noticias ia para empresas agentes de ia automatizacion de procesos

Today, the money points to where automation is heading: one agent company raises 77 million promising to take over what companies currently spend on software and external services, and another sees its valuation soar to 3.5 billion by selling what is in shortest supply—expert judgment. In between, OpenAI slashes the price of its workhorse models by half, the benchmark test for business processes reminds us that no model completes more than four out of ten from start to finish, and Amazon opens its sellers' back office to outside agents, starting with Claude. This follows Wednesday's radar, where we noted that price per token no longer predicts the cost of a job.

77 Million for "AI Employees": Agents are Already Competing for Your Software and Services Budget

Ema, founded in 2023, has closed a $77 million Series B led by Creaegis, and its valuation, which it does not disclose, has more than quadrupled compared to the previous round. It sells teams of agents that execute complete HR, IT, and finance processes on top of the applications the company already uses: they plan, review their own results, and request approval for critical decisions. Its figures, as stated by the company itself: revenue multiplied by 50 in two years, over 150 million in multi-year contracts, a gross margin near 80%, and clients like Hitachi, ADP, PwC, or Wipro; at one of them, a large outsourced services provider, it handles over a million incidents a year with less than 10% escalation to a human. Its CEO sums it up: "Enterprises don't need more software. They need work to get done." For your company, that sentence is a budget review: take your three largest recurring items for licenses and external services—IT support, accounting services, HR tools—and note down what work you pay for with each and how many units per month: tickets, payrolls, onboarding. With that number, you can compare the first offer that charges you per result rather than per user, and ask that offer for one piece of data before the demo: what part goes back to a person. That percentage tells you if a process can move up to the UNATTENDED level. Source

OpenAI Slashes Prices by 50% and No Model Finishes Even Half of the Test Processes

Ninety minutes after Anthropic launched Opus 5.5, OpenAI introduced GPT-6 Sol, for complex work, and GPT-6 Luna, for high-volume administrative tasks like data extraction, with a 50% price cut: Sol drops from $4 and $20 per million input and output tokens to $2 and $10, and Luna from $0.20 and $1.20 to $0.10 and $0.50. The news compared to yesterday is that the manufacturer itself is now selling its models by cost per task: on AutomationBench, Zapier's test with over 600 flows from six business areas spread across 47 applications, Sol completes 33.2% at $0.27 per task. OpenAI compares itself to Opus 5, not the Opus 5.5 from that same afternoon; on Zapier's public leaderboard, GPT-6 Astra leads with 41.4% at $1.73 per task, and Opus 5.5 follows with 40% at $1.28. For your company, an attempt now costs pennies, so price is no longer an excuse not to test; but in the most demanding test, even the best model fails nearly six out of ten end-to-end processes. Don't automate the entire process at once: break it down, leave the segments whose results you can verify without looking—the amount matches, all fields are present—to the AI, and reserve a person for the rest. And when comparing, divide the cost of each attempt by the successful ones: with Sol, the quick math gives about $0.80 per completed flow, not $0.27. Source

Amazon Opens Seller Back Office to Outside Agents, Starting with Claude

At its Accelerate seller conference, Amazon introduced a plugin that connects account data—listings, inventory, sales, and performance metrics—with external assistants: it launches on Amazon Quick and, in beta, on Claude, so that users can check stock, adjust prices, or update listings without opening Seller Central. The seller chooses which data the plugin sees, approves each action before it is executed, and everything is logged; the connection, according to Amazon, takes about 60 seconds and requires no coding. The company claims that 90% of its sellers already use some form of external AI, and its VP of Seller Experience summarizes it as "never having to go into Seller Central." For now, it is only for US stores, with international expansion dates TBD, and any account holder can request 12 months of Quick Plus for free until December 31. For your company, if you sell on Amazon, today's job is to write the rule before the button arrives: what the agent can do alone—restock below a certain inventory level—and what needs your signature—any price change above a certain percentage. Even if you don't sell on Amazon, watch the pattern: platforms are opening their back offices to external agents, and the question for your ERP, CRM, or bank is whether they already have that connector. Without it, your AI remains at the ASSISTED level. Source

Snorkel Worth 3.5 Billion: The Scarcest Resource in AI is Expert Judgment

Snorkel AI closed a $350 million Series E led by Insight Partners and S32 at a valuation of $3.5 billion, nearly triple the $1.3 billion from 17 months ago. Its annualized revenue has reached $375 million, eighteen times higher than a year ago, due to a shift in business: it stopped just selling data labeling software and now delivers finished task sets, test environments, and grading rubrics to AI labs and large enterprises—pieces that take a specialist hours or days to create. It's not alone: Mercor reports $2 billion in annualized gross revenue, and the sector pays between 60% and 70% of its billings to the experts it hires. For your company, what labs are paying top dollar for is something you already have but haven't written down: how your best person knows a job is well done. Pick a process—a complaint response, a quote, a reconciliation—gather 20 or 30 already resolved real cases, and note down the criteria used to approve them. That is your private test bench: with it, you can verify any new model in an afternoon, and it is the automatic verification without which no process moves from the INTEGRATED level to UNATTENDED. Source

What to watch for tomorrow?

The echoes of Wednesday's UN Security Council session, with Altman, Amodei, DeepSeek, and Moonshot discussing common standards: it sets the tone for what your clients will eventually ask of you. And whether Anthropic responds with a date for Sonnet 5.5 and Haiku 5.5: with Luna at $0.50, the price war is already being fought in the models used by SMEs.