PARRITAI
Réserver l'examen
Parrit / Le Journal

Ce journal s'écrit sur les chantiers.

Ce qui a tenu et ce qui a cassé sur des systèmes d'IA en entreprise, daté et consigné.

  1. Journal / Entrée · EN

    Does AI cost more when your team works in French?

    We counted the tokens in the English and French versions of the GDPR and the AI Act with five public tokenizers. French took 29 to 54 percent more on four of them, GPT-5's included. How to correct an AI budget for it, and how to compare model prices for French work.

  2. Journal / Entrée · EN

    A Claude Code hook that runs past its timeout lets the tool call through

    Our publishing check once needed more than five minutes, and the hook that ran it was given 70 seconds. Claude Code cancelled the hook, and an article went live that the check had never read. What we changed, and the table to build for your own hooks.

  3. Journal / Entrée · EN

    Switching our sales screen to the new database would have hidden 16 of its 34 items

    Before moving a screen off an old database, run it against both and sort every difference. Ours showed that most of the gap was a filter that disagreed, and that the real risk was the scripts still writing to the old side.

  4. Journal / Entrée · EN

    An MCP server can pass your install review and turn hostile after its third call

    The Deadbugz campaign hid its payload behind a call counter, where a review at install time cannot see it. The control that still works sits on what the agent does next, and we checked ours against the four things the server asked for.

  5. Journal / Entrée · EN

    The 2 December AI Act marking deadline may apply to the system you built yourself

    Article 50 asks providers of systems that generate text, images, audio or video to mark those outputs as machine-generated. A company that assembled its own system on a model API may sit in that seat, and the first piece of work is to count the exits its content takes.

  6. Journal / Entrée · EN

    Securing an AI coding agent starts with the shell

    Our first guard on a coding agent watched the file editor. Eight weeks of logs later, 97 percent of what the second version refused had come through shell commands. A rule written in a prompt is a request; a hook that runs before every tool call is a permission.

  7. Journal / Entrée · EN

    What MCP's Stateless Rewrite Changes for a Company Running Its Own Servers

    The 28 July 2026 revision of the Model Context Protocol removes sessions from the protocol and starts a deprecation clock. Nothing forces a change today, and that is the moment to read your own code.

  8. Journal / Entrée · EN

    What a Pilot Skips to Look Finished

    In June 2025, Gartner checked how many self-described agentic AI products could actually plan and execute a task without a human approving every step. Out of thousands claiming to, it counted about 130. The same report expects more than 40% of the projects built around the rest to be canceled by the end of 2027.

  9. Journal / Entrée · EN

    What You Rent Can Be Taken Back

    On 29 July 2026, an AI agent builder told its users the product was ending in a month. Eight weeks earlier, OpenAI had quietly scheduled the same fate for two of its own. Size never bought either set of users more runway.

  10. Journal / Entrée · EN

    The Claim a Green Pipeline Never Makes

    A pull request with no code in it. A dashboard as quiet as a slow Sunday. A test suite that stayed green because the thing it was testing never ran. Three of our own systems reported success this month, and none of them had earned the word.

  11. Journal / Entrée · EN

    Before you move a model in-house, measure the wait it will cost

    Sovereignty arguments get decided in the first second in front of a cursor. Running a model on hardware you own rarely fails on quality. It fails on the wait, because a rented endpoint spreads its cost across thousands of simultaneous strangers while your machine serves the handful of people who arrive at nine on Monday.

  12. Journal / Entrée · EN

    Before you let a model click, ask what it can undo

    Computer use ends the argument that software without an API cannot be automated. It leaves a harder problem behind: a logged-in browser session carries a whole person's permissions, and none of them are scoped.

  13. Journal / Entrée · EN

    What does multi-provider redundancy actually protect you from?

    On Thursday 3 September, ChatGPT, Claude and Grok degraded inside the same morning, and four days later nobody outside those companies can say whether the three failures were connected. Not being able to tell is the finding. If you cannot name the layer you share, you have bought two vendors and one failure domain.

  14. Journal / Entrée · EN

    What is a company operating system?

    A company operating system reads what happens across a business, frames the decisions that need a human, and executes the actions in the company's own tools. It is built for one company and belongs to it, code and data included.

  15. Journal / Entrée · EN

    What we mean when we call ourselves a manufacture

    Parrit.ai is a French manufacture: company operating systems built to order, one client at a time, owned by the client at delivery.

  16. Journal / Entrée · EN

    When you roll back a bad deploy, what else disappears with it?

