded engineer design the lean system or the token-hungry one? Enterprises are not stupid, and by Karp’s account many of them are livid. I hear milder versions of the same sentiment in every industrial boardroom I sit in.
The second reason is China. Chinese open-weight models are improving fast, they are cheap, and undercutting the economics of American closed labs is not a side effect, it is a stated objective. Coinbase reportedly cut its internal AI spend by nearly half by defaulting to Chinese open-weight models. When the model layer is being actively commoditized by a competitor that does not need it to be profitable, building your entire enterprise strategy on token margin is a fragile position.
Where I land
Put the pieces together and here is my prediction for the next five years.
The winners of enterprise AI will not be the model companies. They will be verticalized, AI-native companies, one or a few per vertical, that own the ontology of their domain: what a changeover actually is, what a stockout actually costs, how an R&D loop actually runs. They will do the unfashionable first-equipment work, deployment by deployment, and productize the modules as they go. And rather than fighting the Capgeminis and Accentures of the world, they will partner with them, because the dollar demand for deployment is about to be larger than it has ever been in the history of enterprise technology, larger than any one company, even a $2.5 billion Frontier Company, can absorb.
In other words: the AI era’s enterprise winners will look less like the labs and more like IBM circa 1988. Prestigious deployment forces. Ad-hoc systems with enormous returns. Annuities built on the modules that emerge. It was a very good business the first time.
Nobody remembers the AS/400. Somebody is about to rebuild it.