Training is not adoption

Nearly every enterprise I have worked with has bought AI training, and most of them have little to show for it. The pattern is consistent. Seats get licensed, a vendor runs a workshop, usage spikes for a few weeks, and within a quarter the dashboards are flat and the interesting work has moved back to personal accounts. The training was fine. The problem is that training was treated as adoption, and they are different things. Adoption happens when a person ships a governed workflow into production and answers for it afterward. Anything short of that is awareness, and awareness does not change how work gets done.

Why the seats go quiet

The seat license is the tell. Enterprises have been here before, with collaboration suites, analytics platforms, and every cloud console that arrived with a free tier of enthusiasm. The program declares victory on the day the invitations go out, and then the usage curve does what usage curves do when nothing in anyone’s job has changed.

The workshop produces a person who knows the tool exists and has watched it do something impressive on a clean example. Production requires a person who has taken a messy workflow from their own team, scoped it, built it inside the controls the organization actually has, and signed their name to its behavior. The gap between those two people is not more content. It is a shipped artifact and the accountability attached to it.

What certification should test

Most AI certifications I have reviewed test whether someone knows the API. That is the wrong exam. Knowing the API is a week of reading. The scarce skill is the sequence that turns a capability into a governed workflow, and that is what a practitioner should be certified against.

Scope a workflow. The candidate picks one real step from a real process, names the input, the output, and the definition of right, and states what the system will not do. “Add AI to intake” is a mood. “Classify inbound requests into these six queues and route the ambiguous ones to a person” is a scope.

Produce its two numbers. Before anything is built, the candidate forecasts what the workflow returns and what it will cost to run, with a ramp and a range, following the method in the two-numbers essay. A certification that does not require the second number is certifying people to write the business cases finance eventually freezes.

Build it on the paved road. The workflow uses the approved model access, the approved data classes, the approved identity layer. No personal keys, no side channel. If the paved road cannot carry the workflow, that is a finding about the road, logged and escalated, not a reason to drive around it.

Instrument it. Every model call is tagged with the step and the unit of work it served, so the forecast can be checked against telemetry within the first month. A workflow with no attribution is one whose cost nobody will ever state.

Hand it off. The candidate writes the runbook, names the production owner, and walks that owner through the failure modes. If the candidate is the owner, they document what it would take to hand it to someone else.

Nothing on that list is about prompt wording. All of it is about whether the organization can trust the next thing this person builds.

The artifact is the exam

The structure that follows from that list is a cohort, not a course. Each cohort builds a real workflow in a governed environment, and certification follows deployment, not attendance. In my work leading enterprise AI delivery, the artifact is the exam. There is no written test at the end, because a written test cannot ask the question that matters, which is whether the thing runs in production and whether someone is answering for it.

This changes who shows up and why. A cohort that has to ship picks problems its own teams care about, because the work has to survive past the last session. It changes what the facilitator does: less lecturing, more design review, more sitting with a team while they discover that the data class mapping does not cover the field they need. That discovery, made in a governed sandbox with an instructor in the room, is the point. Made for the first time in production, it is an incident.

Certification, in this model, is a record that this person has shipped one governed workflow and handed it off cleanly. The next workflow is the renewal.

Build in a governed sandbox

The environment the cohort builds in matters more than the curriculum. Habits form where the work happens, and a workshop that runs on a laptop with a personal key and a toy dataset teaches the habits the program exists to replace.

The environment I insist on is shared, multi-tenant, and governed the way production is governed. Each team gets an isolated tenant: its own data boundaries, its own cost attribution, its own audit trail. The data classes are real: the same classifications and access rules the enterprise applies everywhere else, even when the records are synthetic or de-identified. The telemetry is real: the dashboards the cohort learns to read are the dashboards the production owner will read later. If tenants can leak into each other through shared indexes or shared caches, the cohort learns that isolation is optional, and that lesson follows them into production.

The closest prior era is cloud. Learning a provider in a personal account teaches you the service. Learning it inside the organization’s landing zone, with the tagging policy and the budget alerts already switched on, teaches you how your company uses the service. Only the second transfers to Monday morning.

Measure enablement like delivery

Most training programs I have seen report completion rates, satisfaction scores, and seats activated. Those are attendance metrics, and the flat dashboard a quarter in is what attendance metrics look like from the other side. The numbers I track are delivery numbers.

Governed workflows shipped per cohort. Not started, not demoed. Running in production, instrumented, with a named owner.

Time to first production deployment. How long from a cohort’s first session to its first workflow carrying real traffic. This is the number that says whether the paved road is actually paved.

Residual shadow usage. How much AI work is still happening in personal accounts and unapproved tools after the cohort has been through. If this does not fall, the program taught people that the approved path exists and failed to make it the path they take.

I do not have a defensible cross-industry benchmark for any of these, and I would be suspicious of anyone who claims one. What I can say is that a program reporting completion rates alone has not decided to measure itself.

Make the approved path the easiest path

Governance decides which path is approved. Change management decides whether anyone takes it, and the two are halves of the same discipline. A paved road nobody drives on is a budget line, not a strategy.

The leader’s job here is specific. Make the approved path faster to start on than a personal account. Make the sandbox available the day someone asks. Make the certification mean something, so a certified person ships the next workflow without relitigating the controls. Every point of friction on the paved road is a point in favor of the shadow path, and the shadow path never needed a workshop.

What I am still learning

Settled, in my experience: training without a shipped artifact decays, and the artifact-as-exam structure has held where I have used it. Open: how long a certification should stay valid, how much of the cohort structure survives when the facilitators are internal rather than external, and whether residual shadow usage can be measured honestly, since the people best placed to report it are the people doing it. My prior is that the structure transfers and the calibration does not, and I would like to be corrected by data.

Training is bought by the seat. Adoption is shipped by the owner.

Related: The first ninety days of an AI program and Governance is the strategy, not the checkbox.