The Agenda — 2026-10
Two French firms in three use generative AI; one user in three sees a gain
Banque de France, 7,000 firms: 67% use it, 31% of users report a productivity gain, 6% a "significant" effect.
6394 articles read · 4091 from a credible source · 239 cleared the editorial judge · 17 kept.
The Banque de France surveyed about 7,000 firms with 20 or more employees between 25 February and 3 April 2026. 67% say they use generative AI; among those, 31% report a productivity gain, and 74% deploy it first in support functions.
The through-line
The items kept from 30 August to 30 September agree on one point: AI use has spread faster than the proof of what it returns. The Banque de France measures 67% of firms using it and 31% of those reporting a gain. A manufacturer that required 200 queries a month found nearly a third were repeated or off-task, and in Créteil a judge suspended job cuts for want of a trial in real conditions. On liability, a German court attributed a chatbot's false claims to the company, and the product liability directive, software included, must be transposed by 9 December.
Everyone's business
Banque de France · Tier A
The fact. The Banque de France surveyed about 7,000 firms with 20 or more employees between 25 February and 3 April 2026. 67% say they use generative AI; among those, 31% report a productivity gain and 6% a "significant" effect.
What it changes for you. If you are asked what AI has returned, the most common answer among your peers is that they have observed nothing. 74% of users deploy it first in support functions and 14% in production.
The question to ask on Monday. "Which figure shows what AI has gained us, and in which activity?"
Worth quoting. « Une minorité des entreprises utilisant les technologies d'IA constate des gains de productivité : 31 % chez les utilisatrices de l'IA générative, 41 % pour celles de l'IA non générative »
Self-reported answers; "use" includes experimental use. Scope: manufacturing, construction, non-financial market services excluding trade. As a share of all firms surveyed, the 31% becomes 22%.
Technology's Legal Edge · Tier B
The fact. On May 12, 2026, the Higher Regional Court of Hamm (OLG Hamm) ruled that a chatbot's misleading statements about Aesthetify's founder qualifications were attributable to the company itself (Case No. 4 UKl 3/25).
What it changes for you. If a chatbot answers your customers in your name, what it states can be attributed to you as the company's own commercial statement, including when the error is the model's and not your data's.
The question to ask on Monday. "How do our customer-facing AI systems ensure that information provided is accurate and not misleading, and who is accountable for verifying it?"
Worth quoting. “The court rejected the argument that the AI system operated as an independent third party.”
German ruling, on an action brought by the consumer association Verbraucherzentrale NRW. It is final: an appeal was allowed and none was lodged. It does not bind courts elsewhere.
Executive
arXiv · Université de Hong Kong · Tier B
The fact. A medical-device maker required roughly 5,000 employees to send at least 200 queries a month to its AI assistant: 31% of queries were repeated or off-task. Lowering the target to 100 cut volume by 30%, almost entirely on those queries, and salespeople's monthly sales rose by 7%.
What it changes for you. A usage-volume target produces volume. If you track AI adoption with a count of queries or logins, part of what you read is manufactured to hit the number.
The question to ask on Monday. "Which AI adoption indicator do we track, and how would we know it is inflated?"
Worth quoting. "query counts bunched at the threshold, and 31 percent of queries were repeated or off-task"
Preprint, not peer-reviewed, one company, in China. The sales increase concerns salespeople only.
Tech & Data
The Pragmatic Engineer · Tier B
The fact. Uber published figures for its coding agents: cost per session is down 52% from its June peak, and cost per 1,000 requests down almost 34%, with the model held fixed.
What it changes for you. Uber got this reduction without changing model, by working on how its agents call it. Pinterest says it runs open models, trained on its own data, at under 8% of the cost of comparable closed models.
The question to ask on Monday. "What does one session of our agents cost today, and what have we tried before considering a change of model?"
Worth quoting. "With open models, we are achieving cost per transaction at less than 8% of the cost of comparable closed proprietary models." — Bill Ready, CEO of Pinterest
Figures published by each company about itself. The source article credits Uber's reduction to open models; Uber's own post measures it with the model held fixed.
JDN : Derniers contenus · Tier B
The fact. In September 2026 the OECD published a working paper based on interviews with 25 organisations running AI agents. None reports letting them act without limits, and a person's approval is typically required for irreversible actions such as a payment or a data deletion.
What it changes for you. You do not have to choose between handing everything to an agent and reviewing everything: the organisations interviewed set a perimeter where errors can be corrected, and keep control of the rest.
The question to ask on Monday. "Which actions of our software cannot be undone, and who validates them today?"
Worth quoting. "No participating organisation reported deploying agentic AI systems with unrestricted autonomy."
The OECD states that its sample is illustrative, not representative, and that the examples are described by the organisations themselves. Checked against the paper.
