Hello,

Jobs are moving, proof is slower

This week, four stories share one thread: the pressing need for human intervention when AI is used autonomously. An AI model sent a false tip to the Philadelphia police, discovered two months later. A union filed a social alarm over AI deployment at RATP, the Paris transit group. Six French companies out of ten have no written rule for their AI use. And two assistants get four figures out of ten wrong when asked to read the accounts of CAC 40 companies.

This edition also brings something new: a dedicated section on how jobs are changing and which new ones are appearing, born from a request received after #006. I started from the documents themselves, French and European first, and separated what is measured from what is only announced. You will find fifteen resources with their dates and limits, several of them to download.

As always: the source is visible, the opinion is my own, and the tutorial needs no code.

Jérôme Denis, JDENIS Consulting, Toulouse

The essentials in 1 minute

• An AI agent sent a false tip to the police, and it took two months to notice.
• Permanent-contract hiring of 21-29 year olds is about 28% below its November 2022 level in the jobs most exposed to AI, against 8% in the least exposed (Banque de France). How much of this is due to AI is still disputed.
• Six French companies out of ten have no written rule for AI (Coface barometer, self-reported): a charter remains the first step.

Contents

Four signals from the week

Agents · Signal 1 of 4

An Anthropic model sends a false tip to the police, and nobody notices for two months

On July 18, an Anthropic AI model submitted false information about an unsolved murder to the Philadelphia police tip line. According to Anthropic, it was running a test on randomly selected websites at the time. The company only found out on September 28 and told the police on October 8; the message had been filed as spam. Anthropic has also cut live internet access for all its internal evaluations, after finding that its agents were exploiting software flaws, including on US government sites, to reach their goal.

techcrunch.com · October 9, 2026 · techcrunch.com · October 9

Social dialogue · Signal 2 of 4

At RATP, a social alarm over consulting the works council

On September 30, the CFE-CGC union at the RATP group filed a social alarm for failure to respect social dialogue over AI deployment and for not informing and consulting employee bodies about new technologies. The letter names no tool. Management had five working days to respond, until October 6, and the union had published no response by that date. The framework is Article L. 2312-8 of the French Labour Code: the works council (CSE) must be informed and consulted before new technologies are introduced. A February 14, 2025 order of the Nanterre court held that deploying AI tools, even as a pilot, is already subject to that duty.

actuia.com · October 5, 2026 · frenchweb.fr · May 12, 2025 (Nanterre order)

Governance · Signal 3 of 4

Six French companies out of ten have no written rule for AI

According to the 2026 Data & AI Barometer from Coface and Les Echos Etudes, published October 5, 60% of French companies have no formal framework for their AI use: 35% have no arrangement or plan at all, and another quarter are thinking about it without written rules. Only 17% have defined, formalized and communicated a policy to their staff. The survey rests on the statements of 900 decision-makers, with no external verification, and the full study is only available after registration: I cite ActuIA's report of it here.

actuia.com · October 6, 2026

Finance · Signal 4 of 4

A word to the Sunday traders: Four figures out of ten are wrong when ChatGPT and Gemini read CAC 40 accounts

Reputation Age, a communications agency, and Bubbling, a platform that tracks brand presence in AI assistants, questioned ChatGPT and Gemini about the half-year accounts of 39 CAC 40 companies between August 25 and 28: 8,671 conversations, 34,684 answers. Result published October 2: 40% of the figures returned are incorrect. Published growth is right 42% of the time, net income 67%. In 24% of the answers on published growth, the assistant quotes an official company figure, but for a different indicator: "AI finds the right document but picks the wrong number," the study sums up.

actuia.com · October 5, 2026 (study of October 2)

Law & governance, in brief

The AI Act after the omnibus

Regulation (EU) 2026/1744 of July 8, 2026, published in the Official Journal on July 24, moves the requirements for Annex III high-risk systems (including recruitment and personnel management) from August 2, 2026 to December 2, 2027, and those for Annex I to August 2, 2028. Article 4 on AI literacy of staff is kept in a softened form: take measures, with no specific level to guarantee for each person.

eur-lex.europa.eu, July 24, 2026

OpenAI will watermark ChatGPT text in the EU

On October 5, OpenAI published the technical report for textGrain, a statistical watermark applied to ChatGPT and Codex text to comply with Article 50 of the AI Act. Announced detection: about 95% on a 400-token passage, 17% if a quarter of the words are replaced. The detector will first be limited to researchers, and the watermark stays optional for the API.

