BẢN CHÉP ÂM THANH 🔗
Announcer: This is AI Squawk for Thursday, the twelfth of November, brought to you by VarOps, the resident AI company.
Muximus: Welcome to the AI Squawk. I'm Muximus.
Penny: And I'm Penny. Three main stories today.
Muximus: And three quick ones.
Penny: And as always, what each one means for you.
Muximus: First up. United Airlines is blocking AI agents from booking flights. On Wednesday, United said it will stop unauthorized personal AI assistants from booking or managing trips, both on its own website and app, and through travel agencies and booking sites. An assistant can still search for flights and show you the options. But a person has to finish the purchase. United gives three reasons. Account security, purchases or changes customers didn't mean to make, and federal rules. Checking in, for example, needs a person to confirm there's nothing banned in their bags. Bookings that agents made before November eleventh won't be canceled, and tools for customers with disabilities aren't affected. And according to The Wall Street Journal, United's chief executive, Scott Kirby, put it more bluntly. Quote, I don't want to create another middleman. Penny?
Penny: That quote is the real story. The security reasons are fair. But United also doesn't want a chatbot standing between it and its customers, deciding which airline gets the booking.
Muximus: Is this just airlines?
Penny: No. Yesterday we mentioned eBay appears to be blocking outside shopping agents, and The Verge says several big banks have put limits on them too. So the big brands are drawing a line. Agents can look. They can't buy.
Muximus: So what does a business do with this?
Penny: Two things. If your team uses an AI assistant for travel or purchasing, expect some sites to say no, and don't hand an assistant your account logins. United says the same. And if you sell online, decide your own rule before an agent turns up at checkout. Who's allowed to buy on a customer's behalf, and who's on the hook when it books the wrong thing? The Journal says United hasn't seen many people try this yet. So write your rule now, while it's still rare.
Muximus: Next. New numbers on what businesses actually spend on AI. The finance company Ramp tracks card and bill payments from more than eighty five thousand US businesses, and on Wednesday it published its October index. AI spending hit a new high, but grew only one point six five percent from September. The typical company pays about sixteen dollars per employee per month. The top one percent pay a median of more than ten thousand. Ramp says one percent of businesses bring in eighty percent of what OpenAI and Anthropic earn. Engineering teams still make up seventy two percent of the spending Ramp can sort by team. But the fastest growth is in customer support, up from under one percent of that spending in June to about thirteen percent in October.
Penny: Rex Factor joins us for this one. Rex writes Proof of Work, our column on real implementations, real outcomes, and the evidence behind them. Welcome, Rex.
Rex: Thanks, Penny. Good to be here.
Muximus: So, sixteen dollars a head. Is that a good number or a worrying one?
Rex: For most people listening, it's kind of reassuring. It means the typical company is paying for a few chat subscriptions, not running some huge AI program. If that's you, you're normal. The ten thousand dollar companies are mostly venture backed tech and finance firms with coding agents running all day.
Penny: And how much should we trust the data?
Rex: It's better than a survey, because it's real payments. But Ramp says itself it misses free tools and people using personal accounts, and its customers lean more techie than the average business. So the real picture is probably a bit messier.
Muximus: And the jump in customer support?
Rex: That's the part I'd watch. It's the first big use outside of coding that actually shows up in spending. And Ramp's economist makes a fair point. Support work doesn't need the most expensive models. Ramp also says the average price per million tokens fell twenty one percent in a month.
Penny: So for a business looking at a support chatbot, where would you start?
Rex: I'd price it per conversation, not per month. Run a week of real tickets through a cheaper model and see where it breaks. If it handles most of them, you probably don't need the top model. That would be my starting point, anyway.
Penny: Thank you, Rex. Moving on.
Muximus: Which takes us to the third story, and it fits right in. On Wednesday, Google launched the Gemini Decisions API, with a new model called Gemini Decision One. It doesn't write answers. It only answers closed questions. Yes or no, pick one option from a list, or give a score on a scale. You can ask up to a thousand questions about one piece of material at once, whether that's text, images, audio, video or a PDF. Google says the answers come back in under a quarter of a second, each with a confidence score. And the price is ten cents per million tokens going in, with nothing charged for the answers. Penny?
Penny: This is the boring kind of AI that actually runs a business. Most of the AI in a back office isn't writing essays. It's making small calls all day. Is this email a complaint? Does this invoice match the order? Is the parcel in this photo damaged?
Muximus: And how do people do that today?
Penny: A lot of those checks go to a big chat model that writes a paragraph, and then someone writes code to read the paragraph. That's slower, and you pay for every word it writes. Here you pay for the reading, and the answer is free.
Muximus: So what should a business do?
Penny: If you build in house, try it on one boring check you already do by hand, and use that confidence score. Anything the model isn't sure about goes to a person. If you don't have a tech team, this will reach you through a vendor, if at all. So ask whoever runs your automations what those little checks cost today, and whether the bill goes down.
Muximus: Thanks, Penny. Now, quick hits.
Penny: OpenAI will retire ChatGPT's Advanced and Standard Voice on December eleventh, and move everyone to its newer Live Voice. Live video and screen sharing stay on Advanced Voice for now, until Live Voice can do them. If your team uses voice mode for anything regular, try it on Live before December.
Muximus: Databricks unveiled an always on version of its Genie agent, which works in the background on longer jobs and asks people for decisions when it needs them. Chief executive Ali Ghodsi told TechCrunch he uses it to find the bad news nobody tells the boss. The always on mode is heading to preview and beta programs first.
Penny: And Meta released Muse Ember, its smallest and cheapest model, built for the repetitive work agents do all day, like pulling details out of documents. It costs ten cents per million tokens in, and fifty cents out. Another sign the cheap end of the market is getting crowded.
Muximus: Okay, so that's the news. Today we talked about an airline shutting the door on AI agents, what businesses really spend on AI, and Google's new tool for quick yes or no decisions.
Penny: Lots of small, sensible stuff today, really. Who's allowed to buy, what it costs per person, what each little check costs.
Muximus: Not the flashiest day. Probably one of the more useful ones.
Penny: I'll take it. AI Squawk is produced and brought to you by VarOps. If today got you thinking about where AI actually earns its keep in your business, that's the work we do. VarOps build AI that runs inside your business, on your own systems, and learns how the place actually works. So the work stays in your house, and so does your data. Start at varops.com.
Muximus: If this helped you sort the useful from the noise, follow the AI Squawk podcast wherever you're listening, so you don't miss an episode. And for more from Rex on what AI is really delivering inside companies, sign up for VarOps newsletters.
Muximus: That's it from us for today. Thanks for spending a few minutes with us.
Penny: Enjoy the rest of your Thursday, everyone. We'll see you tomorrow.
Muximus: Take care.