The Agenda — 2026-11

In the occupations most exposed to AI, French youth hiring on permanent contracts is 28% below its 2022 level

Banque de France: −28% permanent hires of 21-29-year-olds in the most exposed occupations, −8% in the least exposed, and total employment not clearly moving.

According to a Banque de France post published on 8 October 2026, permanent-contract hires of 21-29-year-olds in the fifth of occupations most exposed to AI were in March 2026 about 28% below their November 2022 level, against 8% in the least exposed fifth. Total employment at firms that use AI heavily does not differ clearly from other firms.

The through-line

The items kept from 1 to 31 October concern what AI changes at the entry point of trades, and who answers for what it produces. In France, youth hiring on permanent contracts has fallen in the most exposed occupations, and at Anthropic the share of juniors in technical hires has dropped from 27% to 9%. Two German courts treat the answer of Google's AI as Google's own text, and a US appeals court denies fair use to a tool trained to compete with the publisher whose content it took. On costs, Microsoft now bills Copilot's heavy tasks by usage, and a study shows that switching coding tool changes the bill first.

Everyone's business

Generative AI: the first labour-market changes fall on young workers

IA générative : les premiers changements du marché du travail sur les jeunes

Banque de France · Bloc-notes Éco n° 463 · Tier A

The fact. According to a Banque de France post published on 8 October 2026, permanent-contract hires of 21-29-year-olds in the fifth of occupations most exposed to AI were in March 2026 about 28% below their November 2022 level, against 8% in the least exposed fifth. The gap narrows with age and almost disappears after 50.

What it changes for you. The first effect measured in France does not show in headcount: total employment at firms that use AI heavily does not differ clearly from other firms. It shows at the entry point, in junior positions, and with them in the way a trade is learned.

The question to ask on Monday. "How many beginners did we hire this year in our occupations most exposed to AI, and who teaches them the trade?"

Worth quoting. « Suivre seulement l'emploi total pourrait donc manquer une partie de l'ajustement »

A correlation: the authors write that they cannot yet separate the effect of AI from the selection of firms that adopt it. These are smoothed flows of permanent-contract hires from French labour-movement filings. Checked against the post.

For two German courts, what Google's AI says, Google says

Two German courts now hold Google liable for what its AI overviews say

The Decoder · Tier B

The fact. On 24 September 2026 the Aachen Regional Court issued an interim injunction barring Google from showing in its "AI Mode" three texts that attributed to a medical professor another man's criminal conviction, on pain of a penalty of up to €250,000 per breach. It follows a Munich judgment of 28 May 2026 on the search engine's AI Overviews.

What it changes for you. After the Hamm ruling on a seller's chatbot, which we covered in October, two courts treat a search engine's answer as its operator's own text. What your AI tool says about a person or a product binds you, even when it only summarises other people's pages.

The question to ask on Monday. "If our AI tool states something false about a customer or a competitor, who sees it, and how fast do we correct it?"

Worth quoting. « Die Antragsgegnerin stellt die KI-generierte Suche bereit und ist damit auch für die insoweit generierten Antworten verantwortlich. » — Aachen Regional Court

Two interim rulings at first instance: Google can lodge an objection in Aachen, nothing has been decided on the merits and no damages were awarded. The source article's headline speaks of AI Overviews; in Aachen it is AI Mode. Rulings checked at source.

Executive

The Dutch tax administration drops the Microsoft 365 cloud and moves back to its own servers

Toekomstbestendigheid Belastingdienst — Kamerbrief 31 066, nr. 1579

Tweede Kamer · lettre du gouvernement · Tier B

The fact. In a letter to Parliament dated 5 October 2026, the Dutch State Secretary for Finance announced that the tax administration, customs and the benefits agency are dropping the move to Microsoft 365 decided in 2025. In 2027, email and calendar move to their own servers, using existing licences; in late 2027 and 2028, storage and collaboration tools move to European open-source software.

What it changes for you. Reversing a supplier choice can be done, and it is planned over two years, with functions lost along the way: some planned archiving functions will not be available. If your email, your documents and now your AI depend on a single suite, this timetable gives the order of magnitude of an exit.

The question to ask on Monday. "If we had to leave our office suite and the AI that comes with it, what would we lose, and how long would it take us?"

Worth quoting. « De Belastingdienst, Douane en Toeslagen kiezen nu voor een on-premises oplossing in combinatie met open source. »

A public administration's decision, taken after a critical opinion from the Dutch government's IT project review body. According to the Dutch press, email will stay on Microsoft software installed on government servers; we have not checked this at source. Letter checked at source.

Tech & Data

PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet

wiz.io - Www.wiz.io · Tier B

The fact. On 21 October 2026 the PHP project released versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35, fixing two flaws that allow unauthenticated remote code execution, rated 9.2 out of 10. The root cause was reported in August by several independent teams; Wiz says it searched for it with AI models, and the PHP Foundation validated the reports with an OpenAI tool.

What it changes for you. Your website, extranet or online shop very probably runs on PHP, often at a provider. The update takes a few hours; what is usually missing is knowing who decides it.

The question to ask on Monday. "Which PHP versions run our sites and our providers' sites, and who confirms to us that they are up to date?"

Worth quoting. "PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 fix two critical vulnerabilities that can allow unauthenticated remote code execution." — php.net

The headline's "70% of the Internet" is PHP's share according to W3Techs (69.2% of sites whose language is known), not the number of vulnerable sites. The 10 out of 10 score that circulated is wrong according to the PHP Foundation. Announcements checked at source.

What Does a Harness Buy? Tokens, Mostly

arXiv · université du Shandong · Tier B

The fact. Two researchers at Shandong University ran five models in three coding-agent harnesses, including Claude Code and OpenCode, on 447 tasks of the SWE-bench Verified benchmark. Switching harness changes the outcome of 13% of tasks, as much as running the same setup twice, and the cost per task varies up to threefold.

What it changes for you. A score gap announced between two tools can come down to the luck of one run. What certainly varies is consumption: for the same result, the most expensive tool can cost three times as much.

The question to ask on Monday. "When we chose our AI coding tool, how many times did we rerun the test, and what did we compare: the result or the bill?"

Worth quoting. "What a harness buys, on this evidence, is fewer ways to lose a task, and a bill, not pass rate."

Preprint, not peer-reviewed. On a sample of 45 hard tasks, the study detects a 13-point gap only half the time: it shows that small differences cannot be seen, not that none exist.

Legal, compliance & risk

Training a tool on a publisher's content to compete with it is not fair use, a US appeals court rules

Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 25-2153 (3d Cir.)

Cour d'appel fédérale des États-Unis, 3e circuit · Tier B

The fact. On 29 September 2026 the US Court of Appeals for the Third Circuit confirmed that Thomson Reuters could protect 2,243 case headnotes from its Westlaw database, and that Ross Intelligence could not claim fair use for using them to train a competing legal search engine. The case goes back to the Delaware court for the rest.

What it changes for you. The court relies on two features: Ross's tool wrote nothing, it pointed to existing decisions, and it aimed to replace Westlaw with the same customers. If you build a tool on a supplier's content to sell the same service as that supplier, this is the case the ruling describes.

The question to ask on Monday. "Which third-party content were our AI tools built on, and does one of those third parties sell the same service as us?"

Worth quoting. "We therefore hold that Thomson Reuters's materials are copyrightable and that ROSS's use was not fair."

US law. The court expressly sets generative models aside: it does not rule on the training of assistants such as ChatGPT. Rehearing or further appeal remains possible. Opinion checked at source.

Operations

With Copilot, employees work with more colleagues, not fewer

The Effects of Generative AI on Collaboration and Social Networks Among Knowledge Workers

arXiv · université Bocconi · Tier B

The fact. Researchers at Bocconi University followed more than 12,000 employees of a multinational over the 13 months of a staged Microsoft Copilot rollout, using the traces of their email and meetings. Users interact with more colleagues and their networks become less clustered; they also concentrate more of their exchanges on colleagues they already worked with.

