Joe Procopio's recent column in Inc. declared that generative AI is over. I agree with more of it than the headline suggests. His core points hold: AI is a tool, not a replacement. Workflows that aren't stress-tested in the real world fail badly. And too many consultants skipped past the fact that data is the real moat.

But “over” is the wrong word. What's ending is the phase where enterprises bought generative capability and expected it to produce business performance on its own. What hasn't started, in any serious way, is the work that actually creates that performance: building intelligence that is specific to the organization.

Generation is not intelligence

Frontier models are extraordinary generators. They draft, summarize, code and converse. They are trained on the world's public knowledge, which is exactly why they know nothing about your enterprise. They don't know why your releases slip, which requirements your teams consistently miss, or which of your incidents are symptoms of the same design flaw.

That knowledge lives in your history. It's in your tickets, user stories, test cases, bug records, change logs and incident timelines. It is the only dataset no competitor and no model provider has. Procopio is right that data is the moat. I'd go further: the moat is not data sitting in systems. It's the intelligence you build from it.

What organization-specific intelligence looks like

A concrete example from my own work in SDLC.

We connected user stories, epics, test cases and years of historical Jira bug data for a retail e-commerce platform. Then we classified the gaps: missing test coverage, architecture gaps and data gaps. One pattern stood out. Localization and currency issues kept surfacing in production, and they had never been captured in requirements. No generic model would have found that, because it wasn't a coding problem. It was a pattern in that organization's history.

We then trained an open-weight model on that classified history. The requirements-to-deployment pipeline gained contextual intelligence specific to that organization. Gaps were flagged at the requirements stage, before a line of code was written. The results were shorter release cycles, fewer requirement gaps, better code quality and bugs eliminated rather than fixed.

That is the difference. Generative AI helped write the code faster. Enterprise intelligence prevented the defects from being designed in at all.

The right model for the right job

Frontier models for generation. Use them where breadth matters: drafting, reasoning across open-ended problems, interacting with people.

Purpose-built open-weight models for decisions. Use them where the decision depends on your history. Models like Mistral or Qwen, trained or fine-tuned on an organization's own classified data, can run inference that is cheaper, more controllable and governed inside your own boundary. The intelligence they encode becomes an asset you own rather than a capability you rent.

This isn't a “frontier versus open-weight” debate. It's an architecture decision: which model does what. Most enterprises haven't made it deliberately. They default to one frontier API for everything, then wonder why usage rises while performance doesn't.

Why so little is happening here

I see enterprise AI moving on three tracks.

  • Checking the legitimacy of work — should this task exist at all? This has been a perennial question since runbook automation and scripting.
  • Operating model, architecture and governance — a perennial question in every innovation wave.
  • Building enterprise intelligence and memory as an independent track — the one that's barely moving.

Most AI programs fold intelligence into tool deployment. A copilot here, an agent there, each learning almost nothing from the organization's accumulated history. Intelligence deserves its own track, its own data foundation and its own ownership, because it is what makes every other AI investment perform.

Where this goes next

If the hype phase is over, good. The questions that matter now are harder and more interesting:

  • What does your organization's history already know that your people and your tools don't?
  • Which decisions should run on intelligence you own, and which on capability you rent?
  • Are your AI investments eliminating the causes of poor performance, or just generating output faster?

Generative AI isn't over. It has simply stopped being the interesting part. The interesting part is enterprise intelligence, and for most organizations it hasn't started.

This is the first piece in a new series on enterprise intelligence: building intelligence into every application, process, data layer and infrastructure discipline so it drives peak business performance.