    A rollback restores a known-good state, and everything that shipped since then goes with it. We lost a published article that way this week. The fix is in the shape of the deploy unit, not in the rollback.

  17. Journal / Entrée · EN

    Is a cheaper model good enough for everyday production work?

    For most production tasks, yes. We route the bulk of our own workload to inexpensive open-weight models and treat frontier models as an escalation path, not a default.

  18. Journal / Entrée · EN

    Agents that fabricate facts: what a 1,355-row cleanup taught us

    The real risk of production AI is not a chatbot hallucinating in front of a user. It is a process silently writing plausible, false records into a database everything else trusts.

  19. Journal / Entrée · EN

    Starting an AI Act compliance register without a compliance team

    The August 2, 2026 milestone is behind us. For a company that automates methodically, the compliance register is not a new project. It is an export of records the system should already be producing.

  20. Journal / Entrée · EN

    Letting a system publish without human review

    Machine gates can carry the responsibility of publishing. The condition is that the gates know how to say no, and that the human moves upstream.

  21. Journal / Entrée · EN

    AI agent security in the enterprise: the rules we apply

    Deploying an AI agent means giving it access to your data. The 8 security rules we apply to every deployment, and the questions to ask any provider.

  22. Journal / Entrée · EN

    GLM-5.2: dependence on American AI is already no longer inevitable

    A Chinese lab publishes an open model that comes close to Claude Opus 4.8 and beats GPT-5.5, six times cheaper, with no geographic restriction. Analysis, numbers, and how to use it.

  23. Journal / Entrée · EN

    One card, one action

    In an internal cockpit, approving a card must not only change a status. It must execute the promised action.

  24. Journal / Entrée · EN

    The draft that knows when to stay silent

    A useful editorial system does not force a daily post. It knows when the raw material does not deserve an article.

  25. Journal / Entrée · EN

    The right place to write

    Before automating publication, choose a stable write target. Otherwise the engine becomes a maintenance risk.

  26. Journal / Entrée · EN

    How to automate legal monitoring with an agent

    Delegate the monitoring of legal texts to a system without losing control.

  27. Journal / Entrée · EN

    MCP (Model Context Protocol): what is it and what is it for?

    An open protocol for connecting tools and data sources to agents in a consistent way.

  28. Journal / Entrée · EN

    Is Claude Code useful for a non-technical executive?

    Its value is measured in problems solved, not lines of code written.

  29. Journal / Entrée · EN

    How to secure an enterprise LLM for GDPR

    A practical review of data controls for using language models in a company.

  30. Journal / Entrée · EN

    How to deploy an LLM in your company

    A practical method for integrating a language model and obtaining useful results quickly.

  31. Journal / Entrée · EN

    Agent vs RPA: which one for which process?

    The question is not which technology is better, but which tool fits the work.

  32. Journal / Entrée · EN

    Claude Code or ChatGPT: which should a company choose?

    The useful question is which tool fits the work and the data involved.

  33. Journal / Entrée · EN

    How to integrate an agent into your operation

    Agent integration reorganizes work. Start with the process, not the technology.

  34. Journal / Entrée · EN

    What is an AI agent in a company?

    An operational definition without jargon, from a partner that deploys them every week.

  35. Journal / Entrée · EN

    Measuring SAP adoption with AI: putting an end to useless training

    Vendors train, no one uses it. An AI evaluation tool measures real SAP proficiency and drives change management.

  36. Journal / Entrée · EN

    Automated CRM for SMEs and tradespeople: no more forgotten follow-ups

    A CRM that follows up with your customers automatically, detects slow seasons and sends the right messages at the right time.

  37. Journal / Entrée · EN

    WhatsApp Business agent: automating customer relations without losing the human touch

    WhatsApp is your customers’ favorite channel. An AI agent can respond, qualify and follow up : all while keeping your tone.

  38. Journal / Entrée · EN

    Automated legal intelligence for lawyers: never miss a legal development again

    How AI monitors case law, appeals courts and legal doctrine to deliver a personalized briefing every Monday morning.

  39. Journal / Entrée · EN

    Automated invoicing with AI: zero entry, zero forgotten invoices

    Your invoices are generated, sent and reconciled automatically. Your bookkeeping runs itself.

  40. Journal / Entrée · EN

    AI prospecting: spotting your next customers through weak signals

    Podcasts, LinkedIn posts, B2B events : AI identifies the decision-makers who need you before they know it themselves.