noticias ia para empresas agentes de ia precios ia

Companies keep using more AI, yet their spending grew just 1.5% over two months, according to Ramp: prices are falling, and that gives buyers leverage. Two deals show what can happen to your vendor: Five9 is paying $325 million for Level AI, and Conductor, now headed to Meta, shuts its cloud on November 20. Three more items bring things back to earth: agents already cost money to the websites they visit, the Bank of England sees only 0.5% in productivity gains over three years, and the UK is drawing up powers to investigate when an agent goes off the rails. This follows Friday's radar.

Companies use more AI but barely spend more: divide your bill by headcount and negotiate

Ramp, which sees the payments of more than 85,000 U.S. businesses, told TechCrunch on Friday that spending on tokens, the unit AI models are billed in, grew only 1.5% across August and September, after 42% in July. In the latest week measured it rebounded 8.9% against the same week a month earlier, on 37.1% more volume, because the blended price per million tokens from OpenAI and Anthropic fell 26.8%, from $0.76 to $0.56. Ramp's lead economist, Ara Kharazian, calls it a buyer's market: price cuts, contract discounts, cheaper models for simple jobs and per-employee budgets. The typical company paying for AI spends $15.70 per employee per month, the top 10% about $813 and the top 1% about $9,568, and Kharazian doubts a doctor's office or a car dealership needs to get anywhere near that. These are U.S. figures, from a firm that sells token spend management. For your company: divide your AI bill by your headcount. If you're around that $15, you're probably paying for chat licenses and sitting on the ASSISTED rung: your problem isn't price, it's the processes you haven't touched. If you're renewing, ask for today's price and avoid long lock-ins. Make a light model the default, keep the expensive one for work whose results justify it, and set a cap per person. And don't drop the margin test at 30% higher prices: that competition will keep prices from rising is Kharazian's opinion, not a clause in your contract. Source and Source

Five9 pays $325 million for Level AI and Conductor shuts its cloud in 16 days: when your vendor gets bought

Five9, which sells contact center software, announced on Thursday that it's buying Level AI, whose tools listen to and score calls, coach human agents and automate simple requests. It's paying $325 million in cash, almost half of its $738 million in cash and investments, and promises to keep selling those tools on their own. Its shares fell more than 23% in late trading, with soft guidance in the mix. That makes three customer service deals this week, after Cresta and Ada. Conductor shows the other side. It runs coding agents in parallel and is going to Meta at an undisclosed price: no new sign-ups since November 4, its cloud goes read-only on the 13th, and on the 20th it starts deleting data. The local app gets critical fixes until early 2027, and after that the company deletes what it holds there too. For your company: if anyone on your team uses Conductor Cloud, export now: on the 13th it stops taking new work, and on the 20th deletion begins. For everyone else, the lesson is the timeline: 16 days from announcement to deletion. List the tools from small vendors that hold up a real process, and in every contract demand 90 days' notice, export in a format someone else can read, and a refund of what you've paid. If they won't give you that, write down what you'd replace them with. Source and Source

Agents already cost websites money and now shop by text: decide which ones you let in

Cloudflare CEO Matthew Prince said on the company's earnings call that some customers see agents as "a bane to their business." His example: an agent that checks 1,000 restaurants to pick one leaves 999 of them with the cost and no customer. He says about 5% of the web has opted in to Cloudflare's content protection products since July, and the company expects 1,000 bot visits for every human one; keep in mind Cloudflare sells the remedy. On Friday, Amazon opened Alexa for Shopping by text to invited U.S. customers: it searches, compares and buys with your approval, or on its own when a price hits the target you set. Amazon blocked Meta's Muse agent back in September, and DoorDash said last month that third-party agents have placed only a few hundred orders on its platform. For your company: check your site logs or ask your provider how much of your traffic is bots and what it costs you. Don't block blindly: separate the crawlers that put you in front of customers, like ChatGPT's search, from the ones that just take content. If you sell online, keep price, stock and delivery times in plain text. And don't spend yet on selling to shopping agents: at those order numbers, it's still a bet. Source and Source

The Bank of England sees just 0.5% in productivity over three years: measure your process before you expand

This week's Monetary Policy Report says British firms in the Bank's survey estimate AI has raised their productivity by around 0.5% over three years, one or two tenths of a point a year. AI-related activity accounts for just under a fifth of GDP growth since the start of 2026, but the Bank attributes that to the companies building and implementing the technology, not the ones adopting it: for now, selling AI looks better than using it. The effect on total employment is close to zero, though the Bank's contacts say AI makes firms less likely to hire or replace people who leave, not more likely to fire. Firms expect far more, 0.9 points a year over the next three years, and its contacts say firms using AI proficiently and at scale are already automating routine admin work. For your company: this is a national average, but it says buying licenses isn't enough. Pick one process, write down today's hours and errors, and compare in 90 days. If they don't fall, don't expand. And if you're leaving a vacancy open because "we have AI now," put it in writing: which task goes to which tool and who checks it. Source and Source

The UK is drawing up powers to investigate when an agent goes rogue: list yours

An amendment to Britain's Cyber Security and Resilience Bill, added this week in the House of Lords, according to The Register, would let ministers regulate how certain organizations deploy AI agents: safeguards, notifying those affected, preserving evidence and reviewing incidents. Stuart Davey, a partner at Pinsent Masons, told the publication that could include appointing inspectors, entering premises and seizing documents. It imposes nothing yet: it needs consultation and follow-up legislation. It could cover gas and electricity suppliers, managed service providers and cloud providers deploying agents for customers, but not model makers simply for building the technology. It follows incidents with OpenAI's agents. For your company: it doesn't bind you outside the UK, but it previews the question you'll get if something goes wrong: who granted those permissions? Build a table with each agent, what systems it reaches, who answers for it and where its logs live, and keep those logs for at least 90 days. If your IT provider runs agents on your systems, ask for the same table in writing, plus how it would alert you to an incident. That's also what moves you toward the UNATTENDED rung: without that table, there's no process you can let go of. Source

What to watch tomorrow

On Monday the 9th, Copilot Chat's caps reach large organizations, and that week should show whether Anthropic starts its investor roadshow. Also watch Ramp's next index: it will tell us whether October's rebound holds.