Briefings de terrain pour les opérateurs de l’IA.
Notes d’exploitation sur la gouvernance de l’IA, les systèmes agentiques, l’architecture d’entreprise, la réglementation et le risque d’exécution. Écrites depuis le terrain. Publiées quand c’est défendable.
-
Visa Agentic Commerce Benchmark: 54% of failed AI shopping runs still bought.
Visa graded 22 AI models on 15 shopping tasks. Walmart says its returns process applies when an agent errs. Shopify reports AI orders at 3.3x. We work out what a retailer measures before the channel grows. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
Microsoft pushed delayed Azure GPU capacity into 2027. Quota does not reserve yours.
Microsoft disclosed delayed third-party GPU capacity and demand above supply. IREN filed the delivery terms Microsoft holds. We read what protects an enterprise workload on Azure until the capacity lands. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
Kinsale and Hanover bet on AI triage as E&S premium per filing falls 12%.
Stamping office data show E&S carriers filing more work for less premium per unit. We read Kinsale and Hanover third-quarter AI claims and name the metric that settles them. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
Pennymac has 22 AI agents live. Its production unit cost still rose.
Pennymac says AI is uncoupling its platform from headcount. Its own table shows production expense up on every volume base. We show what would settle the claim, and when. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
FICO cuts about 15% of positions and cites AI‑driven product development. HubSpot cuts nearly 660 roles and says AI‑related efficiencies are not the driver.
On October 6, 2026, FICO and HubSpot each filed an 8‑K on a workforce reduction built on fewer organizational layers. FICO cites AI‑driven product development and states no saving; HubSpot says the cut is not driven by AI‑related efficiencies while it reorganizes product around outcomes it delivers with AI. We compute the charge per role, read the Challenger September count, and name the test a board applies to either label. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
Microsoft 365 Copilot Business bills agents by usage on December 1. The default Copilot Credits cap exceeds the license price.
From December 1, 2026, new Microsoft 365 Copilot Business licenses sold through CSP default to pay-as-you-go Copilot Credits, capped at 4,000 credits per user per month. At Microsoft's $0.01 rate that is $40, against a $21 list seat. We price the cap, read the defaults behind it and name what a CFO and a CIO decide before December 1. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
C.H. Robinson prices RXO on Lean AI. Its own filings show headcount fell faster than cost.
C.H. Robinson agreed on October 5, 2026 to buy RXO for $5.8 billion, priced on $300 million of Lean AI synergies. Its own filings show headcount fell far faster than cash cost. We recompute what the price asks of RXO and name the metric to track. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
Amazon blocked Meta's Muse with its Conditions of Use. The Ninth Circuit pointed it there.
Amazon began blocking Meta's Muse agent on September 20, 2026, citing its Conditions of Use. We read the Agent Terms, the August 4 Ninth Circuit ruling in Amazon v. Perplexity, and the decision every enterprise with a customer login now faces. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
Paychex counts 2,000 AI agents. Its 10-Q attributes no savings to them.
Paychex reported more than 2,000 AI agents and features on September 23, 2026. Its 10-Q attributes no savings to them. We rebuild the 1% expense growth line by line and name the metric a board should track. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
Salesforce projects 3x to 4x ARR from wall-to-wall Agentforce customers. Build your AI case on that bill.
Salesforce's September 16, 2026 investor day projects 3x to 4x ARR from Agentforce customers. We turn that into a support-labor break-even and a contract. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
Astrana Health credits AI with a 210 bps G&A gain. Earn-out accounting explains 76 bps.
Astrana Health's September 15, 2026 deck credits AI with a 210 bps G&A gain. We rebuild it from the 10-Q: earn-outs, Prospect, full-risk revenue. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
Oracle's $664 billion backlog is a third‑party risk input for banks on OCI.
Oracle reported $664 billion of RPO on September 10, 2026. We read its filings against bank third‑party risk guidance: backlog timing, prepayments, leases. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
Accenture is staffing Gemini Enterprise with 1,000 forward deployed engineers. It employs 799,000.
On September 8, 2026 Accenture and Google Cloud announced the Accenture Gemini Enterprise Business Group, which will establish a 1,000-person forward deployed engineer workforce built on Accenture's existing 50,000 Google Cloud-skilled professionals. That is 2 percent of the practice and 0.13 percent of the firm. Accenture stopped publishing its advanced AI bookings and revenue after the quarter reported on December 18, 2025, so no vendor series exists against which to check whether the embedded engineering model converts into delivered work. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
Anthropic is moving Claude data retention into your cloud account. The misuse flags come with it.
