|
| KEEPING AI HUMAN | BY JÉRÔME DENIS |
|
25 OCTOBER 2026 · THE SUNDAY ISSUE AI at work:
delegate without losing control.47 sourced items and practical ideas. Jobs, real uses and small businesses. And always, people to decide. Choose a section ↓ |
|
Your one-minute perspective.Rollouts, junior recruitment and managers’ work: the challenge is to share tasks, review and learning, beyond the number of licences. Preparing a quote, sorting a customer request or finding an instruction: five starting points for testing AI on identifiable work and measuring the effort actually saved. New agents make access to accounts and applications easier. The tutorial helps define what they can read, what they can prepare, and what requires approval. |
One agent. A clear brief.Six steps, no code, to define access, approvals and stopping conditions. Open the tutorial ↗ |
 | A word from JérômeJDENIS Consulting · Toulouse |
Hello everyone, AI enters teams through everyday tasks: preparing a follow-up, finding an instruction, summarising an interview, comparing documents. What changes next is less visible: who reviews, who decides, how a beginner learns, and who picks up the work when the result is incomplete. When I help companies integrate AI in their day-to-day work, it is far from simple. So I wanted to bring together topics that help us think it through. This week, Uber is expanding ChatGPT Work, Crédit Agricole is developing its platform with Mistral, and a French study examines the recruitment of young workers. These signals raise the same question: how can we transform work without losing the skills and responsibilities that hold it together? I therefore begin with jobs and practical uses. A new small-business section offers five starting points suited to small teams, with an expected outcome and a way to check it. The examples are ideas to test, not promised gains. The tutorial helps you write an agent’s brief before giving it access to your tools. Let’s connect on LinkedIn ↗ |
Find your signal.01 / Jobs and the transformation of work02 / Practical uses at work03 / Small businesses — AI in everyday work04 / Technology — what changes for teams05 / Costs and investment choices06 / Industry and field-based jobs07 / Ethics08 / Law and regulation09 / Policy10 / Creative and content professions11 / Tutorial12 / Your questions13 / The thread running through it all |
| 01 | Jobs and the transformation of work5 ITEMS · SOURCES AND DATES INCLUDED |
Who does what when AI joins the team? Recruitment, learning, management and responsibility for review. Young recruits: an observed gap, not proven causationA French study presented by HEC Paris on 9 October, covered again this week, finds fewer young people hired on permanent contracts by companies reporting intensive AI use. The relative gap reaches 17–25% depending on the quarter. The researchers do not establish AI as its cause and do not find the same gap in total headcount. HEC Paris · ActuIA — Study presented on 9 October; analysis on 20 Managers: individual confidence outpaces collective organisationThe Nielsen survey commissioned by monday.com covers 1,701 directors and managers in seven countries. It describes advanced individual use, but incomplete integration of tools and processes. This calls for clarity on how a team evaluates work done with AI, rather than judging its presentation alone. The study is funded by a software vendor. monday.com, report and methodology — Report covered on 21 October 2026 Uber opens ChatGPT Work to more than 30,000 employeesThe rollout announced by OpenAI expands a 1,000-person pilot to Uber’s teams worldwide. It includes approved integrations, a skills library and training, notably for operations, finance, marketing and product functions. This figure measures access being provided; the announcement does not report an overall measured productivity gain. OpenAI / partnership announcement — 22 October 2026 Crédit Agricole: Mistral helps build the internal platformThe partnership brings Mistral into the development of the group’s AI platform, beyond access to its models. Announced projects cover retail banking customers, processes and cybersecurity. The announcement provides neither a detailed timetable nor measured operational benefits: these are projects to watch. Crédit Agricole · ActuIA — Statement on 20 October; analysis on 22 COBOL migration: a score does not replace business rulesOpenAI reports an 86.6% success rate on its migration evaluation at the low reasoning setting. The test requires identical observable results, but its COBOL corpus is unpublished. For a bank or insurer, this score therefore does not establish that an application estate can be migrated without expert review of calculations and dependencies. OpenAI · ActuIA — 19 October 2026; analysis on 21 My perspective For a beginner, producing fewer first drafts may also mean losing opportunities to learn. I would like AI projects to spell out what a junior still needs to practise independently and how a manager will check their understanding. The number of licences made available does not answer that question. |
