noticias ia para empresas anthropic agentes de ia

The week opens with a European rule that now lives inside the text itself. Starting Tuesday, November 3, Anthropic adds its invisible watermark to four older Claude models, so, with one technical caveat, everything Claude writes will carry a machine-readable signal. Pipedrive, the CRM used by more than 100,000 businesses, is cutting 267 jobs in Estonia, 70% of its staff there, and its CEO says AI has changed how companies buy software. On the money side, Aramco is paying $2.6 billion for 49% of HUMAIN, the Saudi AI company it will also buy services from. And DeepSeek has released an agent that decides on its own which tasks to pick up. Catch up on Sunday's radar.

Claude watermarks everything it writes from Tuesday: the AI label now travels inside the text

Anthropic told its API customers by email that starting Tuesday, November 3, it will add its text watermark to four older models: Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. The Claude Platform gets it first, with Amazon Bedrock following a few days later; every model released since August 2 already shipped with it. The mark is an imperceptible pattern in word choice. It adds no characters, doesn't change pricing, carries no information about the user or the company, and can't be switched off. Because it sits in the model itself, it also shows up when you use Claude through Amazon, Google Cloud or Microsoft, it travels with the text when you copy and paste it, and it can survive some editing. Anthropic is doing this to comply with Article 50 of the EU AI Act, whose code of practice it has signed, and ties the retrofit to the December deadline. Two caveats. On Opus 4.6 and Sonnet 4.6, the mark only applies if you leave temperature and the two other randomness settings at their defaults. And the text detector is in private preview, limited to regulators, media, researchers and companies the law requires to verify it; for images and files there's a free checker. OpenAI is taking the same route with ChatGPT, as we covered in the October 6 radar. The takeaway: the emails, product pages and proposals you draft with Claude will carry a signal a regulator or a journalist can read, so don't tell a customer a person wrote something if that isn't true. If you've built a product or an automation on the API and you're counting on that mark for compliance, write down which model and which settings each workflow uses, and leave the defaults alone on the 4.6 models. Anthropic is explicit that working out what Article 50 requires of your product is your job: knowing which model sits behind every process is part of the INTEGRATED rung of the ladder. Source · Source

Pipedrive cuts 70% of its Estonian staff: AI is changing how customers find you

Pipedrive, the sales CRM used by more than 100,000 businesses in 180 countries, announced the biggest restructuring in its 16-year history on October 29, and Estonian business daily Äripäev has now put numbers on it: 267 jobs gone in Estonia, 219 in Tallinn and 48 in Tartu, or 70% of its 380 employees there; its Lisbon office, with about 220 people, and its Prague office will close, and developers in Berlin will be hit too. The company, controlled since 2020 by US private equity firm Vista Equity Partners, booked €237 million in revenue last year and €32.7 million in operating profit, so it isn't losing money. CEO Paulo Cunha says organic traffic to its website has dropped sharply because buyers now ask AI assistants which software to use, that paid acquisition has become more expensive, and that a CRM that just records sales is no longer enough: customers expect it to automate them. He denies this is "an exercise in using AI to reduce costs" or about replacing people with AI. A former CTO thinks Vista is getting the company ready for a sale, which nobody has confirmed. The takeaway: if Pipedrive is your CRM, export a copy of your contacts and deals today and ask your account manager what changes in support; a vendor rebuilding itself this fast can move prices and features. Then apply the lesson to your own business: compare how many visits you get from search engines with a year ago, and ask ChatGPT or Gemini about your category in your city to see whether you get named. If you win customers through your website, that's the channel that's shifting. Source

Aramco pays $2.6 billion for nearly half of its AI supplier: big companies are buying AI, not renting it

Saudi oil giant Aramco closed a deal on Sunday, through a wholly owned subsidiary, for 49% of HUMAIN, the AI company owned by Saudi sovereign wealth fund PIF, which keeps the other 51%. Aramco will put in about 9.9 billion riyals, roughly $2.6 billion, in cash during 2026, plus AI, compute and digital assets; the announcement doesn't value those assets or the company. HUMAIN, launched in May 2025, is targeting 3 gigawatts of data center capacity by 2030 and wants to sell everything from infrastructure to models and applications. Aramco will be both shareholder and customer: it plans to use that infrastructure for industrial applications, and CEO Amin Nasser says its technology initiatives have delivered $11.3 billion in value since 2023, much of it from AI. The takeaway: you won't buy your vendor, but you can copy Aramco's method, which is to track realized value. Keep a sheet listing every automation, what it cost and what it saves or earns each month, with dates. Without that ledger you won't know what to scale and what to shut off, and it's exactly what the UNATTENDED rung demands: a process that runs on its own has to pay for itself on its own. Source

DeepSeek releases an agent that picks its own tasks: the limit is no longer the model, it's whoever is watching it

On Saturday, October 31, DeepSeek's Harness team released dsh-superagent, an MIT-licensed research prototype: a coding agent that listens to a GitHub repository, and optionally the Feishu chat app, works out which messages are allowed work requests and spins up a separate agent session for each one, in parallel. Its argument is that one person can only really follow three to five agent sessions at once, so the bottleneck is human attention, not the model. The team tested it on simulated streams of about 300 messages, of which only about 20 were real work requests. The takeaway: this isn't a product, it's an early build from a Chinese vendor, so don't hook it up to a repository that holds customer data. What's useful is the idea: if someone on your team spends the day watching what the agents are doing, you haven't automated anything. The UNATTENDED rung means the agent filters and executes within written rules about what it may do alone, and a person reviews the result at the end. Source · Source

What to watch this week

Whether Tuesday's watermark rollout reaches all four models without changes your customers notice, and when Anthropic opens its detector to more companies. Whether Pipedrive spells out what changes for its customers in Europe. And whether OpenAI closes the first tranche of its funding round.