On June 9, 2026 Anthropic made 30-day retention mandatory for Covered Models and unavailable under zero data retention. On September 1 it announced Enterprise Frontier Safeguards, which moves that window into the customer's own cloud account under the customer's keys and audit log. The platform documentation carries a second sentence, in a section called Retention regardless of arrangement: even under ZDR or a HIPAA arrangement, a flagged session can be retained for up to 2 years. The announcement does not say whether that exception survives the move. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
Nvidia booked $96 billion. Most US AI adopters stop at 3 business functions.
Nvidia's 10-Q for the quarter ended July 26, 2026 discloses one direct customer at 16 percent of revenue and 5 direct customers holding 70 percent of accounts receivable. The US Census Bureau, measuring November 2025 to January 2026, found 18 percent of US firms using AI in a business function, and 57 percent of those adopters running it in 3 functions or fewer. The variable their regression associates with commercial performance is breadth, which is the variable most adopters have not moved. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
Cboe redefined non-display usage to cover LLMs. Your research agent reads at a different price.
On April 1, 2026 Cboe amended the definition of Non-Display Usage in its North American Data Policies so that machine access counts unless it is solely in support of display for a natural person. The filing says plainly that the change targets firms leveraging Large Language Models. 4 months later Google Cloud shipped a research agent for capital markets with connectors to LSEG, S&P Global and FactSet. Entitlement controls answer who may see the data. They do not answer what a machine may do with it. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
Claudeforce puts Salesforce behind Claude. What remains is the permission model.
On August 26, 2026 Salesforce and Anthropic announced Claudeforce. The first product, Salesforce in Claude, ships 37 prebuilt sales skills that read live revenue context and take governed action from inside Claude, in pilot now with an open beta expected in September. The announcement moves the value of the CRM from the interface to the permission model, and that reading applies to every company that lets agents act, not only to Salesforce customers. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
The Linux Foundation now governs TRACE. Your agent audit log is not evidence.
On August 25, 2026 the Linux Foundation welcomed TRACE, an open specification for hardware-attested AI agent governance records contributed by OPAQUE and backed by AMD, Intel, Microsoft and TII. The specification is at v0.2. Its own limitations file states that the software-only tier does not satisfy EU AI Act tamper-evident logging or DORA. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
OpenAI cut GPT-5.6 prices 80 percent. The frontier tier did not move.
In late July 2026 OpenAI cut GPT-5.6 Luna by 80 percent, Terra by 20 percent, and left the flagship Sol at $5 and $30 per million tokens. The gap between the cheapest and the most expensive tier widened from 5x to 25x. Canva and Wix both published what that costs. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
Synchrony put promotional financing inside ChatGPT. Regulation Z came with it.
On August 17, 2026 Synchrony put promotional financing offers inside ChatGPT. Promotional financing is 35.3 percent of a $103.8 billion book, deferred interest alone 17.7 percent. Regulation Z attaches to the advertisement, and its exemptions name envelopes and banner ads, not generated answers. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
The US does not have an AI Act. It has regulators.
No federal statute governs enterprise AI in the United States. A voluntary NIST framework, a Colorado law rewritten twice, 5 sector regulators, an FTC accuracy policy, and the EU AI Act's extraterritorial reach all apply to the same enterprise at once, on 7 different clocks. Franck Ohrel · Founder & CEO · 10 min read Lire le briefing -
Google signed 1 gigawatt of demand response. AI is cut first.
Google has signed 1 gigawatt of data center demand response and made machine learning workloads the first thing it cuts. On August 17 the 6 US grid operators file the rules that turn conditional service into a tariff category. Which of your AI workloads may pause is becoming a procurement question. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
The regulator now reads the trial while it runs.
2 sponsors already send endpoints and safety signals to the FDA while their studies run. Selection for the broader pilot closes this month. The prize is cycle time. The constraint is whether your evidence is stable enough to be read while it is still moving. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
The cheapest integration is now a logged-in browser tab.
From August 3, an agent drives desktop Chrome signed in as your employee. It automates the portals integration never paid for. It also acts under a person's identity. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
AWS deprecated Bedrock Agents without a deadline.
AWS moved Bedrock Agents to maintenance mode on July 30 and froze its model catalog. Existing agents keep running. That is precisely why nobody escalates. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
The carriers bought the workflow, not the model.
A top-five insurer scaled AI subrogation in the second quarter. The vendor discloses its gain. The carrier discloses nothing and answers the examiner alone. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
The audit trail is the product.
