noticias ia para empresas agentes de ia gestoría

Business AI money is coming in through the back office: bookkeeping, collections and spend control. France's Dougs has raised €50 million for an accounting firm where agents do the work and a chartered accountant signs off, Ramp has confirmed its round at a $60 billion valuation, and Intuit is buying Sequence. In Spain, Banco Sabadell is getting ready to roll out agents in 2027 on one condition: no lock-in to any single model provider, on the same day Nvidia released open models you can run in-house. And ChatGPT is now finding hundreds of millions of new pages thanks to Cloudflare. This picks up from Thursday's radar (October 15).

AI agents do the books, an accountant signs off: Dougs raises €50 million as the money moves into the back office

Dougs, a Lyon-based accounting firm registered with France's Order of Chartered Accountants, has raised €50 million from Expedition Growth Capital, an investor since 2023, which becomes the majority shareholder while the founders keep a majority of the voting rights. It serves more than 50,000 businesses, brings in over €50 million a year and says its gross margin tops 80%, versus under 50% for the industry. Its model: agents already handle bookkeeping and entire stages of production on their own, and the next step is financial statements generated by AI, then reviewed and signed off by an accountant. It plans to use the money to finish automating production and to get closer to clients through a network of local partners. Also on Thursday, Ramp confirmed the round we reported on Sunday, October 11: $1.85 billion at a $60 billion valuation, with more than $2 billion in annualized revenue, 85,000 business customers and a tool to control AI spend by provider, team, model and API key, the kind of caps we suggested yesterday. And Intuit, the maker of QuickBooks, is buying London-based Sequence, which automates everything from quote to cash, for an undisclosed sum. For your business: bookkeeping and collections are where automation is already paying for itself, and investors are backing that up. Ask your accountant three things: how much of your bookkeeping a machine already does, who reviews and signs off on the output, and whether those savings show up in your fees or in how many days it takes to close the month. An agent doing the work with a professional signing off is UNATTENDED with someone accountable; someone hand-keying invoices you already have as PDFs is MANUAL at expert rates. Source

Banco Sabadell will roll out AI agents in 2027 without locking into any model: kill switch and human sign-off first, scale second

Banco Sabadell has signed a deal with T-Systems, Deutsche Telekom's IT services arm, to use its European agentic AI platform, and plans to start rolling it out in the first half of 2027 in three areas: customer assistants, text first and voice later; "augmented relationship managers" that feed bankers context in real time while they serve clients; and internal process automation, including some risk analysis. Today the bank mostly uses OpenAI and is evaluating Google, and the platform lets it mix models by task, including open-weight models for the most sensitive data. CTO Rüdiger Schmidt puts it plainly: "we're not wedding ourselves to any model provider." That independence is key to complying with DORA, the EU regulation that requires banks to have exit plans and resilience for critical processes. The bank also sets conditions: agents must meet the same security and control requirements as its critical systems, with an emergency shutoff and human sign-off on critical decisions, because "you can't just let them run without protection." For your business: you don't need a bank-grade platform to copy the approach. Before you turn an agent loose, answer four questions in writing: what credentials it logs in with, which data and apps it can touch, who watches what it does, and how you stop it within a minute. Then flag which decisions need a human yes. We covered the exit plan on Wednesday, October 14. An agent with no kill switch and no human sign-off on critical calls isn't UNATTENDED: it's off the leash. Source

Nvidia releases open models for agents you can run in-house: an option for documents you don't want leaving the building

Nvidia has launched Nemotron 3.5 Super VL, an open model with 120 billion parameters, only 12 billion of them active at a time, built for always-on agents: it reads documents, tables, charts, PDFs and web pages, handles up to one million tokens of context, lists Spanish among its languages, and its license allows commercial use. Any company can post-train it on its own data and run it on its own infrastructure; Ramp is tuning it to read financial documents, and EY to run M&A due diligence. Alongside it comes Nemotron 3.5 Bolt, a version that fits in 24 GB of memory and runs on a PC with an Nvidia graphics card, with a 4 GB version due on October 23. Nvidia says Super VL is five times faster than the next-best model in its class; that's Nvidia's own figure. For your business: this is exactly the use case Sabadell describes. If you have documents you'd rather not send to an outside vendor, like payroll, contracts or customer records, take 50 real ones, run them through the model you use today and through an open model on your own machine or a European cloud, and compare accuracy, speed and cost. Picking the model for each task, not out of habit, is part of being INTEGRATED. Source

Cloudflare helps ChatGPT find hundreds of millions of new pages: make sure your site lets its search crawler in

Cloudflare and OpenAI have shared early results from a pilot that began in July: using signals from about 1% of Cloudflare's network, OpenAI's search crawler has discovered and indexed hundreds of millions of pages it had never seen, which ChatGPT can now draw on and cite. The signals flag which pages have changed and are worth another visit, and which don't exist or just redirect elsewhere, so crawlers waste fewer requests. Both companies stress that a signal doesn't grant permission to access a site and that no page content is shared for model training. Their examples of valuable pages are very specific: the manual for an exact appliance model, the requirements for a passport application, a restaurant's menu and opening hours. For your business: your customers are already asking ChatGPT before they call you. Check your robots.txt today: OpenAI separates OAI-SearchBot, which gets you into ChatGPT search results, from GPTBot, which collects content for training, so you can allow the first and block the second; changes take about 24 hours to kick in. If your site runs on Cloudflare, also check what its AI Crawl Control is blocking. And give the questions you get most their own pages: prices, lead times, hours and terms. Source

What to watch tomorrow

Whether OpenAI and Anthropic respond to the document demands the US FTC has now sent them over their safety incidents, according to Axios, in the investigation we covered on October 1. And on the money side, whether the automated-accounting wave reaches a Spanish accounting or advisory firm.