What Commodum's 923 pre-scored roles reveal about real AI exposure in French companies

Breaking 923 roles down into elementary tasks and scoring each one against available AI capability produces, at scale, something no sector study can offer: a picture of exposure that rests on neither job titles nor branch averages. This reference library is the foundation every Commodum diagnostic checks a client’s roles against. But taken as a whole, it also says something about how AI exposure actually distributes across French organizations — and what it says doesn’t always match the prevailing narrative.

923 roles, one method: how the library was built

Every role in the library was processed the same way a client’s roles are: broken down into elementary tasks based on job descriptions and real activity data, then each task scored on its current automation potential. No shortcuts by job title, no sector average injected after the fact. The result is a homogeneous base, built with the same granularity from the first role to the last, spanning support functions as well as specialist roles, from execution to management.

That methodological consistency is what makes it possible to compare a finance role to a customer-relations role on a common basis — the type of task — rather than on a sector classification that lumps very different realities under the same label.

First finding: the job title lies more often than you’d think

The clearest signal coming out of the library is this: two roles carrying the same title, at two different organizations, can show a considerable exposure gap once broken down into tasks. A “payroll administrator” who still manually enters variables into a spreadsheet has almost nothing in common, automation-score-wise, with a “payroll administrator” whose system is already largely integrated and who spends most of their time handling complex cases and disputes.

This finding undercuts a large share of the job-title reasoning that dominates public debate. Ranking roles by name to estimate automation risk is a bit like sorting books by cover color: the ranking exists, but it says nothing about the contents.

Second finding: exposure clusters by task type, not by role

What actually predicts a role’s score isn’t its title, or even its sector — it’s the composition of its elementary tasks. A structured data-entry task behaves similarly whether it’s performed in accounting, HR, or logistics. A high-stakes relational arbitration task holds up comparably whether it belongs to law, complex sales, or management.

The library reveals cross-cutting task families as a result — structured entry and reconciliation, first-pass drafting, structured analysis, high-stakes arbitration, high-emotional-load relationships — whose behavior under AI is far more stable than that of the roles that contain them. Two roles that look very different on paper can carry an almost identical exposure profile because they share the same underlying task architecture dressed up differently.

The role isn’t the unit that matters. The task is — and it cheerfully ignores the boundaries between job titles.

Third finding: middle management is underestimated

Public debate treats middle management as a relatively protected zone, on the grounds that it rests on relational and decision-making skills that are hard to automate. The library seriously qualifies that assumption. A large share of many mid-level management roles is actually taken up by reporting, data consolidation, and decision-preparation tasks that have already been largely worked through — tasks whose automation potential is comparable to that of administrative functions, but which stay invisible in the debate because they’re buried inside a job title with a strategic ring to it.

That doesn’t mean the middle manager disappears: arbitration, team support and decision-making remain low-automation-potential tasks. But the genuinely protected share of the role is often narrower than the title suggests — a gap only a task-by-task breakdown can reveal.

What these patterns don’t replace: your own diagnostic

These three findings are useful because they break widely held intuitions. They don’t, however, excuse you from measuring your own organization specifically. The library provides a benchmark — it tells you what a task of a given type looks like when it has already been observed elsewhere. It doesn’t tell you, on its own, how your roles are actually composed, or in what proportion your employees really perform those tasks.

It’s the difference between a map and a GPS. The library maps the typical terrain of hundreds of task families. The diagnostic applies that map to your actual position — your job descriptions, your volumes, your level of tooling — to produce a score that concerns you specifically, not a general trend you happen to belong to statistically.

What 923 roles say, and what they still don’t

A library of this size changes the nature of the question you can ask. It lets you move past the sterile debate over “will AI replace this job” and ask the question that actually matters: what share of this role’s tasks, in this specific organization, belongs to a task family that’s already largely automatable elsewhere? That’s a question you can answer with data, not with an opinion about the future of work.

What the library will never do, on the other hand, is tell you what’s happening in your own company without being handed your own data to check against it. The number 923 only becomes valuable the day it meets the number of roles your organization actually has.