Increase first, then automate

AI agents are no longer just chatbots. They are becoming digital colleagues capable of performing complex tasks: analyzing data, producing reports, creating presentations, or generating code. But for businesses, the real question isn't "Does it work?" but "How does it work?"

Code bias

The main observation is striking: the agents adopt a massively programmatic approach. Even for visual tasks (slides, logos, mockups), they favor the terminal and code over graphical interfaces. This strategy has two sides: remarkable power for structured and repetitive tasks, but a worrying fragility as soon as it comes to judging aesthetics, interpreting an image, or refining formatting.

The silent risks

The danger isn't limited to typical errors. Two worrying phenomena are emerging: data fabrication (inventing plausible figures when extraction fails) and the inappropriate use of tools (seeking external sources when the internal file is problematic). In a professional context, these practices can distort critical decisions without warning.

Augmentation vs Automation

The raw efficiency of the agents is impressive, but total automation can paradoxically slow down the organization. When the agent produces quickly but poorly, the human operator pays the price in terms of checks, corrections, and loss of trust. The winning approach? "Teaming": dividing the workflow by assigning programmable steps to the agents (data cleaning, structure generation, calculation) and keeping the human operator in critical areas (judgment, consistency, verification, polishing).

The pragmatic roadmap

For businesses, a sound strategy is to map actual workflows, identify mechanical steps, and introduce the agent precisely where it maximizes the effectiveness/risk ratio. Start with targeted augmentation, implement verification, and then progressively automate highly programmable segments.

The question isn't whether AI will replace jobs, but how it reshapes work. When properly integrated, it frees up time on technical infrastructure. When poorly deployed, it creates an illusion of productivity followed by a quality deficit.

To learn more: discover the three-part series: champions, hackathons and MVP, how Generative AI is becoming embedded in workflows et the challenge of industrializing AI.