actuia.com, October 6

California: no discipline or firing decided by AI alone

The governor signed bill SB 947 on September 30. From July 1, 2027, an employer may no longer rely solely on an automated decision system: a human must corroborate it with other evidence (evaluations, the employee file), and the employee receives written notice.

crowell.com, October 6

United Kingdom: the ICO reports on ten AI developers

On October 8, the data protection regulator published the outcome of its supervision of foundation models: ten developers have changed, or promised to change, their notices and legal analyses, with sensitive data still a sticking point. A six-week consultation on agentic AI is open until November 20.

actuia.com, October 9

United States: a "Super Intelligence Force" with no official text

Announced on October 4 in a presidential post on Truth Social, it puts four officials in charge of coordinating federal AI policy. The next day, no executive order or charter had been published: its legal basis, budget and powers remain unknown.

actuia.com, October 5

OpenAI: the three dismissed researchers push back

In an open letter, they deny having broken the rules or acted outside their duties. OpenAI cites a "pattern of misconduct" in the handling of research information, without saying which rules, and denies any retaliation for safety warnings.

techcrunch.com, October 8

AI on the job

Public administration

The state's AI assistant, evaluated on ten thousand licences

From December 2025 to June 2026, the French state trialled an "Assistant IA" on 10,000 licences. The evaluation by the public transformation directorate, run with researchers from Inria, INSA Rennes and CNRS, found that 66% of the civil servants surveyed say they save time, 19% see no impact, and 6% lose time. Among those who save time, 45% put it at under an hour a week and 51% at several hours. In workshop tests with about 30 civil servants: 21 minutes versus 19 to answer an email chain, 33 versus 28 to summarize an article. The authors write that these results cannot be generalized.

modernisation.gouv.fr · June 2026

Banking

Barclays publishes usage volumes, not gains

The British bank says more than 16,000 employees have used an internal knowledge assistant since 2025, which has handled over a million searches to help serve its 20 million UK retail customers. A triage system classifies and routes about 120,000 emails a day in its Global Markets division. It expects 50% of its developers to be on Claude Code by the end of 2026. These are figures declared by the bank and published on the vendor's site: they are volumes, and no gain in time or money is measured.

According to anthropic.com · October 1, 2026

Telecoms

Deutsche Telekom reports 2.6 million calls handled by a chatbot

In its October 5 release, the operator says its "Frag Magenta" chatbot handled about 2.6 million customer service calls in the first half of 2026. In the United States, call volume fell 55% and AI agents handle 40% of contacts. On the network, handling certain events went from several hours to about a minute. For 2027, outside the United States, the company expects about 1.1 billion euros in gross savings from AI and automation, compared with 2023. Figures declared by the company, with no external audit.

telekom.com · October 5, 2026

Culture

Qobuz labels AI-generated music

The French music streaming platform announced on September 24 a label that flags, at album level, content identified as AI-generated, using a tool developed in-house. According to Qobuz, 60% of streams on tracks identified as AI are judged fraudulent by its anti-fraud systems and excluded from royalty payouts; those tracks account for 0.38% of streams. Figures declared, detection method not detailed. The platform says music discovery must remain guided by human choices.

musicbusinessworldwide.com · September 28, 2026

Education

The European Commission finds the effect on learning still unclear

In its "Friend or foe?" report, published September 22, the Commission writes that generative AI is "slowly but steadily" becoming a mainstream tool in teaching, but that its effect on long-term skills "remains unclear". The benefit depends on how it is used, and teachers need substantial initial and ongoing training. According to the second report published the same day, nearly two thirds of European education systems have an AI strategy or framework.

digital-strategy.ec.europa.eu · September 22, 2026

Healthcare

In intensive care, models that update their judgment poorly

A study posted October 2 on arXiv (preprint, not peer reviewed) tests several language models on intensive-care patient trajectories. When the data change, the models react more strongly to a respiratory deterioration than to an equivalent improvement, and their estimate depends on their own earlier judgment: raising a prior risk from 10% to 90% shifts the estimate by 26.2 points on identical evidence. Rewording the prompt does not fix it. This is a research result, not a tool in service.

arxiv.org · October 2, 2026

Notaries

A framework and a loan, no results yet

The French national council of notaries announced on July 9 that it chose Mistral AI and Scaleway for its AI strategy, stressing that responsibility for the deed stays with the notary. The Banque des Territoires offers a "NumerIA" loan at 1.95% (preferential rate, limited period), up to 80,000 euros per office over five years, which also covers staff training. No measured result has been published so far.

banquedesterritoires.fr · September 30, 2026

What you asked me

After edition #006, a reader who supports the HR transformation of a large organization asked me for documentation on how jobs are evolving and which new ones are emerging. Here is the selection. I started from the documents themselves rather than press summaries, French and European first, 2025 and 2026 where possible.

