AI redistributes responsibility, it does not remove it
13 July 2026On 30 June, at Le Village by CA in Paris where our French offices are located, we brought together a room of compliance, risk and data leaders for Compl.IA.nce as a Service, the first edition of an invitation-only executive morning.
We opened not with slides but with a live demonstration: an AI agent working inside the KYB and KYC chain, in vivo, in front of the room. The contrast between what AI promises and what actually happens when you test it in front of practitioners set the tone for everything that followed. Three roundtables and a closing conversation took the room from the data layer to the strategic question.
Three convictions stayed with us.
💡 Key takeaways
- AI redistributes responsibility rather than removing it, which makes robust compliance an architecture question rather than a tooling one.
- What works at scale is not a single all-purpose AI but a governed squad of six specialised agents, including a contradiction agent, which cut alert handling from roughly 45 minutes to 8.
- The compliance officer's role is shifting from case-by-case analysis to supervising decision-making systems: arbitrating exceptions and validating changes of logic rather than re-reading every transaction.
- Know Your Counterparty rests on four dimensions, identity, relationship, transaction and event, and AI built on a fragmented foundation reveals the gaps rather than resolving them.
- No one can anticipate the next six months, so compliance must be treated as critical infrastructure with governance you can exercise, including the ability to stop an agent in production without stopping everything else.
1. The human in the loop remains essential
That redistribution is precisely what makes this an architecture question rather than a tooling one: no single product carries that responsibility, a coherent and governable system does.
As one speaker put it, a single AI that does everything is like a chef cooking alone, and what works at scale is a kitchen brigade or what we would call a squad in business. The morning described that squad concretely: six specialised agents, from collection and quality control through analysis and continuous vigilance to alert handling, with a contradiction agent whose only job is to challenge the chain's own decisions. Internal control, the panel argued, should sit inside the chain, not after it. In production, that orchestration moves the analyst from reconstruction to arbitration, with alert handling cut from roughly 45 minutes to 8.
2. The compliance officer's role is shifting, not disappearing
The work moves from case-by-case analysis toward supervising entire decision-making systems. Human in the loop no longer means an analyst re-reading every transaction. This is neither feasible nor what supervisors ask for. It means supervision that is effective rather than exhaustive: the human arbitrates the exceptions and validates the changes of logic, a new threshold, a new rule, a new model version.
One paradox drew real debate. Junior staff comfortable with AI sometimes struggle to assess its output, while senior professionals who watched the technology emerge are often better equipped to challenge it. The lesson: effective AI is judged not by the fact that it is used, but by what supports it, i.e. quality data and governance designed from the outset by the compliance, data and risk teams together. This is where the data roundtable landed too. Know Your Counterparty rests on four dimensions: identity, relationship, transaction and event. Most institutions today master only identity and transaction. AI added on top of a fragmented foundation does not resolve that fragmentation. It reveals it and it turns gaps into confident but false certainties.
3. No one can claim to anticipate the next six months
AI is evolving faster than the frameworks around it. The real question is not to guess its trajectory, but to build systems structured enough to absorb it, whichever direction it takes.
In practice that means treating compliance as critical infrastructure, and it means governance you can actually exercise: an agent holds a revocable licence to act, and the real test of maturity is whether an institution can say stop to a system in production without stopping everything else.
This was only the first edition
Our thanks to the speakers who made the morning what it was: Didier Alleaume (Artimon), François Jaussaud (SAS), Matthieu Boussard (Craft AI), Francis Goter (Clever Age) and Marielle Fernainé (KYC Consulting).
The conviction we carry out of the room is simple. Tomorrow's financial institutions will not be steered by case files. They will be steered by trust architectures, and the advantage will belong to those who start building them now.
Harmoney offers a cutting-edge digital platform that streamlines intricate onboarding and compliance procedures, featuring automated screening functionalities. Interested in discovering more about our innovative solution? Reach out to us for further details or stay in touch via our newsletter ⬇️.