This is not an argument against platforms. The next wave of enterprise AI will run on them or it will not run at all. It is an argument about how they get built, because the industry is about to repeat a mistake it has already made twice, at precisely the moment the stakes are highest.
Two forecasts that cannot both be casual
Start with the apparent contradiction, because it is doing more work than either number alone.
Gartner says more than 40% of agentic AI projects will be canceled by end of 2027. Not paused. Canceled. The stated reasons are the classics: cost, unclear value, missing risk controls. Gartner's Anushree Verma adds that most agentic propositions "lack significant value or ROI," and that of the thousands of vendors selling agents, only around 130 are legitimate. The rest are doing what Gartner calls agent washing. Rebranded chatbots. Yesterday's process automation in a trench coat.
IDC, looking at the same horizon, sees agent use at the world's 2,000 largest companies growing tenfold by 2027, with the compute behind it growing a thousandfold. By 2029 it expects over a billion deployed agents worldwide executing 217 billion actions a day.
Read together, the two forecasts are not in conflict. They describe a sorting. A minority of organizations will carry enormous agent workloads on rails built to hold them. A large minority will cancel their projects after the pilot demo, having discovered that fifty agents with no shared identity, no shared observability, no shared cost accounting, and no shared kill switch are not a capability. They are fifty incidents waiting for a root cause analysis.
The sorting variable is the platform. Which raises the uncomfortable question of why platforms fail, because it is rarely the technology.
The infrastructure reflex
Enterprises know how to build infrastructure. Scope it, fund it, build it, migrate everyone onto it, declare victory. The model works when users have no choice. Nobody opts out of the network layer.
An AI platform is not that. Every intended user of an internal AI platform has an alternative, and the alternative has the best user experience money can buy. It is called a browser tab. The consumer AI tools your data scientists and business analysts reach for are one login away, and the shadow AI numbers show exactly how willing people are to take that exit: 65% of decision-makers already use unapproved AI tools when the sanctioned path is slower than the unsanctioned one.
So the platform team ships v1. It is secure, compliant, and architecturally correct. It is also harder to use than the browser tab. Adoption stalls at the two teams who were in the pilot. The steering committee responds the way steering committees do: with a mandate. Usage becomes policy. And the platform crosses the line from product to tax, at which point the best people route around it and the platform serves, in practice, as an elaborate compliance documentation system for work that happens elsewhere.
This is the trap. Not a failed build. A successful build that nobody chose.
The pattern has run twice before at full scale. In the 2010s the same reflex produced Hadoop clusters that became data swamps: technically operational, organizationally dead, eventually written off. In the early cloud years it produced private cloud platforms whose main achievement was making the public cloud look even better. Both times the diagnosis was the same. The platform was built for the org chart, not for the user.
What the research actually says
The field has, usefully, started measuring this. DORA reports that internal platforms, the shared foundation a company's own developers build on, are now close to universal in large organizations, with dedicated platform teams established at three-quarters of them. The performance findings are more sobering than the adoption numbers. Platforms improve productivity, but DORA found they can decrease throughput and change stability when they are not carefully managed. A platform is not automatically an accelerant. Badly run, it is friction with a budget line.
Then comes the finding that should reorganize every AI platform business case being written this quarter. In DORA's analysis, platform quality moderates the entire relationship between AI adoption and organizational performance. High platform quality, and AI adoption shows a strong positive effect. Low platform quality, and the effect is negligible. Not smaller. Negligible.
That is the quantitative version of something practitioners have felt for two years. MIT's GenAI Divide study found only 5% of AI implementations producing measurable profit-and-loss impact, and the difference was never model access. Everyone has the same models. The difference is the quality of the paved road between a model and a production workflow: data access, evaluation, deployment, monitoring, cost control. That paved road is the platform. Organizations spending millions on AI licenses while their platform is an afterthought are buying engines without building the car.
Run it like a product or write it off
What separates a platform that gets adopted from one that gets mandated is not a technology choice. It is an operating model, and it has a name: platform as a product.
The concrete version fits in one paragraph. The platform has users, not conscripts, and someone talks to them every week. It has a roadmap driven by what blocks those users, not by the architecture diagram's sense of completeness. It has golden paths, DORA's term for the paved route that makes the secure, compliant way also the fastest way. It starts as a minimum viable platform serving one high-value use case end to end, rather than a two-year foundation program that meets its first real user in year three. And it has one metric that matters: voluntary adoption. Not registered accounts. Teams that chose the platform when they had an alternative.
The same logic has already been priced for data. McKinsey's analysis of data products found that managing data as a product, with owners, consumers, and quality standards, delivers new use cases up to 90% faster and cuts total cost of ownership by 30%. Nothing about that finding is specific to data. It is a finding about what happens when internal capabilities acquire the disciplines external products have always needed to survive: a user who can say no.
For agents, the product framing stops being good practice and becomes the safety system. An agent that books nothing is a demo. An agent that acts needs an identity, permissions, an audit trail, a budget, and an off switch, and it needs them from one shared layer, because fifty teams each building their own version of those five things is how Gartner's 40% earns its cancellation. The organizations on IDC's side of the ledger will be the ones where shipping an agent onto the platform was easier than shipping one around it.
The go-live is the starting line
The historical rhythm is familiar by now. ERP taught the industry that implementation is not adoption. The cloud taught it that a migration without an operating model change just relocates the mess. The data lake decade taught it that storage is not strategy. Each lesson cost billions, and each produced a discipline that is now considered obvious.
The AI platform wave is early in the same curve, with one difference: the sorting will be faster. Agent workloads growing a thousandfold do not wait for a five-year platform program to mature. The organizations that treat their platform as a product with a first user this quarter will compound. The ones assembling infrastructure for a user population that exists only in the business case will join the 40%, and the write-off will be booked as a technology failure, and it will not have been one.
A platform's value has never been what it can do. It is what people do with it.