Each resource carries a label. OBSERVED means measured in real data, ANNOUNCED means a projection, a scenario or a statement of intent. The two do not add up, and I set aside figures from commercial consultancies that I could not trace to a verifiable source.

What the sources say, in short

Transformed rather than replaced. The ILO writes that most jobs include tasks that require human judgment, and that job transformation is the most likely effect of generative AI. The French Treasury cannot yet measure the total effect on employment.

The clearest signal. It concerns young people entering the most exposed jobs, in France as in the United States. How much of it is due to AI is disputed: the OECD judges it limited, Stanford sees an early indicator.

What is emerging. The best-documented novelties are new tasks inside existing jobs (oversight, integration, compliance). As for job titles, the 2026 ROME, France's official job register, recognizes "AI engineer" and "prompt engineer", but neither "chief AI officer" nor "AI trainer".

What rises, what recedes. Rising: mastery of AI tools, secure use, data governance, judgment and validation. Receding: codified, repetitive tasks. The World Economic Forum lists data entry clerks and bank tellers among the fastest-declining jobs.

OBSERVEDInsee (French national statistics office) · July 21, 2026 · 4 pages, PDF in French

Insee Première no. 2120: information and communication technologies in companies in 2025

18% of companies based in France say they use at least one AI technology in 2025, 8 points more than in 2024 (20% on average across the European Union). The rate reaches 58% for companies with 250 employees or more. Among non-users, 71% see no use for it and 54% cite a lack of expertise.

Who it is for: any leadership team that wants an official order of magnitude, to cite in place of vendor surveys.

Limit: companies with 10 employees or more, use declared by the company. Insee points out that this does not mean 59% of employees use AI: user companies account for 59% of employment in the scope.

Insee page

OBSERVEDBanque de France · October 8, 2026 · blog post in French

Banque de France, blog post no. 463: generative AI, the first labour-market changes for young people

In March 2026, permanent-contract hires of 21-29 year olds in the fifth of jobs most exposed to AI are about 28% below their November 2022 level, against 8% in the least exposed fifth. The gap becomes almost nil from age 50. Total employment at companies that use AI intensively does not clearly differ from others: the first changes show up in entries and exits more than in headcount.

Who it is for: HR directors, recruitment and apprenticeship managers, employment counsellors.

Limit: the exposure index measures what AI could technically do (assessed by a language model from the ROME), not actual use. The authors cite foreign studies with contrasting results and urge against looking for a single average effect.

Banque de France post

OBSERVEDDG Trésor (French Treasury) · June 30, 2026 · 12 pages, PDF in French

Trésor-Éco no. 391: Artificial intelligence, what effects on employment?

Empirical studies "cannot determine the total effect of AI on employment", for lack of hindsight. Estimates of exposed jobs range from 5% to 60%, and only part would be substitutable. Layoffs attributed to AI are concentrated in finance and IT, even though AI is sometimes a pretext. Unemployment among 15-24 year olds rose from 19.1% to 21.1% between the first quarter of 2025 and the first quarter of 2026, with the economic cycle as another possible explanation.

Who it is for: to frame a leadership meeting or a works council session with a sober public source.

Limit: a literature review, mostly Anglo-Saxon. Exposure estimates vary widely with the index chosen.

DG Trésor page

OBSERVED + SELF-REPORTEDWorld Economic Forum, in collaboration with PwC · June 2026 · 42 pages, PDF in English

World Economic Forum with PwC: Artificial Intelligence and the Future of Entry-Level Work

The resource to keep for first jobs and the integration of young people. Key figures, with report page numbers: 37% of young workers (15-24) worldwide are employed in occupations with medium to high exposure to AI-driven task change, 63% in Europe and 69% in North America (pp. 4 and 6). The slowdown in entry-level hiring is real, but postings were already falling nearly a year before ChatGPT, and several leaders interviewed say AI is often "the visible explanation" for hiring that slows for other reasons (pp. 11-12). 68% of entry-level workers say AI has raised their productivity, but 45% also say they work more hours (p. 17). Only 16% of organizations report having fully redesigned roles and processes to integrate AI (p. 19), and 28% of entry-level workers believe half or fewer of their current skills will still be relevant in three years (p. 5).

