Cloud-First Was Wrong. AI-First Is "Wronger"

The same lesson cloud taught us, with higher stakes: not every problem wants a probabilistic answer.

Early in the cloud hype cycle, there was a sentiment that public cloud would displace the private data center—not just as the primary way to deliver enterprise IT services, but entirely. The need for a private data center would go away. That didn’t happen. We learned that public cloud and private data center complement one another, and that not every use case is amenable to public cloud.

We’re now entering the phase of the AI hype cycle where we’re learning the same lesson. AI-first is not the position enterprises should take, because not every use case is amenable to AI.

But there’s a difference in how these two transitions actually work, and it matters more than the surface analogy suggests. The difference is in implementing processes with public cloud versus implementing processes with AI.

Public cloud, for all intents and purposes, still ran deterministic workflows. We wrote code, that code ran on public cloud services, and it determined the process. That made the transition a clean complement to the private data center. Once you accounted for the net-new cost, resiliency, and latency factors of moving an application into the public cloud, the compute looked the same. Same execution model, different location.

That is not the case with AI-centric workflows. AI is a probabilistic technology at its core. We try to put guardrails around it with agent harnesses, and even with the best engineering, we still end up with a probabilistic compute mechanism. The execution model itself is different.

This is not a flaw. For certain problems, you need a probabilistic approach. And this is exactly where the robotic process automation movement of a few years ago missed. RPA applied a deterministic underlay to problems that were probabilistic—the kind of problems we had previously put humans in front of. What we’re discovering now is that solving those problems at scale requires both: deterministic compute and probabilistic compute working together.

The market is arriving at this realization the hard way. Mark Zuckerberg reportedly sent a memo acknowledging a staffing mistake in Meta’s transition to AI, with thousands of people affected, on top of Meta’s recent layoffs. The early narrative around IBM, Dell, and other large technology companies cutting headcount because of AI productivity gains is being revised. The clearer picture is that today the technology augments human capability rather than replacing it.

In my own testing, what I’ve found is that AI combined with deterministic code in the loop—DCITL—lets organizations scale probabilistic problems. The AI models do the probabilistic work. The deterministic code is the guardrail.

I built a migration control plane powered by AI as a working example of exactly this pattern. The playbook drives execution, deterministic code performs the known transformations, validators define completion, and the LLM is called only for the bounded tasks that actually need interpretation. The probabilistic agent searches the solution space; the deterministic system decides whether the result is acceptable. Read it here: From Migration Factory to Migration Control Plane.