The sequence is familiar. An organisation discovers it has an AI governance gap. A risk committee asks who owns model oversight, and the honest answer is nobody. So a job description is written. The title is something like Head of AI Governance. The market for that title is thin, the salary conversation is uncomfortable, and after several months a capable person arrives.
Then the second discovery happens. The new hire has experience, but there is nothing for that experience to attach to. No review method. No escalation criteria. No case history. Nobody in the organisation has previously had to decide how grounded an agent's answer needs to be, what evidence should trigger escalation, which controls belong at runtime versus ingestion, or what a human reviewer should actually be inspecting rather than rubber-stamping. The hire understands how similar problems were solved elsewhere, but there is no local method to adapt or extend. The hire spends a year building these answers from nothing, often leaves before finishing, and the organisation starts the search again — convinced the problem is retention.
The problem is not retention. The deeper error is assuming AI governance is already a mature profession — that somewhere out there is a consistent, codified body of practice you can hire into, train into, or buy as a framework. There is emerging practice and a growing number of experienced practitioners, but nothing yet codified or consistent enough for most organisations to rely on. Hiring struggles because there is no mature profession to recruit from at scale. A two-day workshop struggles because there is no settled practice to teach, only fragments. A bought framework struggles because it was built around someone else's fragments, which have not matured into a profession either. Three different remedies, attempted by three different departments, and one root cause underneath them all: organisations are behaving as though the profession already exists.
What a hire actually transports
A senior hire carries individual competence: pattern recognition, accumulated experience, scar tissue from previous mistakes. What a hire cannot carry is institutional competence — the method, the case library, the supervision structure, the shared standards that let twenty people make the same decision the same way. Those live in the organisation the person came from, and they stay there. While the profession itself is still being assembled, even the most senior hire arrives carrying less institutional competence than their title suggests, because most of it has not matured anywhere yet to be carried from.
This is the distinction organisations consistently miss when capability is urgent. A practice is not a collection of capable individuals. It is an apparatus that produces capable individuals, faster than attrition removes them. If you have the apparatus, ordinary hires become strong practitioners within a year. If you do not, exceptional hires plateau, because there is nothing beneath them.
A practice is not a collection of capable individuals. It is an apparatus that produces them faster than attrition removes them.
What building practices teaches
I have spent the better part of four decades establishing labs, analytics functions, and the training frameworks that sustained them. The settings varied. The pattern that produced durable capability did not. Every practice that outlived its founders was built the same way: a small core of senior people, a method codified early and revised often, structured apprenticeship against real cases, explicit gates that certified competence before granting autonomy, and a refresh cycle that kept the method honest as the environment changed.
None of these practices began with mass hiring. Most began with three or four people and a discipline about writing down what they knew. The writing down is the part organisations resist, because the experts are busy and their knowledge feels too contextual to document. It is the same resistance, and the same mistake, every time. Expertise that is not specified cannot be taught, and expertise that cannot be taught dies with the roster.
AI governance is not the first enterprise capability to begin this way. The first ERP implementations looked remarkably similar. Organisations did not start with implementation methodologies, certification programmes, and standard operating procedures — those came later. They started with experienced practitioners making thousands of design decisions that had never previously been codified, and over years those decisions hardened into templates, frameworks, and eventually a profession people could be hired into and trained for. Only after those methods existed did organisations begin hiring ERP specialists at scale, because new practitioners were joining an established capability rather than inventing one. AI governance is on the same trajectory. The uncomfortable part is that most organisations are behaving as though the codification has already happened, when they are still standing at the beginning of it.
What a training framework actually contains
A training framework is not a course catalogue and it is not an onboarding deck. In the practices that worked, it contained five things. A competency map: the specific judgments the role must make, stated as decisions rather than topics. A graded case library: real, anonymised cases ordered from clear-cut to genuinely contested, because judgment is built on the contested ones. Supervised repetition: practitioners deciding real cases with a senior reviewer comparing their reasoning to the standard, not just their conclusion. Certification gates: defined points where demonstrated competence, not tenure, unlocks autonomy. And a refresh cycle: scheduled revision of the method whenever the models, the regulation, or the case mix changes.
Notice what this requires as a precondition: the method must exist. You cannot build a case library for standards that were never articulated. This is why the training framework and the governance framework are the same project wearing two names. Specifying how decisions should be made is governance. Teaching people to make them that way is training. Organisations that treat these as separate initiatives end up with policy documents nobody applies and training nobody can act on.
The arithmetic of grown capability
Hired capability is linear and leaky. Each hire adds one practitioner and a departure subtracts one, taking the accumulated experience along. Grown capability compounds. The case library gets richer with every contested decision. Each certified practitioner becomes a supervisor for the next cohort. The method improves because it is used, challenged, and revised by dozens of people rather than held in the head of one.
There is also an arithmetic of cost that boards rarely see laid out. The fully loaded cost of a senior external hire, the months of vacancy, and the restart cost when they leave will, in most organisations, fund the construction of a training framework that makes the existing team capable — people who already understand the business, which is the half of the expertise no external hire arrives with.
Where to begin
The starting point is unglamorous: pick the single highest-volume judgment your AI programme depends on, sit your best people in a room, and make them argue about ten real cases until the criteria they are actually using are on paper. That document — rough, contested, incomplete — is the seed of the method, the first entries in the case library, and the first lesson of the curriculum, all at once.
Organisations tend to discover that their experts agree less than anyone assumed. That discovery is not a failure of the exercise. It is the exercise. Every disagreement surfaced and resolved in that room is one that will no longer be resolved inconsistently, invisibly, in production. Hiring cannot give you that. Neither can a workshop. Only building can.
Organisations often ask whether they should hire an AI governance leader or train their existing team. The more useful question is what that person inherits on their first day. The organisations that build methods before they build teams will discover something surprising. Hiring becomes easier, training becomes meaningful, and capability begins to compound instead of restarting with every resignation. AI governance does not become scalable because the market suddenly produces more experts. It becomes scalable because the organisation finally creates something worth joining.