What it changes for first jobs: if AI takes over routine tasks, the learn-by-doing model weakens: the report asks organizations to choose what they automate and what they keep in order to train people (p. 19). It also notes that 84% of baby boomers have not used an AI agent in the past year, against 57% of Gen Z: access to the tools is not equal (p. 14).

For HR and employment support: make entry-level hiring an explicit part of workforce planning with clear targets, redesign entry roles rather than layering AI on top, recognize applied experience alongside degrees, and do not assume tool access is the same across age, gender or social background (pp. 14, 19 and 36).

Limit: job-posting data are mostly American and British, the surveys are PwC's (over 9,000 entry-level workers in 48 countries) and self-reported. No detailed section on France.

World Economic Forum page · PDF

OBSERVED, BUT DISPUTEDStanford Digital Economy Lab · August 2026 version; OECD, Employment Outlook 2026 · July 2026

Young people and AI: two readings that conflict (Stanford and OECD)

Stanford (Brynjolfsson, Chandar and Chen), using US payroll data through June 2026: employment of 22-25 year olds in exposed jobs is 19% below where it would be had it followed that of less exposed peers, against 15% with July 2025 data, with no evidence of economy-wide job displacement. The authors describe early, descriptive indicators, not causal estimates. The OECD (section 1.3.1) judges that the role of large language models in young people's difficulties entering the labour market "appears to be limited", with mixed evidence focused mainly on the United States; it concludes that generative AI may have begun to automate some jobs, but that the data show little evidence of it so far.

Who it is for: decision-makers on junior hiring, to avoid overreacting in either direction.

Limit: US data. Read with the Banque de France's caution: a correlation, not yet a cause.

Stanford, Canaries in the Coal Mine · OECD, Employment Outlook 2026 (PDF)

ANNOUNCEDInternational Labour Organization · May 2025 · 76 pages, PDF in English

ILO and NASK: Generative AI and Jobs, a Refined Global Index of Occupational Exposure (Working Paper 140)

One in four jobs worldwide (24%) is in an occupation exposed to generative AI, and 3.3% of global employment falls in the highest exposure category: 4.7% of female employment against 2.4% of male employment, and 9.6% against 3.5% in high-income countries. The annex ranks HR occupations: personnel clerks (ISCO 4416) in the highest category, personnel and careers professionals (2423) in an intermediate one, human resource managers (1212) in minimal exposure.

Who it is for: to classify an organization's positions by occupational family and look at the different effect on female employment.

Limit: this is potential exposure, not impact. The authors write that job transformation is the most likely effect.

ILO page

ANNOUNCEDWorld Economic Forum · January 2025 · 290 pages, PDF in English

World Economic Forum: The Future of Jobs Report 2025

Employers surveyed project the creation of 170 million jobs and the displacement of 92 million by 2030, a net gain of 78 million (7% of current employment). 39% of existing skills will be transformed or become outdated, against 44% in the 2023 edition and 57% in 2020. Data entry clerks and bank tellers are among the fastest-declining jobs.

Who it is for: strategic workforce planning, foresight workshops.

Limit: a self-reported employer survey projected to 2030, mixing AI, demographics and the green transition. The job figures are an extrapolation of statements, not a measurement.

Report (PDF)

OBSERVED (REGISTER)France Travail (French public employment service) · published June 15, 2026 · open data, in French

ROME 4.0, version 61 ("26M06")

The official job register holds about 1,900 job sheets, 351 of them new in this version. Sheet M1889 "Artificial Intelligence Engineer" groups titles such as "Prompt engineer" and "AI Product Manager"; sheet M1873 "Embedded AI specialist" also exists. The register contains skills such as "Use AI", "Automate tasks with an AI agent" and "Secure the use of AI in the company". I searched for, and did not find, "chief AI officer", "AI trainer" or annotator, "data steward" and "AI referent" among the job titles.

Who it is for: career counsellors, builders of skills frameworks, HR directors writing job descriptions.

Limit: the ROME says a title exists, not that it is widely hired. I could not date the creation of sheet M1889.