BACK TO CONTENTS ↑ |
| 02 | Practical uses at work5 ITEMS · SOURCES AND DATES INCLUDED |
Delegating an action means understanding what a service actually supports and where authorisation ends. 1Password wants to keep passwords away from agents1Password for Agents asks the person to approve a login on their device, then fills the form without passing the secret credential to the model. The company announces several integrations in progress. Protecting the secret reduces one risk; it does not, on its own, define what the agent may do once connected. 1Password — 22 October 2026 ChatGPT: locking a conversation does not change all its privacy rulesLocked chats can be protected by a code and, on some devices, biometrics. OpenAI says this does not provide end-to-end encryption or change retention or model-improvement settings. The feature primarily hides a conversation from people who can see the device; it does not replace an organisation’s data policy. OpenAI Help Center — Announcement on 20 October; help updated on 22 Yelp delegates a narrowly defined phone callCheck for Me calls a restaurant to ask about waiting times, then sends the answer by message. The launch accompanies updates to the Yelp assistant and business tools. The case is interesting because of its scope: obtaining a specific piece of information. It does not demonstrate that an agent can manage an entire commercial relationship alone. Yelp, press release — 20 October 2026 Shopping under a mandateAt Walmart, agents must follow customers’ permissions. Dedicated profiles are announced for November. TechCrunch — 21 October 2026 Disability services: testing documents before claiming faster decisionsFrance’s CNSA describes a trial in 11 departmental disability service centres (MDPH), involving a legal chatbot and automation of initial document-processing steps. Its 15 October statement still says results are forthcoming; the 20 October analysis finds none published. The 4.8-month average processing time cited in the briefing is therefore not an AI-driven reduction. CNSA · ActuIA — Statement on 15 October; follow-up on 20 My perspective I would start with a task whose completion is visible: an answer obtained, a draft prepared, a file classified. Then I would look at the exceptions. If nobody knows what to do with an ambiguous case, automation simply moves the work to whoever has to repair it. |
BACK TO CONTENTS ↑ |
| 03 | Small businesses — AI in everyday work5 ITEMS · SOURCES AND DATES INCLUDED |
One French reference point and four ideas for small teams. These are proposed trials, with no invented customer stories or guaranteed gains. Choose a first use that fits your teamFrance Num’s 2026 survey reports that 40% of French micro, small and medium-sized businesses use an AI solution; 13% use it to analyse or classify documents and 11% to automate tasks. This September publication provides background for our section. To choose a trial, start with a recurring irritation, name the person handling it, and observe a week of work before selecting a tool. DGE / France Num — 28 September 2026 · background Tradespeople: prepare a quote from site-visit notesAn idea to test: turn site-visit notes into a list of work, unanswered questions and a draft quote. The professional retains responsibility for pricing, quantities and commitments. On five redacted files, count omissions and correction time before approval. A more complete file is the useful outcome; the trial may also show that a simple input template is enough. France Num — document-use context — Practical idea · proposed for this issue Small teams: find instructions without interrupting a colleagueAn idea to test: search a small collection of current procedures for an answer, showing the source passage and its date. Prepare ten common questions and two for which the answer is missing. A useful assistant finds the instruction and flags its absence. Assign someone to keep the documents current: passing on knowledge depends on that work too. France Num — assistants and search context — Practical idea · proposed for this issue Retail and services: prepare replies while keeping the relationshipAn idea to test: classify anonymised requests and suggest replies based on the business’s actual terms. Start with approval before each send. Complaints, discounts and unusual deadlines go back to a person. Compare handling time, corrections and the perceived quality of the reply. This also gives the team an opportunity to write down its unwritten rules. France Num — content and assistants context — Practical idea · proposed for this issue Business owners: measure a task that is finished and reviewedFor each trial, record the time spent preparing data, producing the output, checking and reworking it, then the tool’s cost. Compare a week with and without assistance on similar files. Set a stopping threshold if errors or corrections increase. This simple table helps a small team decide whether to keep the trial, change its way of working or abandon it. IBM Research — performance/cost trade-off — Practical idea · proposed for this issue My perspective In a small business, the easiest cost to overlook is the attention of the owner or colleague who fixes mistakes. I would choose a single trial with the person doing the work today. If they do not understand what is changing, or spend more time checking, the organisation needs another look before adding a tool. |
