Mercury put an AI agent on your money
ChatGPT for your bank is here, but it won't actually move money on its own. Also: Midjourney's surprise hardware launch, OpenAI's big hire, Apple's supply chain warning, and more.
Immad Akhund doesn’t think the big banks can build what Mercury just shipped. “Even basic things are so hard to do in an incumbent bank,” Mercury’s CEO told me. “Try searching for a transaction. You can’t do a full-text search. There are so many obvious things where I’m like, ‘Wow, I can’t believe it’s so hard to do these at a bank.’”
On Tuesday, Mercury launched Command, an AI assistant that turns a ChatGPT-like interface into the primary way customers run their accounts: checking spend, sending money, issuing cards, creating invoices, and categorizing transactions. It runs on OpenAI’s GPT-5.5, though Mercury says it’s testing every frontier model and treats the intelligence powering Command as “a commodity.”
Plenty of banks have chatbots, and card-first rivals like Brex and Ramp have AI agents for expense management. Mercury is going further, making conversational AI the front door to a full banking product. (Mercury isn’t technically a chartered bank yet; it runs on partner banks and is pursuing a charter.)
The incumbents can’t follow, Akhund argues, because they don’t really care about deposit customers and are “much more lending-focused.” If anything, AI widens the gap, he says: “It took them maybe five years extra to get to a mobile app, and AI is moving really quickly.”
He wasn’t always a believer. “A year ago, I was pretty skeptical. I felt that AI would make mistakes with money, and people wouldn’t trust it.” Tool use and better reasoning changed his mind.
Command’s AI is read-only. Anything that moves money routes to a separate, deterministic service that isn’t AI, and the user clicks to approve it. Akhund contrasted that with agentic coding tools that “do the thing without asking me for permissions.” Eventually, he imagines employee-style controls: let the AI spend up to a set budget, but don’t give it free rein.
It’s a good thing Mercury hasn’t fully connected Command to an account’s money yet. During a live walkthrough that I got ahead of the launch, VP of product Ryan Wiggins watched Command stall, then fail outright on a second task. “Classic demo error,” he said. He’d had engineers freeze code changes for an hour to keep it stable. I got the same connection error trying it myself. In a follow-up email, the company said that a service connecting customers to LLMs had briefly gone down and required a manual reset.
Mercury has more than 200,000 customers, has been profitable for four years, and recently raised at a $5.2 billion valuation. After the launch of Command, the company said 3,000 customers tried Command in its first few hours, mostly for routine banking tasks like pulling statements and checking payments; the most common action was bulk categorizing transactions.
Whether Command makes anyone switch banks, or just keeps Mercury’s AI-native customers happy, is the part even Akhund calls “step one.” For now, it’s telling that even the most tech-forward banking startup won’t let its software touch your money on its own.
(Disclosure: Mercury paid to remove the paywall for this issue for a limited time. It had no editorial influence on coverage.)
A MESSAGE FROM MY SPONSOR
America’s Workforce Academy
The program, built by Meta, will help train those who build America's future.
America's Workforce Academy is the largest private sector commitment to the skilled trades that offers a job upon completion. The program provides paid training, a job and a path to America's future.
Because the future is for everyone.
Feed check
Midjourney is taking on the MRI industry and opening spas. I regrettably couldn’t make it to Midjourney’s big hardware launch event in San Francisco on Wednesday, but I watched the livestream and was intrigued to see that the company’s first hardware bet is for healthcare. Folks at the frontier of AI are increasingly buzzing about biotech, and it’s clear that founder David Holz is in that camp. He said Midjourney is building four hardware products focused on “creativity” and “positive human futures.” He aims to build 50,000 Midjourney scanners, starting with a boutique spa in Union Square late next year that will contain 10 of them. The ultimate goal is to bring “full body imaging to everyone on earth.” He was light on the technical details on how they work, but he said during the stream that his scanner is 10x cheaper and 60x faster than a MRI. Midjourney is a fascinating, mysterious company, and this is one of the most interesting bets from an AI lab I’ve seen in a long time.
OpenAI hired Noam Shazeer. This one shocked the timeline today. Google, and specifically Sergey Brin, paid billions of dollars to acquihire Shazeer away from his startup Character AI to come back and work on pre-training for Google DeepMind. He was co-lead for Gemini and is widely considered one of the most influential researchers in the field. OpenAI research chief Mark Chen says he’ll be the “new lead for architecture research.”
Apple sounds the supply chain warning. Tim Cook giving an interview about actual business news pertaining to Apple outside of an earnings call is about as rare as a lightning strike. When he did it today, he told The Wall Street Journal that Apple will increase prices due to rising memory costs: “There’s less supply at a time when consumers want devices and the memory guys are passing along huge price increases.” Apple has the best supply chain operations of any company on earth. If Cook is saying this now, be prepared for every consumer electronic device to skyrocket in price.
Anthropic may not be able to release Fable after all. The White House reportedly wants Anthropic to ensure that Fable can’t be jailbroken, which, as anyone who works on the security of AI will tell you, is basically impossible to guarantee.
When the seating chart tells the story. From today’s G7 meeting with President Trump: “At the lunch, OpenAI Chief Executive Sam Altman and Demis Hassabis, the head of Google’s DeepMind AI lab, were seated next to Trump. Amodei was seated on the opposite side of the table from the president, next to French President Emmanuel Macron. Down the table were Lutnick and Treasury Secretary Scott Bessent, who were among the officials that targeted the company’s models last week.”
The post-WWDC super nerd convo
We sit down with Ryan Jones, founder of Flighty and a 13-year top indie developer, to talk about what it actually takes to build a beloved app in the age of AI. We talk about Apple's big WWDC announcements, the new Siri, and why every vibe coder who tries to clone Flighty eventually gives up. We also get into Flighty's upcoming AI launch, whether craft and design still matter when anyone can build an app, and whether Apple is falling dangerously behind in AI or not.
Listen or watch wherever you get podcasts.
ICYMI
Sources is a newsletter by Alex Heath about the AI race, featuring scoops, unique analysis, and exclusive interviews. Every week, Sources is read by thousands of decision makers in tech, finance, policy, and media.