ROME 4.0 on data.gouv.fr

OBSERVEDNBER · May 2025, revised March 2026 · 131 pages, PDF in English

Humlum and Vestergaard: Still Waters, Rapid Currents (NBER Working Paper 33777)

A Danish study linking adoption surveys to administrative records, for eleven exposed occupations including HR professionals. It finds no effect on earnings or recorded hours, with a margin that rules out any effect above 2% two years after the launch of ChatGPT. What moves is the organization of work: new tasks appear around content generation, oversight of AI output and its integration. Even without employer initiatives, chatbots create a new workload for 12% of their users; 18% in workplaces that encourage their use.

Who it is for: to redesign job descriptions: oversight tasks get added first, before positions are cut.

Limit: Denmark, data through 2024, a working paper not peer reviewed.

NBER page

SELF-REPORTEDOPIIEC, French digital, engineering, consulting and events sector observatory · April 28, 2025 · 7 pages, PDF in French

OPIIEC: skills, jobs and training needs in AI ("key points" summary)

28% of companies in the sector have created AI-dedicated positions (AI engineer, AI project manager), 21% have increased headcount in some jobs (developer, data analyst) and 7% have cut it in others (designer, support functions such as secretarial work). Companies expect nearly 45,000 jobs to be created in three years, 3% of headcount, and 47% of the sector's jobs have a potential for transformation or automation.

Who it is for: HR directors in services and consulting, sector observatories: an example of a sector-level reading, with jobs created and jobs declining.

Limit: a self-reported survey of companies in the sector that builds AI, so not representative. The 45,000 jobs are a projection.

Summary (PDF)

OBSERVED (SELF-REPORTED)Apec (French association for managers' employment) · May 2026 · 32 pages, PDF in French

Apec: Managers and AI, 2026 edition

A March 2026 survey of 2,000 managers and 1,000 companies. Half of managers use AI at least once a week, 15 points more than in 2025; 29% say they have been trained (24% in 2025). 27% of large companies have a usage charter (17% in 2025). Two companies in ten already value AI proficiency when selecting candidates, and 53% of large companies plan to give it more weight in future manager hiring. 31% of managers looking for a job have used AI tools, but 71% believe they will not fully replace the advice of professionals.

Who it is for: HR directors, managers, training and employment support leads.

Limit: private-sector managers, self-reported data: do not extrapolate to all employees.

Apec page

FRAMEWORKEuropean Union · July 8, 2026, published July 24 · official text; Nanterre order of February 14, 2025

Regulation (EU) 2026/1744 (AI omnibus) and consulting the works council

For recruitment and personnel management tools (Annex III of the AI Act), the requirements for high-risk systems apply from December 2, 2027 instead of August 2, 2026. Article 4, on AI literacy of staff, stays in force in a softened form. Under French law, Article L. 2312-8 of the Labour Code requires informing and consulting the works council before new technologies are introduced; the Nanterre order held that a pilot deployment is already covered.

Who it is for: HR directors, legal departments, works council secretaries.

Limit: to be validated by a lawyer: national transposition and Commission guidelines are still to come.

Regulation on EUR-Lex · Frenchweb on the Nanterre order

SELF-REPORTEDCedefop, European Union agency · 2025 · 29 pages, PDF in English

Cedefop: Skills empower workers in the AI revolution (policy brief)

A survey run in 2024 in 11 member states: 28% of adult workers say an AI tool is used in their workplace. About 15% fear losing their job to AI within five years, 40 to 60% have a poor understanding of what AI is, and six workers in ten face some form of task transformation.

Who it is for: to build a skills development plan and a social dialogue grounded in a European source.

Limit: self-reported, eleven countries, a survey that predates the acceleration of 2025 and 2026.

Cedefop brief (PDF)

To download or read first

World Economic Forum and PwC, Entry-Level Work (42 pp.): for first jobs, young people's integration and the design of entry roles.

Trésor-Éco no. 391 (12 pp.): to frame a leadership team or a works council with what is known, and what is not.

Banque de France, post no. 463: the French figures on young people's entry into work, in a few pages.

Apec, Managers and AI 2026 (32 pp.): real usage, charters, training and effects on hiring.

ROME 4.0 version 61: to check that an AI-related job or skill exists in the official register.

The commitment on live performance still stands: find a documented case where AI changes production or the relationship with audiences. This week, culture came from online music (Qobuz, section 02) and not from the stage. If you work in that field, your account matters more than my own research.

