An advisor learns the job in front of a client. That is also the worst place to learn it: the appointment happens once, the client is real, and the feedback, if any arrives at all, arrives weeks later in a figure.

With Sopra Steria, we built Mon Coach IA for the École supérieure de la banque: a tool where that conversation can be had as often as it takes, against a client who is simulated but does not behave like a script.

A client who reacts to how you ask

The advisor is put in front of a virtual client with a profile of their own, and the appointment unfolds from there. What the client says next depends on what was asked and how it was approached, so the same scenario run twice does not play out the same way. Feedback arrives during the session, not only after it, and a detailed review follows at the end.

Progress is tracked across sessions on a dashboard, so a competency can be seen moving rather than being certified once and assumed to hold.

Four kinds of expertise, four agents

Underneath, it is a multi-agent architecture, and the agents are split along the lines the problem actually has:

Domain, which knows what a banking appointment contains
Pedagogy, which knows when to intervene and when to let someone struggle
Compliance, because this is a regulated sector and an advisor rehearsing the wrong reflex is worse than no rehearsal at all
Performance measurement, which turns a conversation into something that can be scored and followed over time

Splitting them is what makes the thing correctable. A pedagogical choice can be argued with and changed without touching what the system knows about banking, and the compliance constraints stay legible instead of being dissolved into one long prompt.

From 45 learners to more than a thousand

The tool went into testing in June 2026 with 45 learners across two programmes, and launched in September. More than a thousand learners are expected on it by the end of the year: students, apprentices, advisors in continuing education, and bank networks.

It ships three ways: built into a curriculum, deployed as it stands, or adapted with scenarios written for a given network.

Why this one was interesting to build

Simulation is an old idea in training. What has changed is that the simulated party can now hold a conversation that is not on rails, which is exactly the part that used to make role play expensive: it needed another person.

What it does not change is the requirement to be right. A training tool that rehearses a non-compliant answer teaches a non-compliant answer, at scale, to a thousand people. That is why the compliance agent is a component of its own and not a paragraph of instructions, and why measurement was in the design from the start rather than added once someone asked whether it worked.