Schneier on Security · Tier B
The fact. Researchers scanned 6,214 sites of large companies: on 120 of them, the documentation written for AI agents pointed to packages or domains nobody owned. They registered a few: within an hour a machine at a Fortune 500 company was running their package, then a few dozen more, installed by coding agents including Claude, Codex and Hermes.
What it changes for you. Your AI coding agents may be installing code from untrusted sources, creating a supply-chain attack surface your current security systems do not cover.
The question to ask on Monday. "How do our AI coding agents handle code dependencies from unregistered or abandoned domains?"
Worth quoting. "The trust model is broken," Alon Hertz, one of the researchers, wrote in an interview. "Agents treat vendor docs as ground truth and don’t question themand neither do the humans supervising them."
Measurement reported by an unnamed Israeli startup, relayed by Ars Technica and then by Bruce Schneier.
Legal, compliance & risk
JDN : Derniers contenus · Tier B
The fact. On 3 September 2026 the French data protection authority (CNIL) published a €500,000 fine against a private healthcare institution (decision SAN-2026-009). With a single account's credentials, an attacker read the files of 524,867 patients, and nothing allowed the activity to be spotted quickly.
What it changes for you. The failings apply to any software you open your data to, AI agents included: access rights wider than the need, and monitoring that does not catch abnormal activity.
The question to ask on Monday. "If an agent deployed at your organization were compromised tomorrow, would you be able to produce the list of people whose data it actually read?"
Worth quoting. Many organizations are waiting for the European AI regulation to structure their agent governance. It is likely that the first French sanction concerning an agent deployment will not be based on any of its articles.
The CNIL decision concerns an intrusion, not an AI agent. The parallel is drawn by the column's author (SiyadAI).
Operations
AutoGPT · Tier B
The fact. Grant Thornton surveyed 950 business leaders between 23 February and 18 March 2026. Among the 100 manufacturers in the panel, 48% are piloting AI, against 34% overall, and 10% have integrated it into operations, against 14%. None reports a significant revenue uplift or significant cost savings.
What it changes for you. If your AI pilots pile up without reaching daily operations, you are in the most common position in the manufacturing panel, and those in it report no significant gain so far.
The question to ask on Monday. "How many of our AI pilots have reached daily operations, and what stopped the others?"
Worth quoting. Manufacturing experiments more than any other sector and finishes less. That is a peculiar way to lose a race.
Report published by Grant Thornton in the United States; the respondents' country is not stated. The manufacturing subgroup has 100 respondents and its findings are flagged as directional. Figures checked against the report; the source article is a summary of it.
Product
Reed Smith · Tier B
The fact. EU Directive 2024/2853, which counts software and AI systems as products, must be transposed by 9 December 2026. On 8 September three Member States had done so, twelve had a draft, and France had published no text, according to the law firm Reed Smith's tally.
What it changes for you. The new regime covers products placed on the market after that date, software and AI included. Its content is set by the directive: you can prepare your files without waiting for the national text.
The question to ask on Monday. "Which versions of our products ship after 9 December, and what do our files say about their updates?"
Worth quoting. "France has confirmed an interministerial working group, Commission expert-group participation, and stakeholder consultations."
A law firm's tally, as of 8 September 2026; we could not check whether a French bill has been tabled since.
Finance
Towards AI - Medium · Tier B
The fact. OpenAI bills GPT-6 Astra prompts exceeding 272,000 input tokens at 2x the standard input and cache rates and 1.5x the output rate for the entire request.
What it changes for you. Your cost projections for agent loops or long-horizon tasks using GPT-6 Astra are likely understated if they cross the 272,000 input token threshold.
The question to ask on Monday. "Have our teams modeled the financial impact of an entire GPT-6 Astra request being billed at double the rate if it exceeds 272,000 input tokens?"
Worth quoting. “A request at 273,000 tokens bills the entire 273,000 at double.”
Pricing checked against OpenAI's documentation.
Marketing & Revenue
arXiv · Maastricht, Utrecht, Zurich · Tier B
The fact. Researchers from Maastricht, Utrecht and Zurich analysed 1,536 answers from ChatGPT, Gemini and Google's AI Overviews to real purchase questions. For the same question, ChatGPT and Gemini share on average only 5.4% of the sites they cite, and none in 76.7% of comparisons; the products recommended often change from one run to the next.
What it changes for you. What an assistant recommends to your customers depends on the assistant and on the run. A measure of your presence in AI answers taken once, or through the API, does not describe what the customer sees: ChatGPT's API shares only 12% of sites with its interface.
The question to ask on Monday. "How is our measure of presence in AI answers made, and how many times per question?"
Worth quoting. "for the same query, the ChatGPT and Gemini interfaces share only 5.4% of domains on average, with no domain in common in 76.7% of comparisons"
Preprint, not peer-reviewed. Questions in English, asked from the Netherlands in September 2026.
HR & transformation
IBM - Announcements (Artificial intelligence) · Tier B
The fact. A global study from the IBM Institute for Business Value, surveying 1,500 CHROs and 8,800 employees, finds that 60% of employees worry about skills erosion, with critical thinking cited most often as declining.