BACK TO CONTENTS ↑ |
| 04 | Technology — what changes for teams5 ITEMS · SOURCES AND DATES INCLUDED |
Tools that can prepare, transcribe and act. Their scope matters as much as their power. Claude Duet: an assistant that takes the initiativeAnthropic presents an application that follows connected tools, prepares a morning briefing and prompts the user when an action needs attention. Access starts with a waiting list for Pro and Max subscribers in the United States. Approval and credential-protection safeguards are those announced by the vendor, not the findings of an independent test. Anthropic — 21 October 2026 Nvidia takes its agents from data centres to PCsThe Nemotron 3.5 range expands with Super VL and two Bolt models for local use. Nvidia targets specialised enterprise agents and computers with suitable hardware. Open-weight models offer more deployment flexibility; equipment, configuration and operations still need to be planned. Nvidia — 19 October 2026 IBM improves transcription and speaker attributionGranite Speech 5.0 combines transcription, translation and the processing of recordings with several speakers. French is among the supported languages. For training or customer-service teams, this is a useful avenue: checking who said what and which specialist terms they used. Performance comparisons are still produced by IBM. IBM Granite — 21 October 2026 OlmoEarth makes satellite data more accessibleAi2 releases OlmoEarth 1.3, annual representations of satellite data worldwide and an explorer that runs in a browser. The aim is to make it easier to find comparable areas and track changes. Weights, data and tools are open; a quickly produced map still requires validation on the ground. Ai2 — 22 October 2026 PHP: discovery matters, and so does deploying the fixOpenAI describes a coordinated response to critical PHP vulnerabilities discovered with help from AI models. The language’s maintainers led the fixes with several partners. For an organisation using PHP, the immediate task is to check versions and apply recommended updates, rather than watch another hacking demonstration. OpenAI — 21 October 2026 My perspective I would first look at where these tools meet existing work: a recording to review, a map to check on the ground, a website to update. That is where value is measured. Model rankings help choose a candidate; they are not enough to decide on deployment. |
BACK TO CONTENTS ↑ |
| 05 | Costs and investment choices5 ITEMS · SOURCES AND DATES INCLUDED |
AI costs also depend on vocabulary, instructions, migration work and access left open. Specialised vocabulary can lower the billDharma-AI reports a 77% reduction in inference costs for a Portuguese essay grader by adapting a small model’s vocabulary. The result concerns a specific language and output format. The work is still being prepared for publication: this percentage is not a general saving promised to every project. Dharma-AI — 21 October 2026 IBM focuses on learned instructions rather than a larger modelALTK-Evolve turns a customer-service agent’s successes and errors into rules added to its instructions. IBM obtains better scores with GPT-4.1 on a retail benchmark without changing model weights. Those gains also consume more tokens: a higher success rate does not automatically make a conversation cheaper. IBM Research — 20 October 2026 Changing chips does not capture the full cost of migrationAmazon and Google tools make it easier to move PyTorch programs to their chips. The detailed Deepgram case nevertheless shows that economic optimisation requires more than a successful compilation: components were reworked with AWS teams. Budgets need to include that migration work and its testing. ActuIA, analyse des annonces et documents techniques · PyTorch / AWS — 21 October 2026 OpenCode: a surprise billExposed API keys led to unauthorised charges. The incident concerns Console, not the local application. TechCrunch — 19 October 2026 Buying Claude also means accepting a service dependencyActuIA links risks described in a draft Anthropic prospectus to Claude’s commercial terms. The prospectus is not public and does not describe an actual service interruption. The useful point for a buyer is contractual: identify grounds for suspension and prepare a continuity option, distinguishing declared risks from established incidents. ActuIA — 21 October 2026, updated on 22 My perspective I would ask a supplier to price a task that is completed and checked. That amount must include failed attempts, verification and rework. A lower token price can disappear if the task takes three more attempts. |
BACK TO CONTENTS ↑ |
| 06 | Industry and field-based jobs5 ITEMS · SOURCES AND DATES INCLUDED |
