Rgroup’s Pierre-Mathieu Roche explains how Angèle connects AI to existing ERP, CRM and M365 systems, securely and step by step, without a full rebuild.
Angèle is Rgroup’s AI integration layer, linking business assistants and agents to a company’s real data across ERP, CRM, document systems and Microsoft 365, with strict access governance, model-agnostic design, and hosting in Luxembourg or on-premise. Pierre-Mathieu Roche, independent board member and strategy consultant at Rgroup, spoke with Silicon Luxembourg about how the offering works in practice, from a first pilot to full deployment.
“Our design rule is simple: one agent, one model, one data source, clear instructions.”
Pierre-Mathieu Roche, independent board member and strategy consultant at Rgroup
How does Angèle integrate with existing systems (ERP, CRM, M365) without requiring a full rebuild?
Angèle’s starting principle is that intelligence comes to the data, not the other way around. Nobody is asked to migrate their ERP, build a data warehouse, or redesign their information system. Angèle is an orchestration layer that sits on top of the systems already in place and connects to them through their existing interfaces. The data itself never moves.
In practice, everything isn’t connected at once. Our design rule is simple: one agent, one model, one data source, clear instructions. So we start with a single system and use case — documents in M365, or billing data in the ERP — first in read-only mode, then moving to action with human validation. From there, we compose agent by agent. That’s what allows an SME to move forward without tying up its IT department for months.
A second point that matters for the durability of the investment: Angèle is model-agnostic. When a new model comes out, it plugs in without disruption. The company stays technologically current without re-engineering its systems or depending on a single vendor. The integration work doesn’t have to be redone with every technology cycle.
What guarantees exist around access governance and data security, particularly in regulated sectors?
Three guarantees form the very architecture of the product.
The same rules apply to humans and to agents. Access is segmented user by user, aligned exactly with the rights and permissions already in place internally. An agent never accesses more than the user it serves does. This is the point most often overlooked when a company deploys AI in a hurry: an assistant ends up with visibility into data that half the staff isn’t authorised to see.
Routing is arbitrated by data sensitivity. Angèle decides, flow by flow, where inference runs. For sensitive data, the model runs locally, on controlled infrastructure hosted in Luxembourg: nothing leaves, and the question of extraterritorial exposure doesn’t arise. For non-sensitive data, the company can draw on the performance of market-leading models. This isn’t a default setting imposed on the client — it’s defined, justified and logged.
Auditability. Every use, every access, every flow is traceable and controllable. For a regulated player — a PSF, or a company in finance, healthcare or law — this is what turns AI usage into something demonstrable to a regulator or auditor. With the AI Act, it’s better to be compliant from the start than to catch up afterwards.
“[…] it’s adoption that drives return on investment.”
Pierre-Mathieu Roche, independent board member and strategy consultant at Rgroup
Two further operating principles apply. Responses are grounded in the company’s real data rather than generated freely, and can be traced back to their source. And every critical action remains subject to human validation.
Finally, deployment is the client’s choice: SaaS in a sovereign Luxembourg environment, or on-premise on the client’s own infrastructure, through our partnership with Dell. Either way, control over data location, flows and access is preserved.
Can you give a concrete example of a Luxembourg company that has followed this step-by-step approach?
We have a deployment under way with a Luxembourg client, but our contract requires confidentiality, so I won’t name them — which is fairly consistent with what we sell, in fact. What I can describe is the approach itself, which is the same for every client.
It always starts with a short, no-obligation diagnostic: we look at how AI is already being used in the company — there’s always some use, often outside any formal framework — the state of the data and the information system, and then we rank use cases by impact and feasibility. The output is a prioritised roadmap, not a promise.
Next comes scoping: target architecture, choice of models, the pilot’s boundaries, and success criteria defined together with the client. Then a single agent goes into production, on the most profitable use case. Just one. That’s where the project’s credibility is won or lost — we measure real results before considering anything further.
Industrialisation only comes after that: scaling up, monitoring, training the teams. This is also the stage where Luxembourg’s public support schemes — Fit4Digital, the SME Package AI — come into play, depending on the project’s eligibility.
This sequencing isn’t commercial caution. It’s what avoids the four classic failures of an AI project: tackling the wrong problem, choosing the wrong tool, stopping at the prototype stage, and neglecting adoption. Because it’s adoption that drives return on investment, not the technology itself.
