The signal from the field
Across our engagements in Switzerland, Spain and Romania, we keep seeing the same pattern in AI initiatives. A prototype is built, a demo runs well, and then the project stalls. The failure is rarely in the model. It is that the underlying work was never precise enough to be handed to a machine. Decisions were implicit, exceptions were tribal knowledge, and controls existed only in the head of two or three senior people. An agent placed on top of that work does not remove ambiguity - it industrializes it.
- The bottleneck is almost never model quality.
- It is the absence of a clear operating definition of the work.
- Agents amplify what they are wired into, good or bad.
What a process ready for AI actually looks like
A process is ready for AI when the work can be described in a way another human could reproduce with the same inputs. That means the sequence of activities, the decision rules, the exceptions, the roles who own each decision, and the controls that must be enforced are all explicit. Only then is it responsible to introduce an agent - because the boundary of what it may and may not do is defensible. This is where the discipline of process reconstruction, done from the sources rather than from anecdotes, becomes the real prerequisite to any agent design.
- Activities and their ordering are sourced, not inferred.
- Decision rules are explicit and traceable to a policy or a control.
- Exceptions are named and assigned to a role, not to a person.
- Human-in-the-loop points are designed in, not added later.
The sequence that works
In regulated environments, we systematically apply the same order. First, reconstruct the process as it truly runs today, from documented sources. Second, expose the gaps between the written procedure and the actual practice. Third, redesign the target process with governance and controls in place. Only fourth do we introduce agents, and only on the segments where the work is deterministic enough to be delegated with an audit trail. This is slower than a demo. It is what makes the outcome survive an internal audit.
- Reconstruct - Expose - Redesign - Then automate.
- The redesign step is where value is captured, not the automation step.
- Governance is a design input, not a compliance afterthought.
The board-level implication
For Boards and executive committees, the practical consequence is that AI budgets should be split. Part of the investment must go into making the work itself legible before any agent is trained. If a program starts directly at the model, the Board is not funding transformation - it is funding a demo. The right question to ask an AI initiative is not what the agent can do. It is what the work would look like if a new joiner had to run it tomorrow, with the same quality and the same controls.
FAQ
Does that mean AI agents are premature?
No. It means they belong at a specific point in the sequence. On segments of work that have been redesigned and are governed end-to-end, agents create durable value. On segments that have not, they create hidden risk.
How long does the redesign phase typically take?
In a well-scoped pilot on a single business process, weeks - not months - if the sources are available and a business owner can arbitrate. What takes months is not the redesign; it is the misalignment between IT, business and second-line functions.
What is the first deliverable a Board should ask for?
Not a model. A shared, sourced view of how the target process is supposed to run, with named owners for each decision and each control. Everything else can be built on top of that.
Conclusion
An AI agent is not a strategy. It is a capability added to a process. If the process is not defensible on its own, the agent will not fix it - it will only make its problems faster. The most useful first deliverable in any AI initiative is not the agent. It is a shared view of how the work actually moves, and of what has to change before anything is delegated to a machine.