Latest briefing
July 5, 2026 · 5 actus (site) · 5 actus (base)
Imagine 6,000 plumbers specialized in AI, ready to sit in your client's offices and install everything. That is exactly what Microsoft just launched, with a $2.5 billion envelope, to help big companies move from AI experiments to actual production. Until now, many CEOs were spend...
🔥 Top story
01
Microsoft creates a 6,000-strong team of 'AI plumbers' to install AI in big companies - $2.5 billion on the table
Imagine 6,000 plumbers specialized in AI, ready to sit in your client's offices and install everything. That is exactly what Microsoft just launched, with a $2.5 billion envelope, to help big companies move from AI experiments to actual production. Until now, many CEOs were spending millions on AI software without ever really using it - like buying a race car that stays in the garage. So Microsoft sends its engineers on site, at the client, to wire up the tools, train the staff, and bill by the outcome obtained - not by hours spent. In the meantime, Amazon ($1 billion), Anthropic ($1.5 billion) and OpenAI ($4 billion) have made the same bet in recent months: if AI models are ready, the real problem is now knowing how to deploy them. The frontier-lab battle has shifted from the lab to the field.
02
The world's largest machine learning conference opens in Seoul with a record 23,918 papers submitted - and one topic is exploding
In Seoul, from July 6 to 11, 11,000 researchers from around the world are gathering for ICML 2026, the largest scientific conference on artificial intelligence. They received twice as many paper proposals this year as in 2025 - an all-time record. But the most striking thing is the hot topic: 'AI agents', these software programs capable of acting on their own, like an assistant that books your vacation without asking you first. Sixty of the 247 planned workshops are dedicated to it. At the same time, the conference discovered that hundreds of researchers had used ChatGPT to write their reviews illicitly - they were demoted and nearly 500 papers were rejected without appeal. Integrity control is becoming as serious an issue as the topics themselves.
03
Mark Zuckerberg, boss of Meta (Facebook, Instagram, WhatsApp), admits his AI agents are 'behind schedule' despite $145 billion of investment
Mark Zuckerberg, the boss of Facebook, told his employees on July 2: his company's AI agents 'have not accelerated as expected' for the past four months, despite the $145 billion Meta has pledged to invest in AI in 2026 - and despite firing 8,000 people in May, reassigned to these same projects. The market punished the stock: Meta shares dropped 5% that day. This is a rare signal: when the boss of the world's second-largest AI investor admits his agents are not up to par, it means the on-the-ground reality is far harder than the slick conference demos. And it gives weight to the opposite bet Microsoft made the same day: instead of betting on the most expensive models, send 6,000 humans to deploy on client sites. For anyone betting on AI in 2026, the problem is no longer the engine - it is the integration.
04
A New York start-up raises $22 million to build 'analyst agents' that learn the style of each investment manager
Imagine a new hire who watches how you work for two weeks, then does the same work in your place, in your name. That is exactly what LinqAlpha, a New York start-up founded by former Goldman Sachs and MIT people, is offering - after raising $22 million on July 2. Its specialty: AI agents that learn the method of each financial analyst, then scan thousands of market signals (company results, statistics, social media) to spot the ones that will move prices before anyone else. Already 70 major finance players use it, including Causeway Capital and Schonfeld, which together manage more than $5,000 billion in assets. The model: AI does not replace the expert, it learns the expert's style and saves a lot of time. For executives, it is proof that specialist 'agentic' tools per sector are taking over from general-purpose agents.
05
Argentina proposes the world's first law allowing an AI to be a 'chief executive officer' of a company
Imagine a French company whose CEO is an AI program, with no mandatory human shareholder, capable of owning assets, signing contracts and paying taxes. That is exactly what Argentina wants to authorize with a bill submitted to Congress at the end of April 2026 by libertarian President Javier Milei, and which is resurfacing in the specialized press this weekend. The text creates a new legal category - 'Sociedad de Inteligencia Artificial' - without regulating AI itself, and with attractive taxation. A prototype already lives: a tokenized agent called $SAIRI, created by Argentine entrepreneur Santiago Siri, which signs contracts on the blockchain and does decentralized finance. If Argentina passes the law, other low-tax jurisdictions (Dubai, Estonia, Singapore) could copy it. For everyone building commercial agents, a new question arises: who is responsible when an agent signs a contract in your place and something goes wrong?
📡 To watch
Will Microsoft succeed in billing by outcome rather than by hours?
Microsoft is promising an economic model where the client pays only if the agent delivers the expected effect. If the first contracts work, this could shift how all enterprise AI tools are priced. To watch: first signatories, success rate, impact on Azure margin in the coming quarters.
Researchers in Seoul, regulators in Geneva: will the gap between science and governance keep growing?
On July 6, ICML 2026 opens in Seoul while the UN hosts its first global dialogue on AI governance in Geneva. The researchers producing the state of the art and the states wanting to regulate it cross paths on the same day, in two time zones. To watch: joint statements, convergence or open fracture.
Zuckerberg set his own clock: 3 to 6 months to deliver 'meaningful returns' in AI
The Meta boss has publicly committed to a deadline. If in three to six months he has not delivered, it will be a bearish market signal on agent maturity. To watch: Q3 and Q4 2026 earnings calls, new restructurings, key departures in Meta's AI teams.
Is finance opening the door to 'specialist agents per profession'?
With LinqAlpha, finance proves that an agent specialized per profession is worth more than a general-purpose agent. The real question is which sector follows next: health, legal, R&D? To watch: next funding rounds in the same category, acquisitions by generalist platforms (Microsoft, Salesforce).
📊 Trend
July 5, 2026 confirms the agentic AI battle is no longer being fought in laboratories but on the ground. Microsoft sends 6,000 humans to deploy its agents at client sites with a $2.5 billion budget; the Meta boss publicly admits his own agents are falling behind; scientific research quadruples the number of workshops dedicated to agents; and finance proves that a profession-specialized agent can be funded seriously. Three lessons emerge for anyone building with AI: (1) an agent that works in production is no longer a conference prototype, it is an integration project that needs human skills; (2) sector specialization is overtaking generalist agents; (3) the legal framework is starting to catch up - Argentina could become the first country to legally authorize a company led by an AI.