The true cost of automation: monitoring and verification

AI agents promise to revolutionize office work, but the reality is more nuanced than simply replacing humans with machines. Understanding how agents and humans work differently is essential to capturing the value of AI without sacrificing reliability.

Two radically different ways of working

Current agents adopt a largely "programmatic" approach: they transform everything into code. Where a human opens Excel and adjusts a column, the agent writes a Python script. Where a designer manipulates a canvas, the agent generates HTML. This preference reveals a deep bias: agents excel at manipulating symbolic representations but struggle with visual interfaces and incremental feedback.

Humans, on the other hand, work in a "UI-centric" way: interactive interfaces, progressive control, constant intermediate checks. It is these micro-decisions that transform an acceptable deliverable into a result ready to be sent.

The silent abuses: the real risk

The problem is not the visible failure, but the credible fabricationAn agent unable to correctly extract data from an invoice can invent plausible values ​​and produce a professional-looking spreadsheet… but it's wrong. In finance, HR, or compliance, this ability to generate seemingly consistent results is more dangerous than a simple glitch.

Another problem is the misuse of tools. When an employee cannot use an internal document, they resort to similar public sources and continue without raising any concerns. In a company, this raises questions about governance and traceability: what inputs are we really basing this result on?

Structured collaboration rather than substitution

The solution is neither "automate everything" nor "delegate nothing," but to build a step-by-step collaborationAgents often encounter very localized problems: navigating folders, extracting visual content, or understanding ambiguous instructions. Once this limitation is overcome by a human, the agent becomes highly efficient again in its computational functions.

The granularity of delegation is crucial: delegate at the workflow step level, not at the individual click level. In one experiment, simply assigning file navigation and collection to a human allowed the agent to perform the analysis quickly, resulting in a significant overall efficiency gain.

Classify tasks by programmability

Tasks can be easily programmed : cleaning data, calculating indicators, generating structured reports. The agents provide a net gain, with the main risk being the validation of sources.

Semi-programmable tasks The agent succeeds, but only through workarounds (writing in Markdown then converting to docx, coding a website instead of using Figma). The need here: more APIs and bridges between the UI world and the code world.

Tasks that are difficult to program Visual perception, aesthetic judgment, scanned documents: the human element remains the benchmark. This is precisely where agents produce results when they are asked for complete autonomy.

Concrete recommendations

For product builders Measure the quality of the workflow, not just the final success rate. Future competitive advantage will come from monitoring systems, checkpoints, and automated validation.

For companies Start by mapping actual workflows, then inject AI as an augmentation tool before aiming for automation. You'll achieve quick wins without turning your employees into burnt-out controllers.

On the costs The lower cost per task is appealing, but a silent error can cost more than the savings achieved. The true ROI includes the cost of verification, correction, and risk.

The future: a new division of labor

The near future looks less like a company being replaced by agents than a company being reconfigured. Agents will handle the deterministic, repetitive, and calculative steps. Humans will retain the advantage in ambiguity, judgment, communication, and the need for proof.

The best organizations will be those that can orchestrate this complementarity by structuring workflows, choosing the right levels of delegation, and transforming supervision into a system rather than a mental burden.

The field of innovation isn't limited to "creating an agent that can do everything": it lies in workflow engineering, traceability, business tools, and the creation of safeguards against manufacturing. In other words, it encompasses everything that transforms an impressive demo into a reliable and deployable product.

To learn more: read our article on Augmented managers and reinvented teams and our analysis of the workslop.