The dominant story about enterprise AI is a story about generation. The machine writes, summarises, drafts and answers, and the promise is that it will do so faster and at greater scale than people can. That story is not wrong, but it is dangerously incomplete. For the settings I have spent my career studying — public administrations and large enterprises where decisions have consequences and must be defensible — the capacity to generate text is the least interesting thing an AI can offer. What matters is whether the system understands governance. Enterprise AI must understand governance, not just generate text.
What “governance” actually means here
Governance is not a synonym for bureaucracy, though it is often caricatured as one. Governance is the set of arrangements that determine who is allowed to decide what, on what basis, with what accountability, and with what record. It is how an organisation ensures that its decisions are legitimate, traceable and answerable — not merely that they were made. A decision that is fast, well-written and entirely ungoverned is not an asset to an enterprise. It is a liability that has not yet been discovered.
When we say an AI system must understand governance, we mean something quite concrete. It must know that a business case belongs to someone, that it passes through defined roles, that approval is an act with consequences, and that the reasoning behind a decision must survive long enough to be examined later. A system that produces a flawless requirements document but has no notion of who may approve it, or how the decision will be justified afterwards, has automated the easy half of the problem and ignored the half that actually protects the organisation.
A decision that is fast, fluent and ungoverned is not an asset. It is a liability that has not yet been discovered.
— Prof. Yannis Charalabidis
Why generation-first AI struggles in the enterprise
Generation-first tools tend to be adopted enthusiastically and then quietly abandoned in exactly the places where the stakes are highest. The reason is structural, not a matter of model quality. Such tools optimise for producing convincing output, but an enterprise decision is not judged by how convincing it sounds — it is judged by whether it can be defended when challenged. A confident, well-formatted business case with no traceable basis is harder to trust than a rougher one whose reasoning you can follow, because the polish hides exactly the assumptions that need scrutiny.
There is also an accountability vacuum. When output appears fully formed from a machine, it is unclear who owns it, who reviewed it, and who is answerable if it turns out to be wrong. Organisations sense this instinctively, which is why so much impressive AI output ends up as an unofficial draft that a human quietly reworks before anyone important sees it. The technology impresses in the demonstration and disappears from the decision, because the decision needs something the demonstration never offered: a governed path from input to approval.
Governance as a design principle, not a wrapper
The mistake is to treat governance as something you add afterwards — an approvals layer bolted onto a generation engine once the interesting work is done. Governance has to be a design principle from the start, because it shapes what the system is even trying to produce. If the goal is a governed decision rather than a persuasive document, then the system must be built to capture the reasoning, attach it to accountable roles, route it through the right approvals, and preserve the trail. Those are not embellishments on the output. They are the output.
This is what distinguishes governance-first enterprise AI from the tools that came before it. It reaches past ticketing systems, which manage requests but not the justification behind them. It reaches past portfolio management, which tracks initiatives but does not reason about whether they are well-founded. It reaches past requirements tools, which capture requirements but not the approvals, ownership and accountability that make a requirement legitimate. The addition, in every case, is governance — the connective tissue that turns generated material into decisions an organisation can stand behind.
A human-centric direction
None of this argues against the generative capability of modern AI. It argues for putting that capability in its proper place — as a powerful contributor to a governed process, not as a replacement for the process itself. The right role for AI at the demand stage is to improve the quality of the inputs to a decision: to clarify, to enrich, to structure, to make gaps visible. The decision, and the accountability for it, stays with people. That is not a limitation imposed reluctantly. It is the correct architecture for any domain where legitimacy matters as much as speed.
My conviction, formed over years of studying how institutions actually make and defend decisions, is that the enterprises which benefit most from AI will not be the ones with the most fluent generators. They will be the ones whose AI understands the governance it operates within — and is therefore trusted to sit at the point where decisions are made. AI should help organisations think more clearly, decide faster and stay answerable for what they decide. Governance-first is simply what that looks like when you take it seriously.
Key takeaways
- In the enterprise, the ability to generate text is the least interesting thing an AI can offer.
- Governance is about who decides, on what basis, with what accountability and what record — not bureaucracy.
- Generation-first tools impress in the demo and disappear from the decision, because decisions must be defensible.
- Governance must be a design principle from the start, with the human decision and its accountability kept with people.