Brussels moved the calendar on the EU AI Act. It did not touch what Articles 12, 19 and 26 require a governed AI system to record. Franck Ohrel · Founder & CEO · 11 min read Lire le briefing -
The channel that converts best cannot read your site.
More than half of all web requests are now automated. AI-referred shoppers convert better than every other channel, and up to 46 percent of retailer sites cannot be read by machines. The public surface became an interface, and nobody on the org chart owns it. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
When every employee has an agent, routing is the decision.
Cisco is giving 90,000 employees an AI agent that routes to the cheapest sufficient model, on premise. The decision was never the model. It is the layer that routes, prices, and records every model you run. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
The banks quantified their AI gains. The margin will not follow.
In Q2 2026 earnings, US banks proved AI saves real hours. Jamie Dimon told the market why that does not become margin. The operator question boards should ask. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
AI value is moving from the model to the operating layer.
2 AI leaders spun up implementation businesses in one quarter. The value is not in the model. It is in the layer that turns a model into governed production work. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
Can Europe still win the AI war?
The continent that invented modern science risks becoming the regulator of technologies built elsewhere. A signed essay on power, dependence and the narrowing window. Franck Ohrel · Founder & CEO · 18 min read Lire le briefing -
Your model is now a cyber capability.
The EU Action Plan on Cybersecurity and AI moves the model itself inside the security perimeter. What it asks of operators in finance, energy and health. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
The deadline moved. The exposure did not.
The EU AI Act high‑risk deadline slid from August 2026 to December 2027. What still bites in August, and why the risk never moved with the date. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
What actually lands on August 2.
The Digital Omnibus moved the high‑risk deadline to December 2027. The enforcement machine goes live on schedule. The board question that follows. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
The agent no one owns.
Ask each agent in production who owns it and who can stop it. In most enterprises, no one can answer. The inventory is the first control. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
Govern the action, not the model.
Model choice is commoditized. The risk lives at the execution layer, where agents call tools and take action. NIST is standardizing exactly that, and your board will ask for the trace. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
The fox in the henhouse paradox.
When you deploy OpenAI, Anthropic, or any other model provider at the core of your operations, who captures the value of that usage? The board-level question that decides whether AI builds you a margin or absorbs it. Franck Ohrel · Founder & CEO · May 17, 2026 · 5 min read Lire le briefing -
Silence on AI is a decision.
Why strategic silence is becoming a governance failure for executive teams, and the 3 conversations no board can afford to keep postponing. Franck Ohrel · Founder & CEO · 4 min read Lire le briefing -
AI does not replace thinkers. It amplifies builders.
From Brynjolfsson to Musk, the converging frame on AI: a test of sovereignty, not a technology question. Franck Ohrel · Founder & CEO · 6 min read Lire le briefing -
Accenture, Copilot, and the consulting paradox.
Accenture’s clients run the same Copilot as Accenture. Consulting does not disappear. Its middle layer does. Franck Ohrel · Founder & CEO · 5 min read Lire le briefing -
AI governance is an architecture problem. White paper, in French.
20 pages on why generative AI projects fail, and what the few that succeed build first. WP-HB-2026-00426 · FR. Ouvrir le livre blanc
Axes de recherche
6 angles sous lesquels nous publions.
Choisissez un axe pour filtrer les briefings ci-dessus.
Les piliers de fond derrière ces angles
À propos de la Newsroom
La newsroom Hikari Blue publie nos briefings de terrain sur la gouvernance de l’IA, les systèmes agentiques, l’architecture d’entreprise, la réglementation et le risque d’exécution.
Ce ne sont pas des articles de tendance. Ce sont des briefings de terrain pour les dirigeants qui doivent comprendre ce que l’IA change, ce qu’elle expose, et ce qu’il faut construire avant de pouvoir faire confiance à l’autonomie.
Publié par l’équipe opérationnelle de Hikari Blue. Écrit depuis le terrain, pas depuis le fil d’actualité.
- 42Briefings de terrain et livres blancs publiés
- 06Axes de recherche
- 100%Écrit par des opérateurs
Restez au contact du terrain
Recevez le prochain briefing de terrain.
Écrit pour les dirigeants, fondateurs, DSI et administrateurs. Sans cadence fixe : publié quand c’est défendable. Le signal, rien d’autre.
Lu par des dirigeants des services financiers · du luxe · de la santé · de l’énergie
Demandes de presse : écrivez directement à l’équipe · réponse d’un associé désigné, pas d’une agence
Vous apportez un processus plutôt qu’une histoire ? Réserver un appel de 30 min · un associé désigné répond rapidement, par écrit.
