Julien Talbot · activity ergonomist

What AI changes in real work gestures attention decisions agency

I start from real work to decide what AI changes — and what stays under human control.

Email Julien Read the writing

We rolled out an AI tool, but teams aren’t using it. Why?

Usually the tool doesn’t fit the real work loop. We start from the activity, not the tool.

AI saves time, but people seem to think less.

That’s skill atrophy. It can be designed: what you delegate, what stays under human control.

How do we know what actually works before investing?

We observe the real work, test on one activity, document what changes.

Field support See the work →

❯ Task: read and sort unread morning emails.

Analyzing...

▸ 3 supplier invoice emails spotted in inbox

▸ Reading amounts, due dates and attachments

Summarizing the three invoices

Extract amounts and deadlines [done]

Draft a short summary [running]

▸ Reply draft ready — nothing sent

◆ Checkpoints: send, pay, archive

Waiting for your approval [running]

✓ Work ready. Human decision before any action.

Rule: no send, no payment, without your explicit OK

Agents Install an agent →

Stages See stages →

Jun 22, 2026You do not remove a task. You move the real work.

Jun 19, 2026AI Agents Are Operators. They Need Ergonomics.

Jun 4, 2026The Enterprise Agent Problem Is Belief

Jun 2, 2026AI agents in SMEs: start from real work

Jun 2, 2026An AI talk should talk about real work

Jun 2, 2026DUERP, MSDs, psychosocial risks: tie risk to real work

Jun 2, 2026Architects: putting ergonomics into a tender in Réunion

May 29, 2026Raw Traces Are Not Evals

May 19, 2026Tool Use Is Not Task Completion

May 14, 2026The Most Dangerous Agent Failure Is Not Hallucination

May 3, 2026AI Agents Are Not Just Tools. They Are Work Systems

Apr 24, 2026AI Is Not Integrated Into Work. It Reconfigures It.

Apr 21, 2026The Benchmark Lie: Why Grok 4.20 Excels in Benchmarks but Fails in Production

Apr 18, 2026Ergonomics and AI: what the tool displaces in practice

Writing Read writing →

For decision-makers

One loop.
Three steps.

Diagnosis, AI or training — same method: Observe · Decide · Measure.

Budget approved, work unchanged.

Email Julien Read the writing

30 d typical verdict horizon

3 steps — one case file

1 documented verdict

logistics · 12 operators

▸ 4 stations, 2h rotation
▸ shoulder MSDs: 3 / 12
▸ pallet height fixed — back compensation
▸ interview: “I compensate with my back”
▸ AI already there: emails rewritten, no frame

finding: activity is observable

same thread — email integration

▸ delegate   supplier emails
▸ keep       payment, disputes
▸ refuse     generic AI tool

validation: manager before send

D+30 — same thread

▸ email      −40 min/day · errors stable
▸ shoulders  unchanged
▸ attention  less double entry · more dispute watch

verdict: keep email · stop generic tool · fix pallet height

100+: people trained and guided in AI
10+: talks & workshops
6+: years in activity ergonomics, in the field