Situation
A Swiss private bank had licensed enterprise-grade AI tooling but faced the usual paradox: adoption was low, and the little adoption there was happened outside of any control framework. Training people to prompt generic models was not the answer - it would have pushed adoption in directions the bank could not govern.
Friction
The friction sat at the intersection of two constraints. Business teams wanted usable AI now. The regulator, and the bank's own second-line functions, required that anything client-touching stayed inside defined controls. Any adoption programme that ignored either side of that constraint would either produce low usage or produce usage that the bank could not defend.
- Business demand for immediate usefulness.
- Regulatory and internal controls expectations.
- No middle ground available through generic AI training.
Intervention
TSG designed an AI Academy specifically for the bank. The curriculum did not teach prompting in the abstract. It taught operators how to run approved workflows - covering document formats accepted internally, guardrails required by the second line, human-in-the-loop behavior on sensitive tasks, and the boundary between what the tool could do autonomously and what still had to be routed to a person. Learning paths were mapped to the actual roles that existed in the bank.
- Role-based curriculum: relationship managers, compliance, IT, executive assistants with client access.
- Approved workflows taught as the unit of adoption.
- Guardrails and HITL behavior taught as part of the workflow, not as a separate policy.
- Boundaries between autonomous and human-owned decisions made explicit.
Operating model
The Academy is treated as an adoption factory, not as a training catalog. New workflows approved by the bank become new learning modules. Deprecated workflows become deprecated modules. The pace of the Academy therefore tracks the pace at which the bank actually approves new AI-supported ways of working. This alignment is what keeps adoption inside the control framework rather than outside of it.
Evidence
The evidence tracked is whether the operators are running the approved workflows on the tasks those workflows cover - and whether shadow usage on those same tasks is receding. Both are qualitative but observable. Quantified client outcomes are intentionally not disclosed in this brief.
Cas d’usage et impact
Enterprise AI licensed but adoption either low or happening outside of any control framework.
Solution: Role-based AI Academy teaching approved workflows - not prompting in the abstract - with guardrails and HITL behavior embedded in the curriculum.
- Adoption inside control framework: Marginal
- Shadow usage on sensitive tasks: High
- Second-line comfort with rollout: Low
- Adoption inside control framework: Primary path
- Shadow usage on sensitive tasks: Receding
- Second-line comfort with rollout: Alignment achieved
FAQ
Is this a Claude-only academy?
It is currently deployed with the enterprise-grade AI tooling the bank has licensed. The design is model-agnostic; what matters is that the approved workflows are the unit of teaching, not the model.
Why not just teach prompting?
Because prompting in the abstract accelerates adoption in directions the bank cannot govern. Teaching approved workflows accelerates adoption inside the control framework - which is the only kind of adoption a regulated firm actually wants.
How is this different from generic AI literacy training?
It is anchored in the specific roles, controls and workflows of one regulated firm. Generic AI literacy is table stakes; role- and control-aware adoption is where the actual behavior change happens.
Conclusion
In regulated finance, AI adoption cannot be delegated to individual curiosity. It has to be organized as an operating capability, aligned with the firm's own controls, and taught workflow by workflow. That is what turns a licensed tool into a governed way of working - and that is where the value of an AI Academy actually sits.