Every capability plan says the same thing about people, in the same voice. Everyone will be trained. Learning will be continuous. The culture will shift. None of those sentences names a consequence for anyone, which is why none survives a quarter in which something else is on fire.
A consumer and SME banking group operating across several European markets commissioned a board deliverable covering governance, operating model and skills. The skills half is worth reading, and not because the curriculum is clever. It is because the plan names five drivers of adoption, puts them in order, and ends the list with the only one that costs an employee something: certification as a requirement for promotion.
That ordering is the whole argument. It is also the part of the plan that will be argued about, and it should be.
The challenge
Start with the ceiling. The mandatory literacy curriculum for frontline staff is two hours long across five modules, and almost all of it is spent on refusals and detection rather than capability. What not to paste into a prompt. How to handle customer data under privacy rules. How to recognise a hallucination, a biased output, or a phishing attempt written by a model. When to stop and call a manager. Only the fifth module teaches anyone to do anything faster, and it does that with a prompt library.
Two hours is not modesty. It is roughly what an organisation can compel across an entire workforce without attaching a consequence to it. Everything above that ceiling is voluntary, and voluntary effort in a bank competes with month-end, with audits, with a queue. The design accepts this: the frontline targets it sets are behavioural rather than attendance-based. Eighty-five percent of staff using approved tools weekly. Fifteen percent fewer manual data entry errors and compliance breaches. Both are written as targets for a quarter ahead of the document, and neither is a number a completion certificate can produce.
The same discipline runs through the measurement. Knowledge lift is assessed before and after, experience on a recommendation score, and behaviour at ninety days against a target that eighty percent of participants are still using the new tools weekly. That third layer is the one most programmes quietly skip, because ninety-day retention is where a curriculum becomes either a habit or nothing.
The approach
Five adoption drivers, stacked from cheapest to strongest.
First, four protected hours a week per participant. This is the expensive line and it is listed first for a reason: every incentive that follows is unfair without it. Ring-fenced time is what separates a capability programme from a demand that people study in the evening.
Second, role-based tracks, so an engineer, a product owner and a branch employee are not in the same session.
Third, internal badging in three tiers mapped to job families. The foundational tier covers everyone. The middle tier pairs an external cloud vendor certification with an internal module in applied prompting for banking work. The top tier cannot be passed with coursework alone. Its qualifying evidence is a capstone: one production-grade agent, delivered.
Fourth, recognition through awards and bonus.
Fifth, career gates. Certification as a promotion requirement.
At the top of that stack the design does something specific enough to copy. Three consequences are wired to a badge. The top tier is a prerequisite for senior and principal promotion eligibility. The middle tier qualifies someone for the annual innovation bonus pool. And a separate internal licence is mandatory for anyone who owns or approves a model, which converts model accountability from a name in a spreadsheet into a qualification somebody had to earn and keep current.
One further detail gives the badge weight. Certifications expire between eighteen and twenty-four months, capped at twenty-four, and renewal is tied to retaining system access rights. A permanent credential in this field certifies a version of the world that has stopped existing, so expiry is the honest design, and attaching it to access is what makes renewal happen rather than merely being requested.
The governance placement matters as much as the mechanism. The council is chaired jointly from the people side and the digital side, and the operating lead sits inside HR with a dotted line to the AI centre of excellence. That splits two authorities that rarely sit together. HR owns the budget, the promotion process and the compensation cycle. The centre owns technical standards and credibility. Run it purely from HR and the curriculum ages faster than the tooling it describes. Run it purely from the centre and nobody's career depends on finishing it.
The outcome
Nothing here has been run. This is a board deliverable, and the plan inside it is a design.
That distinction matters, because the deck carries a capability dashboard that reads like a report. It shows eighty-seven percent of staff as active learners, an average post-training assessment of 4.2 out of 5.0, two hundred and forty-five certified badge holders against a target of three hundred, onboarding time down thirty-five percent, and a largest single proficiency gain of forty-five percent among product owners. It carries a status marker reading on track.
Every one of those figures sits in a reporting quarter that had not arrived when the document was written. That leaves them ambiguous between two readings, and the slide does not disambiguate: they are either the programme's projections for that quarter, or an illustration of the dashboard's layout with plausible values dropped into the cells. They are not measurements of anything that happened, and no reader should take the on-track marker as a status. A number that expects itself is ambition, not evidence.
The honest reading of the plan is therefore narrower and more useful. Of the five adoption drivers, four are things an organisation offers. One is a thing it withholds. That asymmetry is the finding, and it is why the promotion gate will be the contested item in every version of this conversation.
The objections are real. A gate rewards the people who got their four hours, which is why protected time is listed first and why the design is worthless if that line is cut in the first budget round. A gate turns a curriculum into a filter, which makes curriculum quality a fairness question rather than a training question. And tying expiry to system access means a lapsed credential is an operational outage for a team, not a private inconvenience. None of those is a reason to avoid the gate. All are reasons to build it with evidence underneath.
That is the part we would add. The same deliverable already requires a standard evidence pack at every gate in the model lifecycle: a model card, a risk assessment, a lineage record, test results. A career gate deserves the same treatment. If a badge decides a promotion, the assessment version, the evidence submitted, the assessor and the date have to be recoverable long afterwards, because the answer will eventually be read by a works council or a regulator rather than by a training manager. Most learning systems cannot answer that at all. Synapsa was built so it can, which is why we treat a capability programme as an instrumented system rather than a catalogue of courses, and why our Consult work starts by fixing what a badge is permitted to assert before anyone designs a module.
There is a regulatory floor under this as well. The AI Act's staff literacy obligation is in force, so an organisation deploying these systems has to show that the people operating them were prepared to. A two-hour mandatory module clears that floor. A promotion gate is not compliance. It is the bank deciding that the floor is not where it wants to stand.
Enthusiasm produces a pilot. Consequences produce a workforce. The plan is right that the gate is the strongest instrument it has, and right to place it last, because a gate with no protected time behind it is not a capability programme. It is a filter with a training budget attached.