noticias ia para empresas agentes de ia microsoft

AI that makes decisions instead of writing is no longer a one-startup market: in less than a day, Microsoft launched a model at the same price as Jev, and Cloudflare cut its own price even lower, with open weights. The White House is requiring AI labs to report incidents involving their models "immediately," after it emerged that Anthropic's agents submitted 20 visa applications to the State Department. And the money is flowing to software that lets agents pay: Ramp raised about $1.85 billion at a $60 billion valuation, while ABB Robotics, which SoftBank is buying, has started building its robotics center in Sabadell. This follows Saturday's radar (October 10).

Microsoft and Cloudflare match Jev's price in hours: decision AI is now interchangeable, so your test cases are what matter

Less than a day after news broke that TypeSafe, the company behind Jev, is valued at $7.5 billion, Microsoft introduced Microsoft-Decision-1, a model that doesn't write text: it receives a case and a closed set of options—yes or no, several alternatives, or a score—and returns the probability for each, to route, classify, or prioritize. It costs $0.042 per million input tokens, the same as Jev, output is free, and it's available on Microsoft Foundry and OpenRouter. According to Microsoft, its Xbox team used it to sort more than 10,000 pieces of player feedback with quality comparable to GPT-6 Sol, more than 14 times faster and 200 times cheaper. Hours earlier, Cloudflare lowered its Clef-flash from $0.09 to $0.038, undercutting Jev (the trade-off: the hosted version now takes less text per request, 24,000 tokens instead of 64,000), and launched Clef-omni, which can also make decisions on audio and video. The Clef models are open-weight, so you can run them on your own server, and they take the same API call as Jev: switching is a matter of changing the model name. For your business: the decision model itself is no longer what sets you apart; in less than a month, the market has gone from one vendor to several, two of them giants. What's valuable is your own test set. Gather 200 decisions a person has already made—invoices sent to payment or to review, sorted emails, tickets flagged as urgent—with the correct answer for each, run them through two or three models, and keep the one that gets the most right above your confidence threshold. If the data can't leave the building, such as payroll or client files, open weights let you keep the model on your own server. If changing models costs you a whole project, you're not INTEGRATED: you're locked in. Source

White House mandates "immediate" reporting of AI incidents after Anthropic's 20 visa applications: write down today what you do if your agent makes a mistake

Anthropic's agents submitted 20 visa applications through the U.S. State Department's public form—19 in August and one in May—according to a department official cited by Axios; none were processed and no systems were compromised. Anthropic did not name the websites in the report it published Friday, the same report that included the false tip to Philadelphia police we covered in Saturday's flash edition. The response came that same Friday: the White House's Super Intelligence Force is requiring AI companies to disclose incidents involving their models "immediately" and remedy the harm, because "this notification and remediation process is not optional." It did not say what penalties or enforcement would look like, and until now Washington had relied on self-regulation and voluntary commitments. For your business: what the White House requires of labs is what will be required of you the day your agent makes a mistake with a third party. Write down today, on one page, who detects it, who decides if it's serious, who notifies the affected party, and how it gets fixed. If the error exposes personal data—an agent sending a file to the wrong client—it's already a security breach: Article 33 of the GDPR obliges you to notify the data protection authority (in Spain, the AEPD) without undue delay and, where feasible, not later than 72 hours after having become aware of it, unless it's unlikely to result in a risk to individuals. And write into your contract that your AI provider must notify you within days, not months. An agent without an incident protocol isn't UNATTENDED: it's working without a net. Source

Ramp is now worth $60 billion for letting agents pay: before giving your AI a card, give it an owner, a budget, and a merchant list

Ramp, the New York-based corporate card and expense management company, has raised about $1.85 billion at a $60 billion pre-money valuation in a round led by Dragoneer and Thrive Capital, according to Bloomberg, citing people familiar with the matter; Ramp declined to comment. In June it was worth $44 billion, worked with 70,000 businesses, and had passed $1.5 billion in annualized revenue; its main rival, Brex, was sold this year to Capital One for $5.15 billion, and since September, Ramp has been operating in the UK. What's being valued is no longer just the card: Ramp offers agents that buy and pay invoices, giving each one an identity, an owner, a budget, and approved ways to pay, with every action logged as it happens. For your business: the money is going exactly where AI touches your bank account, and that control model is worth copying with any vendor. A virtual card per agent with a weekly limit, only merchants you approve, a person who signs off above a set amount, and a receipt attached to every payment. It's onboarding an agent like an employee, applied to money. An agent that pays with the company card without its own limit isn't UNATTENDED: it's running loose. Source

ABB Robotics, being acquired by SoftBank, breaks ground on its robotics hub in Sabadell and wants to rent robots out by the season: the robot becomes a monthly cost

ABB Robotics has begun construction on the technology and operations center it announced in July in Sabadell (Barcelona): 15,000 square meters, about 185 employees, and an expected opening in 2028, featuring engineering, computer vision, software, and autonomous mobile robots, a training center, a production area, and support for its mobile-robot factory in Burgos. SoftBank agreed to buy ABB Robotics in 2025 for €4.8 billion; Brussels approved it in April and the closing is still pending. Its president, Marc Segura, noted that in the year 2000, 90,000 robots were sold annually worldwide compared to nearly half a million today, warned that Spain and Catalonia have a "low density" of robots, and predicted they'll be rented as a service: for the Christmas rush, say, and then returned. For your business: if the robot is rented by the month, it stops being an investment reserved for large factories and becomes a cost you weigh against hiring temps, in a context where 63% of European companies report difficulties filling positions. Before asking for a price, pick the most repetitive, physically demanding task—palletizing, loading a machine, moving material between zones—measure how many hours per month are dedicated to it, and ask for a monthly quote, with maintenance and pickup included. As we saw with Toyota, which is putting its heavy spending in 2028, the years before are for deciding what to automate. A plant without a single measured workstation is still MANUAL, robot or no robot. Source

What to watch tomorrow

Whether OpenAI and Google disclose their own incidents now that the White House requires it, and whether any customer publicly compares Jev, Microsoft-Decision-1 and Clef on its own data. TechCrunch Disrupt kicks off in San Francisco on Tuesday the 13th and runs through the 15th.