noticias radar ia agentes de ia automatizacion pymes

The Thursday radar was about the price of AI; today's is about its fragility. Yesterday afternoon, in the middle of the Spanish workday, ChatGPT, Claude, and Grok went down simultaneously, dragging along the tools built on top of them. This happened just as AMETIC finally put a number on the Spanish gap—only one in six SMEs has taken AI beyond pilot stages—and Meta acknowledged that measuring employees by consumed tokens measured nothing at all. In the money chapter, a spin-off from the Polytechnic University of Valencia raised $125 million with Nvidia participating in the round, and the U.S. government wrote to a judge stating that training models with protected material constitutes fair use.

ChatGPT, Claude, and Grok go down at once: having two providers is not having a Plan B

Yesterday, around 5:00 PM in mainland Spain, OpenAI acknowledged "elevated errors in ChatGPT and Codex" that broke conversations, logins, file uploads, voice, search, and image generation; Anthropic declared a partial outage of Claude.ai, the API, Claude Code, and Claude Cowork; xAI confirmed issues with Grok; and Cursor went down as a result of the previous two. Downdetector surpassed 30,000 reports in a single hour, with more than 37,000 accumulated for ChatGPT, 1,365 for Grok, and 1,324 for Claude. Forty-two minutes after the first alert, OpenAI had applied a fix and Anthropic was recovering all its models except Opus 4.8 and Opus 5. No one has confirmed the cause; the suspect cited by everyone is Microsoft Azure, on which all three depend, although its status dashboard reported no incident. Gemini had a spike in user reports, but Google did not confirm an outage. For your company: today, make a list of workflows that stop if the AI doesn't respond, mark which ones have a manual alternative, and calculate what one hour of downtime costs for each; that sheet is your continuity plan, and it can be done in an afternoon. Second, hiring two providers isn't redundancy if both run on the same cloud: ask each one in writing which cloud and region the inference is executed in, because yesterday the shared failure wasn't in the models. And third, the cheap solution that actually works: a retry queue and a notification to a human instead of an error returned to the client. A process that degrades gracefully—gets delayed—is a true Level 3; one that bursts outward remains a Level 1 with an API in front of it. Source

Only one in six Spanish SMEs has integrated AI into its processes: the rest are still in pilots

At the 40th AMETIC Meeting in Santander—the same one that produced the AI Atlas we discussed on Tuesday—Enrique Serrano, president of the employer association's AI and Data Economy Commission, provided the missing number: only one in six SMEs has truly incorporated artificial intelligence into its processes beyond testing. The surrounding figures don't contradict each other; they complement: Carla Redondo, Director General for Digitalization Services, said that 44% of the working population already uses AI and more than 22% of companies have incorporated it, while the Secretary of State for Industry placed industry at 17% and startups at 100%. CEPYME recalled that Spain has lost 25,000 micro-enterprises since the pandemic. And there was a result metric, not just intention: an agentic AI project at the Social Security office has cut delivery times by 30% and freed up 14,600 hours of work. For your company: the gap is no longer between who uses AI and who doesn't, because nearly half your staff uses it whether you ask them to or not; it's between individual use and process integration. That jump is exactly what separates Level 1 from Level 2, and it's not achieved by buying another tool: it's achieved by choosing a process with volume and automatic verification of the result—invoices that match, fields that are or aren't there, orders that match the delivery note—and measuring how many go through without human review. If you can't verify the result without a person, automate the draft and not the loop. And use the Social Security case as a yardstick: 14,600 hours is a process figure, not a "happy staff" figure. Source

Meta stops scoring employees by token usage: the metric was teaching them to waste tokens

