Every business I know is preparing to transform with AI. Very few are actually transforming.
The pattern is always the same: someone forms a working group. They commission a governance framework. They order a data readiness assessment. Consultants arrive to define what “AI-ready” looks like. Eight, then eighteen months pass. The board is still asking for a strategy.
The problem is not the technology. The problem is the sequencing.
Governance-first is a holding pattern
Starting with governance has a comfortable logic. Before deploying AI at scale, you should understand your processes, your data, your risk exposure, your compliance obligations. None of that is wrong.
But governance-first becomes governance-only and governance-forever. Every time you get close to starting, there is another layer to add. The data is not clean enough. The semantic model is not finished. The risk framework needs another review. You can always find a reason governance is not ready, because governance is never fully ready. It is a living system, not a gate.
The enterprises actually moving have reframed this: they do not treat governance as the prerequisite. They treat the pilot as the thing that tells them what governance needs to look like.
What the sequencing actually looks like
The companies making progress pick the highest-ROI use case: specific, narrow, measurable. They build the context layer for that scope. They prove something. Then they expand.
Most AI programmes underestimate the work in the context layer. AI does not just need data. It needs the raw data, a semantic layer that explains what the data means and how key metrics are calculated, and documented workflows that describe how the organisation actually operates. An agent with access to data but not to business semantics produces outputs that sound plausible and are wrong in ways that are hard to catch.
This is why broad programmes stall where narrow pilots succeed. A broad programme asks you to solve the context problem enterprise-wide before you start. A narrow pilot asks you to solve it for one use case — tractable, fast, and it teaches you everything the broader programme actually requires.
The pilot is not the simplified version of the real thing. It is the real thing. Everything else is expansion.
The accountability gap
Governance frameworks are safe. Pilots are not.
A framework can always be refined. A pilot has a result. Someone is accountable for whether it works. In organisations where AI is owned by a committee rather than a person, where success criteria are vague, where the working group reports to nobody who ships anything — governance becomes the perfect substitute for commitment.
I have seen this before. The working group is diligent. The documentation is thorough. The slide decks are polished. Nothing ships, because nobody is on the hook for a specific outcome by a specific date.
Ask the working group: what result are we trying to prove, and when will we know if it worked? If there is a clear answer, governance becomes useful — you are now asking what governance this specific scope actually requires. If there is no answer, no amount of governance work will help.
What this means in practice
The useful frame is not “AI transformation” but “AI sequencing.”
Start with a use case narrow enough to build the full context layer in weeks, not quarters. Define done before you start. Put someone’s name against the result. When it works, use that proof to define what governance looks like at the next level of scale.
I wrote about one version of this in Codify Before You Automate. Organisations that rush to automate before documenting what they actually do discover months later that their agents produce outputs nobody trusts, because there was nothing to ground them. Sequencing is not just about governance. It is about not skipping the unglamorous work.
The governance-as-product-problem angle I explored in The Brittleness Paradox is related — but there is a prior step. Governance does not become a product problem until someone is accountable for the product. Until then it stays a process problem, which means it stays stuck.
Most enterprises have the same models, the same platforms, the same vendors. What separates the ones transforming from the ones perpetually preparing is not what they have. It is the order in which they do things.
Governance is not the enemy of AI transformation. Governance-first is.
Leave a Reply