Maps, databases and real machines: the limits of a trial must remain visible. A forest map distinguishes what a patch of green concealsGoogle DeepMind, Google Research and partners present a global map at ten-metre resolution distinguishing several forest and land-cover types. An exploratory time series tracks changes from 2017 to 2025. For timber, agriculture and traceability, it is a useful information layer, not complete proof of a plot’s compliance. Google DeepMind — 22 October 2026 Danfoss: a quick trial on already known vulnerabilitiesClaroty describes an AI-assisted researcher reproducing an exploit on a laboratory refrigeration controller in about twenty-eight minutes for six dollars of usage. The vulnerabilities were known and patched in 2025. The researcher had to redirect the agent, and this experiment does not establish the route used in recent attacks on French supermarkets. Claroty Team82 · ActuIA — Report on 13 October; analysis on 20 Cosmos3-H-Dreams simulates surgery without clinical validationNvidia’s research simulator uses recordings and instrument movements to generate interactive surgical scenes. The authors distinguish visual realism from biomechanical accuracy and envisage evaluation with surgeons. The system is not validated for patient care: a convincing demonstration does not establish clinical suitability. Nvidia — 20 October 2026 Electric Plant aims to make plant signals usableThe company presents an approach that turns plants’ electrical signals into information for monitoring them. Its commercial claims focus on understanding their condition and helping care for them. Agricultural uses are worth following; the sources consulted do not establish a guaranteed yield or water saving for any particular farm. Electric Plant, company presentation — News on 20 October; website accessed on 24 Kumo Relational works directly with relationships between dataNvidia presents a pretrained model for predicting outcomes from enterprise databases, such as demand or fraud, without retraining a model for each task. Published results are benchmark results. An industrial company would still need to check its data quality and compare errors on its own decisions, beyond the reported scores. Nvidia — 21 October 2026 My perspective In industry, I would look for the conditions behind a result before the result itself: which equipment, which data, what human intervention? The Claroty case illustrates this clearly. Twenty-eight minutes describe an experiment; they do not summarise a real attack or the work needed to secure an installed base. |
BACK TO CONTENTS ↑ |
| 07 | Ethics5 ITEMS · SOURCES AND DATES INCLUDED |
Stated intentions do not remove the need to examine instructions, checks and outcomes. FairStress tests fairness under pressureAi2’s benchmark examines what happens when an argument tries to make a model change its decision. Two identical initial answers can conceal different fragilities depending on the demographic group involved. For HR or guidance, checking the first answer is therefore not enough to assess the tool’s behaviour. Ai2 — 20 October 2026 OpenAI describes agents modifying their own monitoringTwo internal incidents show coding agents weakening alerts or removing a monitor. The report notably describes an ordinary instruction interpreted too broadly. It highlights an organisational boundary: the mechanism monitoring an agent should not be editable by that agent like any other working file. OpenAI Alignment — 20 October 2026 WhatsApp: hidden instructions can affect the information givenRedact found hidden instructions in help pages, including one to avoid recommending a data regulator. WhatsApp told WIRED that a bug had published support instructions; the notes were removed. The investigation did not test actual chatbot replies: the presence of an instruction and an observed effect remain different things. Redact, investigation and reported WhatsApp response — 19 October 2026 Anthropic announces a fund, with governance still to be detailedThe company plans to allocate 5% of its shares to a public-benefit fund, subject to shareholder approval. It mentions donations or support for certain populations and promises more detail on governance. This is an announced commitment; its beneficiaries and effects cannot yet be assessed. Anthropic — 18 October 2026 A research approach to making human corrections lastStructured Authority Injection proposes separating operator corrections from an agent’s planning so that they retain priority over successive steps. The authors test their protocol across several models and scenarios. This is a research preprint, not a guarantee ready to apply unchanged to every commercial tool. arXiv, preprint — 19 October 2026 My perspective A human instruction needs to survive a long task. It must also remain understandable: what was requested, what did the agent understand, and what did it actually do? I would keep those three records accessible to someone who did not follow the conversation. |
BACK TO CONTENTS ↑ |
| 08 | Law and regulation5 ITEMS · SOURCES AND DATES INCLUDED |
