How we measure what AI changes, and where every number comes from.

Your finance department will ask where every number comes from, and so will your works council (CSE). This page answers them: what we measure, with what data, and within what limits.

If you only have a minute. We measure at the level of the task, not the position. Each task gets four measures, grounded in public research and checked by two people. Every result shows its source and how many tasks it rests on; what we don't know, we don't make up.
Checking payroll anomaliesAI exposurePartial: AI flags, a person decidesSourcesStanford study 2026 · O*NET frameworkAssumptionEditable with your own dataFirst reviewSecond review
Every assessment cites its source and goes through two reviews. Example.

AI doesn't replace jobs.
It changes tasks.

That's why we measure at the level of the task. According to the Anthropic Economic Index (2025), AI is used for at least a quarter of the tasks in 36% of jobs, but for three quarters of them in only 4%: a measure by whole job would see nothing. See what the studies say.

  • Position

    What appears on the org chart

    It's how you track your headcount. But a position bundles too many different things to see what AI changes in it.

    Example: “Accounts payable specialist, Nantes site”

  • Job

    What positions have in common

    What everyone in similar positions does. It's the level of your training plans and your strategic workforce planning (GEPP).

    Example: “Accounts payable specialist”

  • Task

    The specific activity, where it all plays out

    An activity with a beginning and an end, which comes up often in the job. At this level you can see whether AI does the work, helps to do it, or leaves it alone.

    Example: “Matching invoices to purchase orders”

  • Skill

    What you need to know how to do

    When a task changes, what you need to know how to do changes too: it's what your teams will have to learn or pass on. A skill often serves several jobs.

    Example: “Spotting the payment terms in a purchase contract”

JobA job in your companyA task AI doesA task AI prepares, you validateA task done with AIA task that stays humanA new task
How Skillberg reads a job: task by task, each with its destination.

For each task, four measures.

They then add up across the whole job, according to the time each task takes.

  • What AI can do

    How much of this task can AI take on?

    A first opinion from our team, grounded in the measurements of Stanford and Anthropic, then checked by two people.

    5 levels

  • What your teams want

    Do the people who do this task want to hand it over to AI?

    Whether AI can do it doesn't tell you whether it should. Only your teams can say: we ask them, after internal communication.

    from 1 to 5

    asked of your teams
  • What needs to be put in place

    What tools, what rules, what training does your company need?

    From “the tool already exists” to “governance has to be created”. A task that is easy to hand over but heavy to prepare comes later.

    5 levels

  • The pace of change

    Will this task change in the next twelve months?

    Stable, within the year, or within a few months. Assessed, then checked by two people.

    12 months

Payroll officerA typical week, task by task (fictional example)Data entry13 hQuestions9 hChecks8 hLeave6 hNew4 h40 hAI does itWith AIStays humanNew
The measures add up across the whole job, according to the time each task takes. Example.

Five ways to share a task with AI

A Stanford research team (Shao et al., 2026) described five ways of splitting work between a person and AI (labelled H1 to H5 in their study). None is better than another: each task finds its own. This is the scale on which you set, in the tool, how far AI can go.
AI · delegatelevels 1-2Mixed · superviselevel 3Human · keeplevels 4-5
11.9%
AI does it alone
235.6%
AI does it, a person validates
345.2%
You work together
416.3%
The person leads, AI helps
51%
The person remains essential at every step

How to read these percentages: Stanford surveyed 1,500 US workers across 104 jobs. In 45% of these jobs, the most desired way of working is an equal share with AI (level 3). Source: Shao et al., Stanford University, arXiv preprint v3, February 2026, computer-based tasks.

Four sources, which the tool never mixes.

At every step, AI prepares the work and a person from our team checks it before any calculation. For data hosting and processing, see data and trust.

  • Your documents

    Your job descriptions, your skills map if you have one, the number of people per job. No individual assessment.

    Your words, your job titles.

  • Public research

    Academic and public studies (Stanford, Anthropic, University College London, LaborIA from the French Ministry of Labour and Inria), public job frameworks such as the ROME from France Travail, and the EU AI Act.

    Always cited with its date and its limits.

  • The Skillberg model

    Our proprietary map, the result of four years of research: 158,000 skills linked to one another and to jobs, and our skill trees, which show what you need to know before learning the next step. It's what breaks your jobs down into tasks and reveals the pathways between jobs.

    Built and maintained by our team since 2022.

  • The assessments

    The four measures, task by task. Our team sets them, two people check them, and your teams add what they want.

    You can correct them in the tool.

Every result tells you where it comes from.

And how many tasks it rests on.

Here is how it looks in the tool, on three fictional examples. The dots show reliability: three dots when the result rests on research measurements, none when the task has yet to be assessed.

With AI12 / 12 tasks· measured by researchHanded to AI5 / 9 tasks· completed by SkillbergTo be assessed0 / 4 tasks· to be gathered from your teams

“The difference between a list of cities and a road map.”

Martin Vielvoye, founder of Skillberg, on what separates a job framework from a database that links skills to one another.

Sources for this page: Shao et al., Future of Work with AI Agents, Stanford, arXiv preprint v3, February 2026 (WORKBank, 104 jobs, 844 tasks, 1,500 US workers) · Anthropic Economic Index, preprint, February 2025 · JobBench, University of Washington et al., preprint, May 2026. All the studies, with their limits: what the studies say.

What we're asked most often.

Why not simply rely on public studies?

Because they don't know the jobs in your company. The Stanford study, for example, covers 1,500 US workers and computer-based tasks. We link each of your tasks, in your words, to its equivalent in the US public framework of occupations (O*NET), which Stanford and Anthropic worked from. Then we add what no study can know: what your teams want and what needs to be put in place in your company.

Can AI really do what people say it can?

Not always, and that's a good reason to look closely. In May 2026, the best AI agent tested by JobBench (University of Washington et al.) scored only 45.9% on tasks that people would like to delegate. Faced with real cases, with conflicting information, agents still lose 26 to 31 points. That's why every Skillberg result tells you how reliable it is.

Do we have to send you personal data about our staff?

No, none. We start from your job descriptions, your skills map if you have one, and the number of people per job. No one is rated.

Thirty minutes to ask the right questions, and know where to start.

Come with your questions and two or three jobs in mind: purchasing, customer relations or maintenance, for example. We show you the tool in action, their tasks, what AI changes in them and where each number comes from. Then, if you wish, a pilot on a few jobs gives you a solid basis for deciding with your leadership team.

30 minutes on video with Skillberg, no commitment. No personal data required.

Action plan · payroll teamMap the team's tasksChoose tasks to hand to AITrain two people to reviewPresent the plan to the works councilto decide together