What We’re Talking About When We Talk About AI Transformation

Foundation models are remarkable, interchangeable, and improving at the same rate for everyone. Competing on model choice is a losing strategy — your competitor has access to the same ones.

↓ dig deeper ↓

A model is a reasoning engine. It is not a system.

Models don’t remember. They don’t connect to your database. They don’t know your policies, your customers, or your edge cases. A model without a harness is an incredibly capable intelligence that cannot actually do anything in your organization. The model answers questions. The harness does work.

Models are interchangeable. Your competitor has the same ones.

GPT, Claude, Gemini, DeepSeek — they are converging on identical capabilities at commodity prices. Whatever advantage one model gives you today will be replicated by every other model within months. If your strategy is “we picked the best model,” you have no strategy.

The industry has already absorbed this lesson.

Every major implementation guide now spends more pages on orchestration, tool integration, and workflow design than on model selection. The conversation moved from “which model” to “which harness” because the people doing the work discovered that the model was never the bottleneck.

Determines what the model can do. Orchestration, tools, memory, guardrails. An important differentiator — but replicable. Your competitor can build one too.

↓ dig deeper ↓

The harness inherits the organization’s ambiguity. It cannot fix what the organization hasn’t resolved.

If nobody can decide who owns a cross-functional decision, the harness can’t either. If the org chart says one person decides but everyone knows someone else actually has veto power, the harness will route to the wrong person. The harness doesn’t know about shadow decision-makers. It only knows what you tell it — and what you tell it is usually the documented process, not the real one.

Governance designed for quarterly reviews cannot govern agents that decide in milliseconds.

An agent making 10,000 decisions per day will make roughly 900,000 decisions between quarterly steering committee meetings. If governance cycle time is measured in weeks and agent decision time is measured in milliseconds, governance is not governing — it is documenting decisions already made. The harness can enforce rules. It cannot create them at the speed the agent needs them.

The human repair layer — undocumented, invisible work — is what actually keeps processes running.

Every organization has people who catch exceptions, translate between incompatible systems, remember why decisions were made, and know where “enough” is versus “perfect.” This work is invisible by definition — if it were documented, it would be part of the process. When you automate the documented process without accounting for the repair layer, the repair layer becomes visible as failures, exceptions, and stalled workflows. The harness didn’t break anything. It revealed what was already broken.

Without organizational learning loops, each deployment starts from scratch.

Most organizations get better at deploying technology. Almost none get better at absorbing it. The difference is whether deployment N changes how you approach deployment N+1 — or whether you start from the same assumptions every time. The harness can be refined. The learning that refines it has to come from the organization. If nobody owns capturing what each deployment taught about how work actually gets done, the organization accumulates technology without accumulating understanding.

The harness matters more than the model. But the organization determines what the harness can do. Every unresolved ambiguity in the operating model becomes a constraint the harness cannot overcome — and a failure the model will be blamed for.

Determines what the harness can do. This is the foundation everything else rests on. Nothing above this layer can compensate for what’s missing here.

The Lesson of the Last Two Years. The model is a commodity. The harness is an advantage, but replicable. The operating model is the ceiling.

2023–2024: We improved the model

2025–2026: We improved the harness

2027→: We improve the organization