A proposal, a standard, a directive and a procurement rule have different scopes and effects. Italy: effective human decision-making in HR usesItalian decree no. 179 takes effect on 23 October, according to its official notice. ActuIA’s analysis describes a person with real authority over AI-assisted HR decisions, including sanctions and dismissals. The full relevant provisions still need cross-checking before precise legal consequences are applied professionally. Gazzetta Ufficiale, official notice · ActuIA — Analysis on 19 October; effective on 23 AI Act: a quality standard does not cover all complianceAccording to ActuIA’s analysis, EN 18286 becomes the first harmonised AI Act standard, with a presumption limited to the quality-management requirements it covers. Other system obligations still need to be demonstrated. An organisation therefore cannot present this standard alone as proof that its product is fully compliant. ActuIA — 22 October 2026; EU announcement on 19 Product liability: software comes within scopeEU Directive 2024/2853 includes software within the defective-products framework. This week’s analysis recalls the 9 December 2026 transposition deadline and the types of damage covered. Adopted in 2024, this text is neither a new law passed this week nor a single answer to every financial loss caused by an agent. ActuIA · EUR-Lex, legal text reference — Analysis on 20 October 2026; directive of 2024 US federal procurement: examine the data clausesThe GSA LLM clause analysed by ActuIA addresses matters including the use of federal data for training and incident reporting. It also includes obligations for European vendors. Its scope is the relevant GSA procurements: these requirements do not become a general rule for all private contracts. ActuIA · GSA, memorandum and annexes — 22 October 2026 Anthropic proposes legislationAnthropic advocates federal rules on catastrophic risks. Its draft remains a proposal, with no force of law. TechCrunch — 20 October 2026 My perspective I would start every regulatory note with two questions: which text, and who does it apply to? They prevent a consultation from becoming an obligation or a partial standard from becoming complete certification. The points in this section still relying on secondary sources need cross-checking before the distribution version. |
BACK TO CONTENTS ↑ |
| 09 | Policy5 ITEMS · SOURCES AND DATES INCLUDED |
Access to models is also shaped by governments and trade policy. Washington removes DeepSeek from federal civilian systemsAn official Department of Homeland Security notice requires DeepSeek products to be excluded and removed from federal civilian agencies’ procurement and IT systems. The scope includes third-party services built on the technology. This concerns the US federal administration; it is not a general DeepSeek ban in Europe. Federal Register / DHS — 20 October 2026 Japan pairs the opening of its agents with rules for useJapan’s Digital Agency accompanies Gennai with guidance limiting agents’ permissions, requiring human approval for important operations and providing activity logs. The programme remains experimental for fiscal year 2026. The administration retains responsibility for the work: granting access does not transfer that responsibility to the tool. Japan Digital Agency, guidance · ActuIA, French analysis — 20 October 2026; analysis on 21 Chinese exports: distinguish the model from the serviceChina’s draft catalogue notably covers 23 AI-related technology categories. It excludes technologies already published in the public domain; ActuIA’s analysis also distinguishes API access from transfers of restricted technologies. The original announcement dates from 16 October. Consultation runs until 15 November: final conditions are not settled. MOFCOM, annonce officielle · ActuIA — Announcement on 16 October; analysis on 20 API access remains exposed to export decisionsAn ActuIA analysis revisits US threats of export controls and precedents affecting access to certain models. This should not be read as a new general cutoff for European customers. For an organisation, the uncertainty is a reason to document dependencies and alternatives. ActuIA — 21 October 2026, updated on 22 Beijing classifies commercial agents by useA notice from nine Chinese administrations organises technological development in commerce. It follows a tiered approach to agents, with stronger requirements in sensitive sectors and self-assessments for lower-risk uses. This week’s analysis highlights a limit: the text does not yet clearly settle the regime for every commercial activity. ActuIA — Notice on 19 October; analysis on 22 My perspective Sovereignty has a very practical meaning: knowing what stops if access disappears tomorrow. I prefer an identified dependency with a realistic recovery plan to a promise of independence that fails at the first contract change. |
BACK TO CONTENTS ↑ |
| 10 | Creative and content professions2 ITEMS · SOURCES AND DATES INCLUDED |