Meta internally communicated this week that it will no longer evaluate engineer performance based on how much AI they use: "we will not use AI adoption dashboards or token counts to evaluate impact." Less than a year ago, the company had said the exact opposite—that it would grade by "AI-driven impact"—and handed out internal labels that today sound like a joke: AI Native, AI First, AI Enabled. What resulted is internally called "tokenmaxxing": employees launching repeated or outright frivolous requests to inflate the counter, and even an internal leaderboard set up by a worker. The contradiction is that while removing the metric, the company is rolling out Hatch, an experimental agent that navigates, manages applications, and completes multi-step tasks without supervision; consumption continues to rise, and some employees are refusing to connect it to their personal email and calendars. For your company: if you measure adoption, you get adoption, just as if you measure an agent incorrectly, you teach it to cheat. Today, replace any goal like "everyone should use AI" with a metric from the process itself: orders processed without correction, emails answered without rewriting, hours from the person who previously did that task. And take note of the uncomfortable part: when the agent starts managing applications from the employee's computer, it must use a business account and work credentials, never personal ones. It's not a lack of trust in the worker; it's that you can't shut down a personal account at three in the morning. Source

125 million for a Valencia spin-off with Nvidia involved, and a 5 billion valuation for agent orchestration

iPronics, a spin-off from the Polytechnic University of Valencia founded in 2019, closed a $125 million Series B yesterday, bringing its total raised to $177 million. It is co-led by Maverick Silicon and Light Street Capital, with participation from Nvidia, Bosch Ventures, the European Innovation Council Fund, and Criteria Venture Tech (La Caixa's venture capital vehicle). iPronics sells iPronics ONE, a rack-mounted programmable optical layer that reconfigures the connectivity of an AI cluster in real-time. A contrast on the same day, to gauge where the money is: the Dutch-Israeli company Wonderful raised $550 million at a $5 billion valuation—double that of March—led by Insight Partners with Salesforce joining as an investor, to sell what it calls an AI operating system that orchestrates agents over existing company systems. For your company, two takeaways. The first is that AI capital does land in Spain when there is deep tech behind it, and that the industry bottleneck is no longer the model but the network connecting the cards. The second is more useful for your budget: what is valued at 5 billion is not the model, but the layer that coordinates agents, flows, and permissions on top of your ERP and CRM. If your automation project has turned into a discussion about which model to choose, you are fighting over the cheap and replaceable part. And use a specific date as leverage: the European Data Act prohibits charging to switch clouds starting January 12, 2027, so ask your providers in writing today to commit to exporting your flows, instructions, and connectors in a format you can use elsewhere. Source · Source

The U.S. Government writes to the judge: training with protected material is fair use

The Department of Justice filed a statement of interest on Tuesday before Judge Sidney Stein of the Southern District of New York in the lawsuit by the New York Times against OpenAI and Microsoft. This is the first time the federal government has taken a stance in the wave of lawsuits filed by authors, publishers, record labels, and media against AI developers. The brief argues that training sufficiently transforms the source material and therefore fits under fair use, and that U.S. leadership in AI is a national security interest whose benefits outweigh any competitive harm. It is not binding: Judge Stein decides, and summary judgment motions are due today. For your company: nothing changes here, and that is precisely the point worth understanding. American fair use does not exist in the European Union, where the text and data mining exception with reservation of rights applies. Therefore, a favorable ruling for OpenAI in New York does not shield your provider in Madrid, nor does it affect the lawsuit Sony Music and Warner Chappell filed against Anthropic a week ago. The takeaway is the balancing lesson: the origin of training data is a liability for your provider that can reach your contract. Look for the IP indemnity clause for model outputs and its cap—it usually doesn't exist in free plans—keep a record of the model, prompt, and date for everything released under your brand, and have the same workflow tested on a second model in case you need to turn off the first one one day. Source

What to watch tomorrow?

Two specific things. Whether OpenAI, Anthropic, or Microsoft publish a root cause analysis of yesterday's outage: the answer will tell us if the single point of failure was the shared cloud or something worse—a common dependency further down the chain. And Judge Stein, who receives summary judgment requests for the New York Times case today with the U.S. Government already positioned in writing.

Watch the 1-minute video