People often say that AI starts from zero, erasing what existed before. It’s the opposite. For Olivier Rouby and Xavier Davila, respectively a Key Account Manager and the Director of Arondor’s Smart Automation division, a specialist document-processing integrator for more than twenty years, AI does not dismantle the automations already in place: it fills their blind spots.
The diagnosis is precise. Historically, document journeys could be automated up to 80% with proven technologies: OCR, machine learning, RPA. The remaining 20% resisted. Not for lack of effort, but because there was a glass ceiling. That’s where LLMs come into play, not to replace what exists, but to unlock what the current setup could not reach.
This shift in perspective changes everything for CIOs. AI isn’t a restart; it’s an additional layer. And companies that built solid foundations—BPM, IDP, robotic workflows—are precisely those who can reap a measurable ROI the fastest. They have a baseline for comparison, processes already calibrated, a reference framework to assess the model’s real contribution.
The real friction isn’t technological. It’s economic and political. Token consumption is hard to forecast, models change versions, results aren’t always stable. Two calls to the same prompt do not yield the same answer. Hallucinations exist. This isn’t an argument against AI; it’s an argument for not using it alone, without safeguards.
Arondor and Tungsten Automation, a partner software vendor whose AI practice predates the current hype, propose a layered approach: classic technologies continue to handle what they do well, while LLMs intervene where they add value, with a human in the loop to validate ambiguous cases. What practitioners call “human in the loop” isn’t a confession of the model’s weakness: it’s responsible engineering.
On sovereignty, the stance is clear. Some groups have already decided, others still hesitate to entrust their document flows to LLMs governed by the Patriot Act. Alternatives exist—open models, sovereign clouds—and they deserve serious evaluation before committing to a dependence on a single provider.
The advice to hesitant CIOs can be summed up in a few words: don’t go it alone. Not because AI is inaccessible, but because implementation missteps are costly and twenty years of experience with document processes don’t get improvised.
Watch the full interview on video:
