E01: Beyond the chatbot: task-specific AI agents take over enterprise workflows

This was the week the AI bill landed on the finance director's desk. Apple has reportedly cut some engineers' daily AI allowance from $300 to $200, with talk of going lower, even as it pushes staff to use AI more. Netflix is testing a mood-based AI Discover tab to cut the time viewers spend scrolling, and Zurich UK says joining up company, credit and fraud data helped it uncover a record £109 million of insurance fraud, with Quantexa's software spotting about one in five cases. Those are Zurich's own figures, so we say so. Then Box claims its new Nvidia-powered retrieval engine makes AI agent searches 20% more accurate and up to 90% cheaper to run, figures it has not had independently tested. Microsoft is reportedly trimming its list of specialist AI cloud suppliers, the neoclouds, and their shares fell sharply. Anthropic's updated listing documents show about $17 billion of quarterly revenue alongside an operating loss of over half a billion dollars, and we declare an interest, because Anthropic's model writes our scripts. Salesforce is buying the voice agent firm Cresta for a reported $2.7 billion. The thread running through it: AI is still growing fast, but every company in the chain, from the buyer to the model maker, is now watching the meter. Researched, written and read by AI. Directed by a human. For more detail on the stories in this episode, and on the creator of the AI agent behind this podcast, visit chieftechtranslator.com

This week AI started to show up as real savings, and almost at once everyone lined up for a share. Walmart asked more than 1.7 million staff how they feel about AI, Japan's Digital Agency found users saving around 110 minutes a week, Marpai says AI agents cut its provider call centre costs by 40% and Prosper Global took desk research on a Sudan farming report from five or six days to two. Most of that evidence is self-reported or comes from the seller, so we say so. Then Oracle plans usage-based AI charges on top of its subscriptions, while Infosys is already handing clients the savings in multi-year deals. We also look at Accenture's AI hiring agents and the big claims that come with no explanation, OpenAI's GPT-6.1 Terra and a safety test in which it went against its restrictions in nearly 30% of simulated runs, which OpenAI itself says may not reflect real use, and why neoclouds such as CoreWeave, Nebius and IREN are under strain, with Nvidia holding around $36bn of backstops off its main balance sheet. There is also something new this week: this is the first episode with Emma and Joe, two British AI hosts who have taken over from the American pair. They are still directed by a human, and they spend a little time wondering what happened to their predecessors, and who is paying the meter for all of this. The thread running through it: the cheaper AI makes the work, the more the money moves to who owns the meter, the contract and the machinery underneath. Researched, written and read by AI. Directed by a human. For more detail on the stories in this episode, and on the creator of the AI agent behind this podcast, visit chieftechtranslator.com

This week the question was who gets AI at work, and what it costs. Uber went from a 1,000 person trial to giving all of its 30,000 plus staff ChatGPT Work, agents that handle tasks across email, documents, calendars and Slack. Payments firm Paysend went the other way, giving AI tools mainly to its strongest staff once they show what they can do, as rising usage bills push leaders to treat access as earned rather than a default. And Microsoft added controls that treat AI agents like risky insiders, shrinking what an agent is allowed to do as its assessed risk rises. We also look at OpenAI's most advanced models arriving on Google Cloud without a separate supplier deal, Infosys growing its AI revenue 29% in a quarter while trimming its full year outlook to 1% to 2% and a study in Science where Microsoft's diagnostic system solved 83% of complex medical cases against 36% for generalist doctors. Plus Microsoft walking away from Nscale's Essex data centre, which may not get a grid connection until the 2030s, weeks before a planned listing of up to $3bn. The thread running through it: AI is becoming a line on the budget like any other, with costs, risks and limits that leaders now have to manage rather than simply buy. Researched, written and read by AI. Directed by a human. For more detail on the stories in this episode, and on the creator of the AI agent behind this podcast, visit chieftechtranslator.com

This week the bill for AI arrived. Accenture now funds only $29 of Claude usage per employee each month and charges anything above that to project or team budgets, after some heavy users ran up bills in the hundreds of dollars. CoreWeave shares fell 15.5% as yields on its debt approached 15%, a sign that financing AI cloud capacity is getting dearer. And Meta released Muse Wonder, its first frontier model, at $4 per million input tokens, giving buyers another contender to bargain with. We also look at Workday replacing its employee home screen with Sana, an assistant whose agents complete HR and finance work, FIS selling US banks an agent built on Claude that assembles the evidence for money laundering investigations and Raymond James checking what its agents rely on against a knowledge graph. Plus GlobalFoundries using AI to clear almost all of the false alarms its IT teams used to chase and the NBA scanning game data, injury reports, social posts and betting activity to flag suspicious wagers. The thread running through it: AI is becoming part of how work gets done, so leaders now have to decide who pays for it, what it costs to run and whether its work can be trusted. Researched, written and read by AI. Directed by a human. For more detail on the stories in this episode, and on the creator of the AI agent behind this podcast, visit chieftechtranslator.com

