AI Can Do the Work. That Doesn't Mean It Wins the Market.
The market keeps making the same mistake: assuming the company with the smartest model automatically owns the market it can touch. Watch AI write production code, summarize a deposition, or close a support ticket without a human in the loop... the conclusion feels inevitable. --> The labs are going to own legal, healthcare, finance, enterprise software, all of it. The logic sounds airtight. It is also missing the hardest part of commerce.
In most industries the hard part is not doing the work. It is getting trusted to do it inside someone else's business. Being able to do the job is not the same as being the company people hire.
The Question Buyers Actually Ask
Forget the demo. Picture the procurement meeting.
The CIO signing an enterprise AI contract wants to know whether it integrates with the existing system of record and who carries the liability when something goes wrong. The GC approving AI for contract review wants to know whether the output will hold up in court. The CFO greenlighting AI for the accounting close is asking whether auditors will accept it. None of them are asking which model scored highest last quarter. Because that is not the question that matters when the workflow breaks and someone has to explain it to the board.
What buyers are really asking is: when this goes sideways, who do I call, and whose contract already covers it? That is a question most frontier labs cannot answer as cleanly as the vendor already embedded in the workflow, already sitting across from legal and compliance, already named in the master service agreement. Which is why "our current provider added AI" keeps beating "switch to a new platform and rewire your systems" in enterprise budget reviews. Not because the incumbent is smarter but because the buyer already knows where to send the escalation. They are not purchasing intelligence. They are purchasing accountability.
What the Evidence Shows
This is not a forecast. It is already the pattern.
Salesforce's AI handled 380,000 support conversations on its own help site (84% resolved without a human, 2% escalated) running inside existing infrastructure, on existing customer data, sold to customers already on Salesforce contracts. Thomson Reuters' CoCounsel, now used by over a million legal professionals, is built around Thomson Reuters' own legal content and sold as AI inside the workflow lawyers already use. Not as a standalone product asking them to change how they work. Oracle embedded clinical AI into its healthcare platform and reduced physician documentation time by roughly 30%. Microsoft pushed ambient clinical documentation into ambulatory care through athenahealth's platform rather than selling directly to physicians. In every case, AI moved through a relationship that already existed. Not because the frontier labs could not do the work, but because the buyer was not shopping for a new vendor.
ServiceNow's insight is simple: in enterprise software, the moat is not owning the best model. It is owning the workflow where the model gets used. Its AI bookings more than doubled year over year in Q4 2025 and that money went to the platform controlling the process. Not the one that trained the weights.
Where the Labs Actually Win
This argument has real limits, and they deserve honest treatment.
The dynamic reverses when the user is also the buyer. In software development, developers do not wait for procurement approval. They try the tool, see whether it makes them faster, and advocate for it afterward. GitHub Copilot has over 20 million users, and Claude Code hit a reported $2.5 billion annualized run rate within roughly a year of launch. When the feedback loop is measured in hours and the person evaluating the tool is the same person doing the work, direct adoption breaks through without waiting for institutional permission. Search follows the same logic: Alphabet's AI Overviews has 2 billion monthly users, search revenue grew 17% in Q4 2025, and the AI layer reinforced the incumbent's distribution rather than threatening it. Consumer assistants (OpenAI at over 800 million weekly users, Meta AI at over 1 billion monthly) built their own distribution through habit and interface, no enterprise contract required.
Across all three, the condition is the same: the user can adopt directly and does not need an institution's approval. Remove that condition, and the dynamic inverts. Healthcare systems, law firms, and corporate finance departments do not move like developers or consumers. They buy through existing relationships, they care more about accountability than benchmark performance, and the person recommending a new vendor owns the outcome if it fails.
The Actual Outcome
The end state is not one AI super-company sitting on top of every industry it can touch. It is something stranger and more structural: intelligence centralizes, but market ownership does not.
Most companies will eventually run on AI, buying it through vendors they already trust, embedded in systems they already use, on contracts they already have. AI may prove to be the most widely distributed technology shift in decades. While the companies that built the underlying capability capture only a fraction of the commercial surface it creates. The labs supply the cognition. The incumbents hold the customer.
The labs may train the model. The incumbent still gets the purchase order.