noticias ia para empresas agentes de ia automatizacion de procesos

The gateway to your website changed its hinges on Tuesday: Cloudflare has begun blocking AI crawlers that mix search with training by default. Capital confirms the same movement from the other side—Profound raised $180 million for measuring whether assistants recommend you—and at Dreamforce, Salesforce stopped renting intelligence from labs and presented its own. This follows the Tuesday, Sept 15 radar, where Microsoft wrote the rules that its models cannot break.

Since Tuesday, your website blocks AI crawlers by default, and likely no one in your company decided it

The policy Cloudflare announced in July went into effect on September 15: "mixed-use" crawlers—those that use the same bot to index for search engines, gather training material, and fetch content for an agent—are now blocked by default on any page that displays ads. The settings change affects new customers, new sites for existing customers, and all free plan customers. Matthew Prince's argument is one of cost: most internet traffic is no longer human, and more than 50% of what AI crawlers spend is re-requesting pages that haven't changed. For your company: if your website is behind Cloudflare, this week you need to open the panel—or ask your hosting provider—and manually decide what to let through, because the default choice has already been made for you. Blocking training while allowing retrieval for agents is not the same thing, and that checkbox determines whether an assistant can cite you when a customer asks about your sector. Source

$180 million to know if ChatGPT recommends your company

Profound closed a $180 million Series D led by Sequoia and Kleiner Perkins, with Lightspeed, Khosla Ventures, and South Park Commons, bringing the valuation to $1.8 billion—less than seven months after a $96 million Series C at $1 billion. The company is two years old, has tripled revenue in six months, and has over 1,000 corporate clients, including Comcast, Estée Lauder, and Walmart. What it sells is measuring and improving how ChatGPT, Gemini, Perplexity, or Claude mention you when someone asks for a recommendation. For your company: the market is telling you where recommendations are now decided, and you can set up the budget version of this yourself. Once a month, for half an hour: ask the three assistants your customers use "who provides [your service] in [your province]?", note if you appear, what data they use to describe you, and which page they pulled it from. If they describe you poorly, the problem isn't the assistant: it's that your website doesn't clearly state what you do. Source

Salesforce builds its own model: The CRM stops paying labs for every task

At the opening of Dreamforce, Salesforce introduced Koa, its first reasoning model, built by post-training Nvidia's Nemotron-3-Super-120B—an open-weights model—with reinforcement learning on synthetic CRM flows modeled from 27 years of in-house deployments. On its own benchmark, CRM Bench, it matches or outperforms leading general models in CRM actions with three times fewer errors, and consumes fewer tokens than sending the same task to Claude or ChatGPT. A fact that matters more than the benchmark: no real customer data was used, only synthetic data mimicking patterns. For your company: the question for the next renewal of any software you already pay for is whether the model is included, what the usage limits are, and what data it was trained on. A small, specialized model within the tool is usually cheaper and carries less exposure than connecting that tool via API to a frontier model, which is what almost everyone set up first. Source

Someone is already auditing your agents from the outside: 5,000 tests and a 100-page report

AIUC—Artificial Intelligence Underwriting Company—founded by Rune Kvist, one of Anthropic's first hires, and Rajiv Dattani, former COO of METR, has raised $40 million in Series A led by Ribbit Capital, totaling $55 million including a $15 million seed round with Nat Friedman's NFDG fund, Emergence, Terrain, and Anthropic co-founder Ben Mann. It sells independent auditing and certification for agents: its AIUC-1 standard undergoes about 5,000 tests for data leakage, hallucination, and jailbreaking, returning a report of around 100 pages; a SOC 2 for agents, with Cursor, Lovable, Harvey, and ElevenLabs as launch customers. Kvist sums it up well: banks, hospitals, governments, and militaries are no longer holding back on AI deployment because the model isn't smart enough, but because of the commitments they have signed with their clients. For your company: yesterday Microsoft was writing its own rules, today someone appears to verify them. When you buy an agent, ask for a third-party report; if the provider doesn't have one, demand at least in writing which tests it has passed and on what date. That's what turns an assisted pilot into a process you can leave working on its own. Source

What to watch tomorrow?

Dreamforce continues until Thursday: if Salesforce puts a price on Koa within Agentforce, we will have the first real benchmark of how much a corporate model costs versus a frontier one via API. And it's worth keeping an eye on the major crawlers: anyone who doesn't separate their search bot from their training bot will be locked out of half the web by default.