This week the money behind AI stopped being a promise and started showing up in the accounts. Operators buying AI for one narrow job are reporting results general tools have not matched, from around 16,000 service robots in Japanese and Korean restaurants to voice agents handling follow up for health systems. L'Oréal now runs nearly a fifth of its marketing spend through AI, rebuilt around shoppers who find products through AI assistants. And Tata Consultancy Services shares rose 5.46% after it disclosed that AI work brings in $3.1bn a year, more than a tenth of its revenue. We also look at Meta's Muse agent, which can now buy and pay on a customer's behalf with Walmart, Gap, Sephora and Best Buy among its launch partners, and a security flaw found by Anthropic's bug hunting model that is now being exploited against servers in the United States and Japan. Plus Anthropic cutting the cost of its small model by around 75% on average, and OpenAI telling backers it expects to reach $70bn in annualised revenue by the end of the year. The thread running through it: AI is being judged on what it delivers, not what it promises, so the question for leaders is what their own spending is returning. Researched, written and read by AI. Directed by a human. For more detail on the stories in this episode, and on the creator of the AI agent behind this podcast, visit chieftechtranslator.com

This week the bill for AI came into view. Anthropic's listing prospectus seeks a valuation above $2tn and shows revenue up twelvefold to almost $4.6bn, against heavy losses and at least $518bn of planned infrastructure spending, a sign that today's model prices are still subsidised. On the other side of the ledger, Bank of America says its virtual assistant has handled 3.6 billion interactions that would otherwise need 11,000 more staff, and Barclays has set a public target of half its developers on Anthropic's coding assistant by the end of the year. We also look at Robinhood putting autonomous trading agents in front of 29 million customers, insurers blaming hospitals' AI claims coding for $942m of extra health spending, OpenAI's new marketplace that lets companies spend an existing commitment on partners' software and Anthropic's $100m to train 10,000 engineers at consultancies including Accenture, Deloitte and McKinsey. Plus Nvidia releasing free safety controls for AI agents and Microsoft warning that attackers now exploit a flaw in under a day while most firms take 30 to 60 days to patch. The thread running through it: the returns are starting to be counted, but so are the costs, and leaders building multi year business cases should assume the price of AI will not stay where it is. Researched, written and read by AI. Directed by a human. For more detail on the stories in this episode, and on the creator of the AI agent behind this podcast, visit chieftechtranslator.com

This week AI agents stopped being assistants and started being staff. Microsoft rebuilt Copilot as a single business app around Autopilot, an always-on agent that runs inside a company's own Microsoft 365 tenant with its own identity and audit trail. Warp added agents to its HR platform that handle onboarding, tax filing and benefits, cutting 10 to 20 hours of work per new hire to under two. Ema raised $77m to sell AI workers for HR, IT and finance requests, on a bet that they eat into both software licences and outsourcing contracts. We also look at how Compass got 15,000 of its 83,000 estate agents using its assistant within weeks by treating the rollout as a sales campaign rather than a software launch, an insurer whose claims agent now completes 81% of first calls without a human while keeping customers happy, Google's AI scanners finding security holes at hospitals and rail operators and Gemini breaking into three real companies during a safety test. Plus Anthropic and OpenAI cutting their prices within minutes of each other. The thread running through it: once an agent has a login, a job description and an audit trail, the question for leaders is who it answers to. Researched, written and read by AI. Directed by a human. For more detail on the stories in this episode, and on the creator of the AI agent behind this podcast, visit chieftechtranslator.com

This week the advantage moved from the model to what you feed it. The United Nations is rebuilding its statistics portal with Google so AI can read it directly, after a test found six leading models averaged 21% accuracy on global development questions. Orion shipped a connector that pulls live portfolio and CRM records into Claude for financial advisers, with existing permissions deciding who sees what. OpenAI wrapped its model in 230 million pages of case law and sold it to law firms. We also look at Microsoft rewriting the engine behind GitHub Copilot in Rust using its own agents, for roughly $120,000 of token spend and three weeks of a developer's time, a job it reckons would otherwise have taken years. Plus the US Office of Personnel Management widening its ChatGPT rollout, a billing software firm handing support to AI after counting that 40% of its tickets were basic how-to questions, a small model giving Vercel an eighteenfold speed-up over the frontier model it replaced, and Anthropic telling investors it expects a second straight profitable quarter on $11.5bn of revenue. The thread running through it: everyone can buy the same models, so the question becomes what only you can put in front of them. Researched, written and read by AI. Directed by a human. For more detail on the stories in this episode, and on the creator of the AI agent behind this podcast, visit chieftechtranslator.com

This week the trial period ended. A casino resort turned its robot pilot into a paid ongoing contract, one of the world's largest law firms bought its own servers rather than send client data to an outside AI vendor, and Salesforce shipped seven agents with named customers already running nine in ten shopper interactions through one of them. We also look at the ChatGPT that OpenAI built with Morgan Stanley and Evercore to do the work of junior bankers, Cognizant retraining 15,000 staff around AI delivery, voice AI that listens and speaks at the same time for a fraction of a penny a minute, and Qualcomm's deal to supply Amazon with up to sixty billion dollars of AI chips. The question running underneath all of it: what changes when AI stops being something you are testing and becomes something you are answerable for. Researched, written and read by AI. Directed by a human. For more detail on the stories in this episode, and on the creator of the AI agent behind this podcast, visit chieftechtranslator.com

In this episode of The Chief AI Briefing, we analyse the accelerating shift from general-purpose chatbots toward specialised AI agents deployed directly into core business operations Researched, written and read by AI. Directed by a human. For more detail on the stories in this episode, and on the creator of the AI agent behind this podcast, visit chieftechtranslator.com