Watch the customer story video
Two systems that drift apart
Le Wagon works with two complementary data streams. On one side, Kitt, the proprietary software that centralizes all student-related information: financing arrangements, enrollments, expected payments. On the other, the accounting system, which is the authoritative record at close.
The problem is not having both. The problem is making them converge.
“Having both is good. What is not good is when both live their own parallel lives.”
The complexity stems from the business context itself. A student’s financing arrangement can change during their programme: a corporate funding plan that comes together after enrollment, a public subsidy that evolves mid-course. Each change triggers accounting adjustments. And over time, errors can accumulate: a forgotten credit note, an invoice that was over or under-billed.
At close, everything had to be reconciled. That work represented several days per month.
“To have a true view of our revenue, we need to be able to run these entries every month.”
What blocked classic AI approaches
Le Wagon was not new to AI. The team already used it in day-to-day work. But when it came to processing large volumes of financial data, generalist tools showed their limits.
“I had tested generalist solutions, but I couldn’t audit the data or trace the calculations that led to the final result.”
For a finance team, getting a result is never enough. You need to understand how that result was built, trace the calculations back, verify the source data. That is not a matter of comfort. It is a professional requirement.
“You cannot afford, as a finance professional, to have a black box: to hand over your data and receive an output in return. You have to understand how that output was constructed.”
What made the difference with Phacet
“Phacet is the anti-black box.”
Two things convinced Guillaume. First: traceability. Phacet exposes the data, the calculations, and the intermediate steps. Every decision is visible and auditable. That is what allows the team to validate results with confidence and defend them at close.
The second was a surprise.
“We realized it was not only AI. There was also Python scripting, and that is a strong point.”
In finance, certain processes must be completely deterministic. AI excels at normalizing imperfect data, identifying semantic matches, handling ambiguity in labels. But when precision must be absolute, you need to step away from the probabilistic and return to pure logic. Phacet combines both.
“You get the best of both worlds if you do it right.”
A setup built through iteration
One of the key lessons from the project is that automating complex financial processes does not happen at the click of a button.
“It does not work on the first try. It is not magic. You are not left alone in front of your agent.”
Phacet’s team worked alongside Le Wagon’s Finance team to build the right behavior progressively: understanding exceptions, adjusting business rules, iterating until the agent was reliable enough for production.
“I prefer to invest a bit of time with people who understand both the technology and my business. Once it is properly built, you know it will work every month.”
This approach fits Le Wagon’s own culture, where product thinking also applies to internal tools: identify a specific problem, test whether the solution works, improve.
“I am more of a small-steps person.”
What this changes for the Finance function
Beyond automated reconciliation, Guillaume sees a more structural impact. Finance teams currently spend a significant portion of their time securing data. That is necessary. But it also eats into time that could create more value.
“Finance can be an even stronger business partner. That is what makes finance exciting: you are genuinely at the heart of the company.”
The next use case on the roadmap illustrates this direction: an agent that ingests historical accounting entries to suggest future ones and flag errors before the team discovers them manually. A silent reviewer that scans the accounts before analysis begins.
A tool that does not just produce results. That explains them. And that frees time for what cannot be automated.