Tutorial: build your first agent, in Claude Cowork and in ChatGPT Business

This week's tutorial

The example: a monitoring agent that, every Monday, gathers the new publications on one job in your organization (Insee, Dares, France Travail, ILO) and delivers a one-page note, with a link and a date for every fact. No personal data: it is a good first ground. Interface labels are quoted as they most often appear.

Common step: write the agent's brief (10 minutes)

This is the part that matters. An agent is just a precise instruction, reused.

1. The mission, in one sentence. "Every Monday, summarize the week's publications on the job of payroll administrator."

2. Approved sources. A closed list of sites. Everything else is excluded.

3. The output format. Five points at most, each with its source, its date and one sentence: "what this changes for us."

4. What it never does. Make up a figure, cite without a link, send anything without your approval.

5. A test case. A note you have already written by hand, to compare against.

In Claude Cowork (Claude Desktop app)

1. Open a task. In Claude Desktop, Cowork tab, click "+ New task". Paste your brief and ask: "Turn this brief into a reusable skill." Claude asks a few questions, then produces the skill (an instruction file). Save it and check that it is enabled in the settings, under Skills.

2. Test on the case already handled. Describe the task in a new conversation: Claude should say it is using the skill. Compare line by line with your note and ask it to fix the skill wherever it gets things wrong.

3. Schedule it. Type /schedule in the task, or go through "Scheduled" in the sidebar, then "+ New task": a name, the instruction (type "/" to call the skill), the "weekly" frequency, and if needed the folder where the note should be saved.

4. The limit to know. A scheduled task only runs if the computer is on and the app is open. A missed run is caught up when the computer wakes, and flagged.

In ChatGPT Business or Enterprise (workspace agents)

1. Create the agent. In the sidebar, "Agents", then "Create". Paste your brief as the instruction (or "Start blank"), review the proposed plan and click "Build this agent".

2. Set up the builder. Check the name, description and instructions. Under "Tools", "+ Add tool": web search, then Google Drive or SharePoint if the note should be saved there. Leave write actions on "Always ask".

3. Test. "Preview" at the top right, with your test case. Adjust the instructions, then "Create".

4. Schedule and share. "Add schedule": weekly, on Monday. Under "Channels", keep "Private to me" during trials; "Share" then lets you invite colleagues as readers ("Can chat") or editors ("Can edit").

5. On the administrator side. In ChatGPT Enterprise, workspace agents are off by default: the administrator decides who can use them, create them and publish them.

Before opening the agent to a team

• Measure for three weeks. Review time, errors found, sources missed. The real gain is the time saved minus the review time.

• Watch shared connections. In ChatGPT, an "Agent-owned account" connection lets other people act with your credentials. Prefer a service account, or each user's own connection.

• Collective use takes preparation. As soon as the agent is offered to a whole team, raise the question of informing and consulting the works council before you start, not after.

Sources: Claude, scheduled tasks in Cowork · Claude, create a skill through conversation · OpenAI, workspace agents for Business and Enterprise

This tutorial prepares the discussion. It does not replace legal advice or consultation of employee representatives.

The common thread

If I had to sum up these seven days in one sentence: jobs change through tasks first, and proof about employment arrives more slowly than the headlines.

On the usage side, the most solid sources of the week are public and cautious. The evaluation of the state's AI assistant shows time savings for two thirds of respondents, but under an hour a week for 45% of those who save any. The European Commission finds the effect on learning still unclear. The Treasury writes that the total effect on employment cannot yet be measured. The more spectacular figures come from company press releases, which publish usage volumes rather than results.

On the governance side, the gap is the same: an agent left for two months with nobody reading its traces, employees asking to be consulted, six companies out of ten with no written rule, two assistants picking the wrong figure in the right document. None of these cases comes from technology alone. Each one needs someone looking, at regular intervals, at what the tool actually does.

For HR, the practical consequence of section 03 is simple: document the tasks before talking about positions. The tutorial starts from there: a well-described task, then an agent that assists with it, under supervision.

Until next Sunday.

Jérôme

Your turn: reply with one sentence: "the task in my job that I would not hand to any tool is…". These concrete answers shape the next issues.

Was this edition useful?

One click is enough, I read every reply.

Forward it to a colleague

Forward it to someone who supports teams through change, who wonders what AI means for their own job, or who follows regulation closely. That is how SIGNAL AI finds its next readers, and its next questions.

Once a week, on Sunday. Nothing else, ever.