What 20 years in large companies taught me about AI adoption

I've been inside large organizations during three major technology transitions: the move from on-premise to cloud, the SAP S/4HANA migration wave, and now the AI moment. Each one was described by vendors as transformational. Each one was implemented by companies with widely varying results. I've been on the inside for all of them: first as a support engineer, then a project manager, now working across SAP, BTP and enterprise technology from inside the business.

Here's what I've learned about why most enterprise technology adoptions fail, and why AI is both similar to and dangerously different from everything that came before.

The pattern never changes

Every enterprise technology transition I've witnessed follows the same arc. A technology becomes impossible to ignore. The C-suite sees competitor announcements and board presentations. A decision to "do something" is made. A project is scoped. An implementation partner is hired. The project delivers something. The something is used by fewer people than planned. The business case is retroactively revised to make the ROI look acceptable.

This isn't cynicism. It's a pattern backed by direct observation. The failure mode is usually the same: the technology conversation starts before the problem conversation.

Every successful technology adoption I've been part of started differently. It started with someone in operations saying "this specific thing wastes our people two hours a day" or "we lose this many orders because of this specific delay." The technology conversation came second, as a response to a concrete problem with a concrete cost.

Why AI is different

Previous technology waves were easier to scope. Cloud migration: we move these servers here. SAP implementation: we configure these modules and train these users. The boundaries were visible. The handoff between consultant and company was clear.

AI doesn't work like that. The boundaries of what AI can do are genuinely unclear, even to experts. The right use cases are specific to each company's data, processes and organizational maturity. There is no standard AI implementation. There is only: what is your actual problem, what data do you have, and what does good look like for you?

This means that the typical enterprise buying process produces bad results with AI. The thing you need first is a discovery phase that most companies don't know how to pay for and many consultants don't know how to sell.

The question isn't "how do we implement AI?" The question is "which three problems, if solved, would change this business, and could AI plausibly solve them?" Those are different projects.

What the winners get right

The companies I've seen make real progress with AI share a few characteristics that have little to do with their technology budget.

They start small and specific. Not "AI strategy." One use case. One team. One process. Something where you can measure before and after. The wins from the first project create the credibility and learning that make the second project faster and smarter.

They own the problem, not just the project. In every failed AI initiative I've observed, there was a project owner but no problem owner. Someone was responsible for delivering the tool. Nobody was responsible for the outcome the tool was supposed to produce.

They treat data as a first-class citizen before day one. AI requires good data. Most enterprise data is messy, siloed and undocumented. The companies that succeed treat data cleanup as part of the real project, not as an annoying precondition.

The SAP-specific problem

I work in SAP-heavy environments, and there is a specific failure mode I see often. Companies have years, sometimes decades, of operational data locked in SAP. That data is enormously valuable for AI applications. But many AI consultants don't understand SAP data structures. And many SAP consultants don't understand AI.

The result: AI projects that ignore the richest data source the company has, while SAP teams watch from the sidelines wondering why the AI team is building something their existing landscape already supports better.

This gap is real, costly, and exactly where practical value exists.

The honest prediction

In five years, every enterprise will have AI running somewhere. The distribution of outcomes will look like every other technology wave: a small group of companies that got something real out of it, and a large group with systems their people work around.

The difference will be determined by the decisions companies make now. The winners will resist the pressure to have an AI strategy and focus first on having an AI use case: specific, measurable and connected to a real problem that real people have every day.