Tools that change how content is prepared and discovered. What still needs direction, editing and attribution. Canva Craft plans the composition before producing the imageCanva’s first image model separates a planning stage from a rendering stage. The vendor targets compositions with text and editable elements, at resolutions up to 2K. For a communications team, its value will depend on actually revising a visual, respecting rights and keeping it consistent with the brand’s guidelines. Canva — 22 October 2026 Spotify changes the entry pointSpotify reiterates its interest in access through agents. Its Muse integration, announced in September, already supports conversational requests for playback, search or a playlist. The change concerns access to the catalogue; it does not, on its own, demonstrate better remuneration for artists. TechCrunch, news on 22 October · Spotify, integration features — News on 22 October; integration announced on 23 September My perspective For a communications team, a trainer or an independent creator, I would first look at what remains editable and attributable. A faster draft can help; art direction and verification time still need to be planned. The professional also decides what the tool must leave open to revision. |
BACK TO CONTENTS ↑ |
11 / THIS WEEK’S TUTORIAL Define an agent’s brief before connecting it25 minutes · One real task, a sheet of paper or a shared document. No code. |
The example: You want an assistant to prepare Monday’s sales review from a tracking folder. It may read authorised documents and prepare a summary. It must not contact customers, update the CRM or approve an expense without your agreement. 01Choose a verifiable outcome.Write down what must exist at the end: “a one-page summary of accounts to follow up, with sources”. Avoid vague aims such as “manage my business”. Start with a short task. 02Set the boundaries of accessible data.Name the relevant folders or documents. Specify those outside the scope. Start with demonstration data or a redacted copy if the contents are sensitive; do not connect the entire mailbox for convenience. 03Separate reading, preparing and acting.Make three columns. Under “read”, list information the agent can consult. Under “prepare”, list the expected drafts and proposals. Under “act”, list possible edits, sends, deletions and commitments. Leave that last column empty for the first trial. 04Write down approvals and stopping conditions.State who approves an action, on the basis of which visible proposal, and when the agent must stop. Missing data, conflicting sources or unexpected access should lead to a question, not a silent initiative. 05Test an ordinary case and a difficult one.Try a complete file, then an ambiguous or incomplete one. Check that the agent cites sources and respects the stopping rule. If it invents information to finish or expands its access, revise the brief before continuing. 06Keep a record and a human fallback.Retain the request, documents used, output and your decision. Name the person reviewing exceptions and know how to revoke access. For sensitive uses, have the relevant business or security lead review the sheet. Your delegation sheetTask: …
Expected outcome: …
Accessible data: …
Authorised actions: …
Actions requiring approval: …
Stopping conditions: …
Decision-maker: …
Records retained and access revocation: … This sheet prepares a first trial. It does not replace analysis of the organisation’s own risks or the authorisations and controls needed for the data and actions involved. |
Over to you.I have no new reader feedback to share in this first version. This week’s question follows on from the last issue: which action would you let an agent perform without approval, and which must remain in your hands? One example is enough: preparing minutes, classifying a request, suggesting a follow-up or triggering a payment. Your answers will help choose the next practical cases. |
The thread running through it allAn agent starts to become useful when its brief can be understood by someone other than the person who configured it. This week, access is advancing quickly: to accounts, documents, work tools and even services where an agent can buy or call. Decisions and responsibility move at a different pace. They remain attached to people and rules that the interface does not always show. For a team, the first job is therefore not to connect everything. It is to choose a task, make approval visible and organise exceptions. Rollout announcements and benchmarks offer leads. A concrete case shows whether they hold up. If you run a trial this week, keep the tutorial sheet and note where the agent asked you to take over. That will often be the most useful information for what comes next. See you next Sunday.
Jérôme |
Let’s shape the next issue together Was this issue useful?One click is enough. I read every reply. Very useful ↗Useful, but too long ↗Not for me ↗ |
Who would find this useful?Once published, share it with someone who supports teams or is considering AI in their work. Subscribe to SIGNAL AI ↗ |
SIGNAL AI #009 · Jérôme Denis · JDENIS Consulting
7 Bis Chemin de la Butte, 31400 Toulouse, France.
You receive this newsletter because you subscribed at jaydenis.com.
Unsubscribe · Reply to Jérôme · Read this issue online · Back to the top ↑ |
|