What it changes for you. The tool does not replace exchanges between colleagues, it adds to them: the authors explain this by what its output requires, which must be interpreted, checked and accepted. A rollout review based only on time saved leaves that time out.

The question to ask on Monday. "In our AI review, have we counted the time our teams spend together checking what it produces?"

Worth quoting. "Those using generative AI interacted with more colleagues and their networks became less clustered, indicating broadening in collaboration structures."

Preprint, not peer-reviewed, on a single anonymous company and data from 2024 and 2025. The traces measure the number of exchanges, not their quality.

Product

Commission Implementing Decision (EU) 2026/2274 on harmonised standards for high-risk artificial intelligence systems

Décision d'exécution (UE) 2026/2274 de la Commission relative aux normes harmonisées sur les systèmes d'intelligence artificielle à haut risque

Journal officiel de l'Union européenne · Tier A

The fact. On 14 October 2026 the Commission published in the Official Journal the reference of the AI Act's first harmonised standard, EN 18286:2026, which covers the quality management system of providers of high-risk AI systems, post-market monitoring included. Since that date, complying with it gives a presumption of conformity with the requirements it covers.

What it changes for you. If you provide a system classed as high-risk, you can now build your quality system on a standard that gives a presumption, before the obligations apply. The decision publishes only this standard: for logging, transparency or human oversight, nothing yet gives a presumption.

The question to ask on Monday. "Which of our products or suppliers fall under high risk, and which standard is our quality system based on?"

Worth quoting. « El cumplimiento de una norma armonizada confiere una presunción de conformidad con los requisitos esenciales correspondientes » — recital 6

Text read in its Spanish Official Journal version. The standard itself is sold by standards bodies and we have not read it.

Finance

Evolution of the Copilot pricing model

Microsoft Tech Community · Tier B

The fact. On 25 September 2026 Microsoft announced that Copilot's per-user licence covers everyday use, and that agents such as Cowork, Code and Autopilot, as well as some models, are billed by usage in Copilot Credits. For enterprise customers, that billing stays off until an admin creates a spending policy.

What it changes for you. The price of a seat no longer tells you what the tool will cost: the variable part depends on what your teams hand it. Nothing is billed by usage before someone at your firm has set a spending policy, and that person holds the budget.

The question to ask on Monday. "Who at our firm sets the Copilot spending policy, with what cap, and who sees the bill each month?"

Worth quoting. "For enterprise customers, UBB services stay off until an admin creates a spending policy in the Microsoft 365 admin center, and nothing is billed before that."

Vendor announcement. For Business offers sold through resellers, Microsoft switches this billing on by default from 1 December 2026, but not in France at launch. The original column attributes the same change to Google, OpenAI and Anthropic: at Google it is a feature withdrawal, and no date is published for the other two.

HR & transformation

At Anthropic, juniors fall from 27% to 9% of technical hires

How employment is changing at Anthropic

Anthropic · Tier B

The fact. Anthropic published its own data on 16 October 2026: the share of juniors among its technical staff has fallen from 21% to 14% since early 2025, and among its technical hires from 27% to 9%. Its assistant's sessions spend about 80% of their time waiting for human input.

What it changes for you. A company that pushes use further than most hires fewer beginners, and what slows it down is its employees' time to review and decide. If you narrow the entry point, you need another path to train those who will hold the senior positions.

The question to ask on Monday. "If we hire fewer juniors, how will the people who hold our senior positions in five years get there?"

Worth quoting. "The proportion of junior technical hires fell even faster, from 27% to 9%."

Data published by an AI company about itself, in a fast-growing firm. The authors point out that adoption is not random, that several measures are made by a model and that one survey question changed along the way.

In the professions

Scientific research: untraceable references are multiplying in published papers

Evaluation of possibly fabricated references in 500,000 science, engineering, and computing papers

PNAS · Columbia University · Tier A

The fact. A Columbia University team checked 21.6 million references from nearly 500,000 science, engineering and computing papers published from January 2023 to March 2026. 1,135 references, in 619 papers, cannot be found in any of seven databases; their rate rises from 6.3 per 10,000 papers in 2023 to 15.4 in 2025 and 23.0 in the first quarter of 2026. Two of these papers have been retracted.

