Buying an AI Platform Is the New SAP
The license-then-adapt playbook is quietly repeating itself in agentic AI. Here is why it fails in workflow-heavy, regulated businesses - and what fit-first delivery looks like instead.
Every enterprise-software era has a moment when a broad platform gets sold as the answer, and the buyer spends the next two years bending the organization around it. Agentic AI is having that moment now. The pitch is familiar because we have heard it before - most memorably during the ERP boom, when SAP and its peers promised that one integrated system would run the whole company. The value was real. So was the tax: multi-year integrations, armies of consultants, and a business reshaped to fit the software rather than the other way around.
The agentic version of this pitch is arriving fast, and it deserves the same scrutiny. Buying a horizontal AI platform and adapting your organization to it is not a shortcut. In workflow-heavy, regulated businesses, it is the slow path wearing a fast path’s clothes.
What the SAP era actually taught us
ERP delivered for companies that already had disciplined processes and clean data. It punished the ones that did not, because the implementation surfaced every inconsistency and handed it back to the customer to resolve - at consulting rates. The platform never removed the hard work of fixing processes and data; it relocated that work into a customization project, on the vendor’s terms and timeline, with switching costs that grew every year. The lesson was not that platforms are bad. It was that a generic platform is only as good as the foundation you bring to it, and the buyers who bring a weak foundation pay the most.
The agentic replay
The same structure is forming in AI. Vendors are racing to ship broad agent platforms and sell them enterprise-wide. Merck signed an agentic deal with Google Cloud reported at up to $1B over more than a decade, standardizing on a single vendor’s agent framework across R&D, manufacturing, and commercial, with the vendor’s engineers embedded on site. Veeva, the incumbent system of record in life sciences, announced Falcon - an agentic platform delivering what it calls “agentic labor” against exactly the regulated workflows its customers already run. Across sectors the arms race looks identical: license the platform, then adapt your workflows, your data, and your controls to it.
The evidence is already in
The early numbers should give any buyer pause, and the through-line in every one of them is the same. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing cost, unclear value, and weak risk controls - and notes that pilots typically run at only 15 to 25% of the eventual production bill. A widely cited MIT study - Project NANDA’s “The GenAI Divide,” 2025 - found that 95% of generative-AI pilots returned no measurable P&L value, and concluded the gap was driven by approach, integration, and workflow fit, not model quality. McKinsey’s QuantumBlack reports that nearly two-thirds of enterprises have experimented with agents but fewer than one in ten have scaled them to tangible value, with eight in ten citing data limitations as the roadblock.
Read those failures closely and a pattern appears. Almost none of them are model failures. They are failures of data readiness, workflow fit, integration, and governance - the exact things a platform purchase does not solve. One MIT Sloan case study of an AI deployment found that 80% of the work was data engineering, governance, stakeholder alignment, and workflow integration. The intelligence was the easy part.
Why a platform relocates the hard part instead of removing it
Here is the structural problem. The value in a domain-specific AI system does not live in the platform layer. It lives in the authoritative data you retrieve over, the workflows you encode, and the evaluation harness that decides what is good enough to ship. Domain specificity is a stack, not a model - and the durable moat is the data pipeline and the eval harness, not the underlying weights. A generic platform hands you the least differentiated layer and still requires you to build the valuable ones. Only now you build them inside someone else’s system, on their roadmap, behind their lock-in. You have bought breadth and inherited the obligation to make it fit. That obligation is the SAP tax, re-priced for AI.
What fit-first looks like
The alternative is to invert the sequence. Instead of licensing a broad system and adapting the business to it, you build the agentic system around the workflow, inside the client’s environment, governed to the client’s standards, and owned by the client. Fit is the point, not a phase-two project. There is no adaptation tax, because there is nothing generic to adapt from, and no lock-in, because the client owns what ships. The hard work - data, workflows, evaluation, guardrails - still has to be done. But it is done in service of the specific workflow that will actually run, not a platform’s idea of it.
This is not anti-platform
None of this makes platforms wrong. A platform earns its place when a workflow is genuinely standard across organizations and the underlying data is already clean - payroll, expense management, and much of core ERP still qualify. The failure mode is narrower and more specific: buying breadth and calling it fit, in domains where the workflow is idiosyncratic and the data is not ready. And the honest way to productize agentic AI is the reverse of the SAP playbook - compose a platform out of real, working, fitted solutions once the pattern is proven, rather than selling the generic system first and discovering the fit later.
How to decide
The build-versus-buy question is answerable before you sign anything. A few signals tell you which way to lean:
| Lean toward a platform when | Lean toward a fit-first build when |
|---|---|
| The workflow is standardized across organizations | The workflow is idiosyncratic to how you actually operate |
| Underlying data is already clean and accessible | Data needs real engineering before any agent is useful |
| The regulatory bar is light or generic | Validation, audit trails, and human-in-the-lead are non-negotiable |
| You can absorb process change on the vendor’s timeline | The system has to fit your existing controls and systems |
| Lock-in and switching cost are acceptable | Ownership and portability matter |
The foundation, not the platform
The SAP era rewarded the companies that fixed their foundations before they bought the system, and punished the ones who expected the system to fix the foundations for them. The agentic era is shaping up the same way. The first question is not which AI platform to buy. It is whether your organization can absorb one - and whether, for the workflows that actually differentiate you, you would be better served by something built to fit. That is the question worth answering before the license is signed.
Sources
- Gartner, “Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” press release, 25 June 2025.
- MIT Project NANDA, “The GenAI Divide: State of AI in Business 2025.” Coverage: Fortune, 18 August 2025.
- McKinsey (QuantumBlack), “Building the foundations for agentic AI at scale,” April 2026.
- MIT Sloan, “5 heavy lifts of deploying AI agents” (Kellogg et al., 2025).
- Merck and Google Cloud, “Partner to Accelerate Agentic AI Enterprise Transformation,” 22 April 2026.
- Veeva, “Veeva Announces Falcon, an Agentic Platform and Standard Agents to Deliver Agentic Labor in Drug Development,” 27 May 2026.