Lire en français

Rethinking processes around AI is a recommendation that is circulating widely at the moment. What it leaves open is where to start.

At the scale of a workflow, two paths are on offer today.

The first starts from the tool. We look at what agents can do, rebuild the workflow around their capabilities, then adapt the human steps to fit.
Redistributing the steps condenses checks that were previously spread out into a few control points. Bottlenecks appear where work once ran freely. The pressure shifts to the time given to judgment. And by the time the workflow has been rolled out to other teams, the tool’s capabilities have already moved on.

The second starts from the human steps we consider irreducible. We set them first, then develop agents that meet that specification. This path also calls for working out in finer detail the outputs judgment needs, rather than choosing a tool that already provides them. The workflow moves at a human pace. We know from the outset that development will take longer, and that in the meantime we set aside the most advanced capabilities currently available.

Neither path removes the human from the loop.

The difference lies in what determines where human judgment enters the process. In the first, it is the tool’s capabilities. In the second, it is the steps on which we do not compromise.

The line itself moves. What the tool cannot do today, it may do tomorrow. What we consider irreducible today may not remain so.

In your view, what implicit assumption guides the choice between these two paths?