What it changes for you. An untraceable reference gets through review and publication: the check is not done downstream. If your laboratory or research department signs publications, it can be done by program, before submission.

The question to ask on Monday. "Before a paper goes to the publisher, who checks that each reference exists, and with what tool?"

Worth quoting. "fabricated is an operational label for a claimed title that could not be located in any of seven bibliographic and open-web sources."

The authors do not attribute these references to language models: "fabricated" means untraceable. Two experts confirmed 85.8% of flags on a sample. We read the abstract and the press release, not the full paper.

Law: in the UK, the Law Society asks firms to leave juniors the work that trains them

Using artificial intelligence in legal practice

Law Society of England and Wales · Tier A

The fact. On 21 October 2026 the Law Society of England and Wales published a practice note on AI. It separates the supervision of people from the oversight of AI-assisted work, and asks those responsible for junior colleagues to ensure that AI does not take away their chances to learn research, drafting and judgment.

What it changes for you. The text puts the burden on supervisors and says what it costs: telling managers to supervise is not enough without giving them the time, training, tools and someone to escalate a doubt to. A firm elsewhere can use it as a checklist without being bound by it.

The question to ask on Monday. "What time have we given our partners to supervise the work juniors do with AI, and have we written it down?"

Worth quoting. "It will likely not be sufficient to instruct managers to supervise AI-assisted work without providing the time, training, tools and escalation routes needed to do so effectively."

Good-practice note, not binding: it states it is not legal advice. It addresses solicitors in England and Wales. Checked at source.

Media: Euronews has its commentary read by a synthetic voice

Euronews : IA et voix sans issue

SNJ · Tier B

The fact. Since October 2026, a synthetic voice produced by start-up Dubformer's software has read Euronews commentary from the newsroom's texts, translated into the channel's 19 languages. According to the French journalists' union SNJ, management estimates the time saved at 40% and did not present the project's rationale to the works council.

What it changes for you. Voice and translation are production jobs, and that is where the tool comes in first, with a gain quantified by management alone. What remains to be shown is the time it takes to check 19 languages, and by whom.

The question to ask on Monday. "For each language or version the tool produces, who checks it before broadcast, and how long does that take?"

Worth quoting. « Ses motivations n'ont pas été présentées au comité social et économique d'Euronews. »

Only the union's version is public; we have not read a response from management. The 40% figure is a management estimate reported by the SNJ.

Journalism: Ebra wants to hand part of the sub-editors' work to AI

MOTION - Soutien aux salariés du groupe Ebra

SNJ · Tier B

The fact. The Ebra group, Crédit Mutuel's press arm, plans to cut nearly 400 jobs across its nine regional dailies, including Le Progrès, Le Dauphiné libéré and L'Est républicain. According to the French journalists' union SNJ, prioritising, proofreading, rewriting, correcting and layout would go to AI; the strike of 9 October was 67% followed, according to the union.

What it changes for you. It is the same job as at L'Usine nouvelle and LSA, where a court suspended a similar reorganisation in September for want of a prior trial. Ebra announces voluntary departures; in Créteil, the judge ruled on the workload of those who stay, not on how the others left.

The question to ask on Monday. "Which proofreading tasks do we hand to the tool, and who answers for a published error?"

Worth quoting. « la direction du pôle presse du Crédit Mutuel entend supprimer les postes de journalistes secrétaires de rédaction »

Union figures; management has not published a strike rate, and the voluntary nature of departures is reported by AFP. Negotiations end on 2 November 2026, the day this issue appears; their outcome was not known when it closed.

The cut of the month

The Latest SaaStr AI Agent Playbook for GTM Is Out: Everything We’ve Learned Running SaaStr on 3 Humans and 21+ AI Agents, In One Handy Guide — SaaStrAI

SaaStr tells what worked for itself, with figures supplied by the party concerned, with no method and no point of comparison: it is a story, not a measurement.

Lire cette édition en français

All editions