What it changes for you. Your team may feel exposed if they cannot challenge AI outputs without fear of blame, as 43% of employees report that when something goes wrong with AI, the blame falls on them.
The question to ask on Monday. "Have we clearly defined accountability when AI fails, and do employees feel safe to question or override AI recommendations?"
Worth quoting. "80% of CHROs believe AI adoption creates “invisible” work for employees, including validating recommendations, fixing mistakes, providing context and managing exceptions."
Figures are self-reported by survey respondents to IBM.
In the professions
arXiv · prépublication · Tier B
The fact. At ICML 2026, which received over 24,000 papers, some reviewers were randomly placed under a policy prohibiting all use of language models, the others under a policy allowing limited assistance. Decisions and scores stayed almost identical; 22.5% of reviewers under the ban say they used a model, and 36.5% of the others report a use that was not allowed.
What it changes for you. A rule on AI use in evaluation, whether it bans or frames, is not followed by part of the evaluators. If you run a committee, a jury or a journal, your written rule does not tell you what is happening.
The question to ask on Monday. "What rule have we written for our committees and juries, and how would we know it is followed?"
Worth quoting. "Policy assignment had near-zero effects on final paper decisions, paper scores, and reviewer confidence, although reviews under the permissive policy were 5.5-7% longer."
Preprint, not peer-reviewed, run with the conference organisers. The share of reviewers outside the rule comes from an anonymous survey of 1,486 answers: it is self-reported.
Sommet Lumière · Tier A
The fact. On 7 September 2026, eleven States, including France and Korea, and film organisations signed a declaration calling for rights holders to be paid when models are trained on their works, for transparency on training data and for generated content to be identified. Xilam, Illumination Studios Paris, Gobelins and CITIA are among the signatories; the Motion Picture Association and its members are not.
What it changes for you. The text binds no one and sets no deadline. It tells you who, among your commissioners and partners, has committed in writing on model training, and who has abstained.
The question to ask on Monday. "In our current contracts, who has the right to train a model on our images, and where is it written?"
Worth quoting. « respecter les droits de propriété intellectuelle et donner lieu à une rémunération équitable des titulaires de droits »
No AI company signed. Transparency is asked for "with due regard for trade secrets".
NBER · document de travail · Tier B
The fact. In a three-month randomised trial, 133 patent lawyers at eleven US firms were or were not given a drafting assistant, and their work was graded blind by fellow patent attorneys. With the tool, quality rises, more so among juniors. On the final test without it, those who had had the tool do better, and all of that advantage sits with the seniors.
What it changes for you. Juniors are the ones whose work improves most with the tool, and on average they keep none of it once it is taken away: some improve, others fall back. The quality of what an associate hands in no longer tells you what they have learned.
The question to ask on Monday. "On which work are our junior associates still assessed without the tool?"
Worth quoting. "The largest gains from AI thus accrued to the lawyers who retained the least."
Working paper, not peer-reviewed. Google funded the experiment, supplied the tool, and six of the seven authors work for it. 91 of the 133 lawyers took the final test, the starting level was not measured, and the junior-senior gap with the tool is not statistically established.
AFP, via Boursorama · Tier B
The fact. On 15 September 2026 the Créteil court, in summary proceedings brought by the works council, suspended the removal of eight sub-editor posts at Gisi, publisher of L'Usine nouvelle and LSA. Management estimated that AI could take over 70% of their tasks; the judge orders full trials first.
What it changes for you. Estimating that a tool will do 70% of a job is not enough to ground a reorganisation: the judge asks for a trial in real conditions. The ruling rests on the risk to the health of the staff who remain, a duty that binds every employer.
The question to ask on Monday. "For each task we hand to the tool, who checks the result, how long does that take, and have we measured it in real conditions?"
Worth quoting. « un risque grave et caractérisé sur la santé physique et mentale »
Summary order, not published: we know it through AFP. Management has said it intends to appeal.
Still standing
JDN : Derniers contenus · Tier B
The fact. The second part of the AI Act began applying on August 2, 2026, with transparency obligations for AI-generated content, while requirements for high-risk AI systems are postponed until December 2, 2027.
What it changes for you. Your teams must now clearly inform users they are interacting with AI, and ensure all AI-generated content is technically detectable as such, even as high-risk system obligations are delayed.
The question to ask on Monday. "Have we identified all customer interaction points where AI is used and verified that AI disclosure is clearly provided?"
Worth quoting. "For example, if a company uses a customer service agent, the customer must understand that they are interacting with an AI system and not a human."
The cut of the month
AI Is Supposed to Be Taking Jobs. Why Are the Jobs Most Exposed to AI Paying 46% More? — Artificial Intelligence on Medium
This claim is too good to be true and flattering for executives, but it lacks method and changes no decision.
Lire cette édition en français
All editions