noticias ia para empresas agentes de ia recursos humanos

AI isn't reaching small businesses as a project anymore. It's arriving inside the software they already pay for: Factorial is opening its One agent for free to its 17,000 business customers. The money keeps following agents that move real cash—Stuut raised $52.5 million to chase unpaid invoices—and the rules for doing business with machines are taking shape: a standard for how your customers' agents deal with your business, and a Google detector, open to anyone, that tells you whether an image was made with AI. Picks up from Wednesday's radar, October 7.

Factorial opens its AI agent for free to 17,000 companies: if you use it for shifts or hiring, explaining it is on you

Factorial, the HR software company founded in Barcelona ten years ago, launched a revamped version of One, its AI agent, on Wednesday and is letting all its clients use it for free: about 17,000 companies and 1.3 million users in more than 90 countries. The new version, which until now was being tested with more than a hundred customers, drafts emails, builds reports and automates tasks using what the company already keeps in Factorial, from shifts to payroll and expenses. Factorial built it in-house on Microsoft models served from Microsoft's European cloud, with paid tiers from 2 to 49 euros per employee per month, depending on usage. "Our clients want to use AI, but they are afraid of not complying with regulations," said its CEO, Jordi Romero. A few weeks ago, Factorial bought Empion, a Berlin startup that evaluates candidates with AI. For your company: what we reported yesterday about SAP for big companies is landing today, for free, in a tool built for small businesses, with no project and no budget. Before you roll it out to all your staff, set three rules. First, put in writing what the agent can do and what it can only suggest: drafting a report is not the same as changing a shift. Second, if you use it to assign shifts or tasks or to decide promotions, Spain's Royal Decree 723/2026 requires you to tell every worker that it exists and what rules it follows; and if you use it to filter or evaluate candidates, the European AI Act considers it a high-risk use (Annex III). Third, set a spending cap before you move to a paid tier: with usage-based pricing, your heaviest users set the bill. An AI that slips in free through the back door, with no rules, isn't INTEGRATED: it's a stowaway. Source

Stuut raises $52.5 million for AI agents to collect your invoices: 81.7% of collections work now runs without humans

New York-based Stuut, which uses AI agents to automate the entire order-to-cash cycle—invoicing, collections, cash application and disputes—announced a $52.5 million Series B on Wednesday, led by Insight Partners, with Andreessen Horowitz, M12 (Microsoft's fund), and Activant. That brings its total to $93 million, ten months after its Series A. Its agents email, text and call customers who owe money, and match each payment to its invoice. According to the company, 81.7% of outbound collections activity now runs without human involvement, 95% of incoming payments are matched automatically, and customers cut their days sales outstanding by 47%: ZoomInfo reduced its average collection period from 51 to 40 days. It has more than 150 customers, every action is logged, and any change in the agents' behavior needs sign-off. The round lands as startup funding in the US and Canada fell 35% from the previous quarter, to $92 billion, and still about two-thirds went to AI, according to Crunchbase. For your company: capital is flowing to agents whose work shows up in your cash balance, and collections is the clearest case. Stuut sells to big companies, but the lesson works for a small business too. Start with reconciliation (matching what hits the bank to each invoice): it's repetitive, easy to check, and your accounting software may already do it. Reminders come next, with rules: approved templates, separate handling for your ten best customers, and a person who signs off before the second reminder. And make sure the agent knows Spanish law: unless the contract says otherwise, businesses must pay within 30 days, never more than 60 even by agreement, and you can charge a flat 40 euros for every late invoice (Articles 4 and 8 of Law 3/2004). If collecting depends on someone remembering to call, you're not even at ASSISTED: you're still MANUAL. Source

Meta, Walmart, Shopify and Stripe set the rules for how your customers' AI agents deal with you: decide now what they can do without a human

On Tuesday afternoon, Meta and Sierra, the customer service AI company co-founded by OpenAI chairman Bret Taylor, unveiled the Personal Agent Protocol, an open standard for how consumers' personal agents deal with businesses. They are developing it with Genesys, Instinct, Rocket, Shopify, Stripe, and Walmart. Taylor compares it to "Sign in with Google": the agent starts as a guest, which is enough to check availability or a returns policy; if the task needs the customer's account, the customer signs in and decides whether the agent can only read or also make changes; and the company decides what the agent is allowed to do and can see what it does, through its website, its API or its own agent. The first spec is due this month; payments without sharing card details come later. It's needed because personal agents are already out there: Meta's Muse has millions of users in the US, and Amazon blocks it over concerns about scraping its site. "It is kind of chaos until such a standard exists," Taylor told CNBC. OpenAI and Anthropic aren't on board yet, though Taylor expects them to join. For your company: more and more, whoever books an appointment, asks for a price or changes an order will be a customer's agent. Don't build anything yet: with Shopify and Stripe among the partners, if you sell or take payments through them, it will likely reach you that way. But write today the list the protocol will ask you for: what an agent can do without identifying itself (check prices, opening hours or availability), what requires the customer to sign in (view or change an order) and what you'll never do without the person (payments, cancellations with fees, sensitive data). And check your website and chat analytics to see whether agents are already showing up. If you don't decide what your customer's agent can do, it will decide for you: your customer service won't be UNATTENDED, it'll be at someone else's mercy. Source

Google opens its AI image, video and audio detector to everyone: use it before you pay out on a photo, never as proof

On Wednesday Google opened SynthID Detector to everyone, in English for now. The site checks whether an image, video or audio file carries the invisible watermark added by its AI tools and those of its partners OpenAI, Nvidia and Kakao; Apple is coming "soon." Until now, only journalists and researchers used it. The checks Google already runs inside Search, Gemini and Chrome top one million a day. It has limits: it only finds that watermark, so it misses content from other tools; the mark can sometimes be stripped; it can't tell whether AI created the file or just edited it; and to use it you must sign in with a Google, OpenAI, or Apple account. For your company, two uses from today. Externally: any customer, reporter or competitor can now check whether the photo in your ad was made with AI, and if it shows people, places or objects that look real but aren't, the EU AI Act already requires you to disclose it (Article 50.4); better to label it yourself than explain it later. Internally: it's a free first filter for photos people send you to get paid, like damage shots for a return or a claim. A hit is a serious red flag; no hit doesn't prove the photo is real. And be careful what you upload: an image showing people is personal data. As we reported on Tuesday, OpenAI will also start watermarking ChatGPT text in the EU: the vendor adds the mark; checking and labeling is on you. A returns process that pays out on a photo nobody checked isn't UNATTENDED: it's exposed. Source

What to watch tomorrow

Whether OpenAI or Anthropic sign on to Meta and Sierra's agent protocol before its first spec, promised for this month; and on Friday, October 9, the last day of the Australian Parliament's hearings on AI agent incidents.