Get tomorrow's brief in your inbox
Today: Salesforce bought AI customer service platform Fin for $3.6 billion to boost its agent offerings. The Trump administration banned Anthropic's Fable and Mythos models over a disputed "jailbreak" that security experts say was just the model fixing bugs as designed. Zapier found 92% of sales teams drop qualified leads monthly because follow-up workflows are broken.
The Government Just Showed It Can Shut Down AI Models Overnight
On Friday afternoon, the U.S. Commerce Department forced Anthropic to pull its newest AI models, Fable 5 and Mythos 5, offline for all users worldwide. The reason: an alleged "guardrail bypass" that turns out to be the model doing exactly what coding assistants should do - fixing security bugs in code.
According to cybersecurity expert Katie Moussouris, who reviewed the report that triggered the ban, researchers asked Fable to "review code for security issues" and it refused. They then asked it to "fix this code" and it complied. That's not a jailbreak. That's a coding model working as intended. Over 76 cybersecurity professionals, including former Facebook CSO Alex Stamos and cryptographer Jon Callas, signed an open letter calling the ban "dangerous" because it removes the best defensive tools from security teams.
The real story here is not technical. It's political and strategic. The U.S. government demonstrated it can unilaterally shut down commercial AI products with 90 minutes' notice and no court approval. For business owners, this means three things: First, relying on a single AI provider is now a documented operational risk. Second, expect more sovereign AI development outside the U.S. as countries realize American AI access is conditional. Third, if you're building critical business processes on frontier models, have a fallback plan that uses older, more stable models or open-source alternatives.
Sources: TechCrunch on the ban, The Verge inside story, cybersecurity experts' protest
Salesforce Acquires Fin for $3.6B to Power Customer Service Agents
Salesforce is buying Fin (formerly Intercom) for $3.6 billion to strengthen its Agentforce platform. Fin's AI agent handles customer queries across chat, WhatsApp, SMS, phone, and Slack using its proprietary Apex model. If you run a service business and already use Salesforce, this matters: you'll likely see tighter integration between your CRM and AI-powered support automation in the next 6-12 months. The bet is on agents that can resolve customer issues without human intervention.
For businesses not locked into Salesforce, Malaysia's Respond.io just raised $62.5M and offers a similar multi-channel AI agent platform at potentially lower cost ($35M ARR at 30% profit margin suggests they're pricing competitively). Respond charges by conversation volume, not seat count, so if you're replacing human support with AI, your costs actually go down instead of staying flat.
Sources: CNBC on Salesforce deal, TechCrunch on Respond.io
AI Agent Identity Management Gets $66M
NewCore emerged from stealth with $66 million (at $300M valuation) to solve a problem you might not know you have yet: how to manage AI agents like employees. As companies deploy more AI agents - McKinsey claims 25,000 agents work alongside 60,000 employees there - traditional identity systems built for humans break down. NewCore treats AI agents as first-class identities with permissions, lifecycle controls, and revocation mechanisms.
If you're using Claude Code, Cursor, or other coding assistants across your team, NewCore's "Agentic Skill" integration package lets those tools access enterprise systems as managed identities instead of sharing API keys. That's the right architecture. The product goes live this summer.
Source: TechCrunch on NewCore
Meta Launches AI Mode Search on Facebook
Facebook now has an "AI Mode" that pulls answers from public posts across Groups, Reels, and regular posts instead of showing search results. You ask a question in plain language, Meta AI synthesizes an answer from what people are discussing. This is the same pattern Google used with Reddit threads.
The concern: you're getting answers from random public posts, not vetted sources. Expect outdated or misleading information. If you manage a Facebook presence for your business, your public posts now feed Meta's AI answers, which means clear, accurate information in your posts may get surfaced to potential customers searching for solutions you provide.
Sources: TechCrunch on AI Mode, The Verge coverage
Why 92% of Sales Teams Drop Qualified Leads
Zapier surveyed over 400 B2B sales managers and found 92% lose qualified leads every month because follow-ups are delayed, inconsistent, or never happen. The problem is not lack of tools. It's broken workflows between those tools.
Key findings: Sales reps spend 3-10 hours per week on CRM upkeep instead of selling (68% of teams). 37% say leads get stuck between marketing tools and CRM or between marketing and sales teams. 42% of teams fail to make a second or third follow-up attempt after initial contact.
The fix is not another tool. It's connecting the tools you have. If a lead fills a form, the next action should trigger automatically: create CRM record, enrich with data, assign to rep, send first email, set follow-up reminder. No manual steps. Zapier's pitch is obvious here, but the diagnosis is correct: you're losing revenue to workflow gaps, not feature gaps.
Source: Zapier survey results
Two People Who Built What Their Job Descriptions Didn't Ask For
Zapier's first monthly Zappy Award winners are worth studying. Rachael Silvano manages a 140,000-member community on a platform with no native points economy. Instead of accepting the limitation, she built a seven-step Zapier workflow using Airtable, JavaScript, and webhook calls to create a complete gamification layer. No development background. Used Zapier Copilot to work through the API calls.
Jeff Hirsch joined Redis and found Zapier wasn't broken - it just wasn't built. He went from zero Zapier experience to a fully governed enterprise implementation in weeks: SSO, SCIM, user provisioning, shared folders, service accounts, restricted actions. Teams are now requesting access and bringing automation ideas before formal rollout.
Both won because they saw a gap and filled it instead of waiting for budget, vendors, or permission. That's the behavior AI tools now enable: non-technical people building technical solutions to specific business problems.
Source: Zapier Zappy Awards
How AI in the Workplace Actually Shows Up
Zapier's deep dive on AI at work breaks it into useful categories. Chatbots handle back-and-forth interactions (support, Q&A). AI-powered automation embeds AI steps inside larger workflows (transcribe voice note, generate blog post, send for review). Agentic AI plans and executes multi-step tasks on its own. MCP (Model Context Protocol) gives AI structured access to your entire tech stack without custom integrations.
The pattern to watch: AI works best when it handles high-volume tasks humans can't sustain. A support team processing 5,000 tickets a month where each requires 15 minutes of research before a human can reply is a perfect use case. AI triages, surfaces relevant information, and drafts a response. Human reviews and sends. You just cut ticket resolution time in half.
Source: Zapier on AI in the workplace
Qualcomm CEO: AI Agents Will Replace Apps
Qualcomm is working on over 40 designs for new AI devices - wearables, smart glasses, earbuds with cameras, pins, watches. CEO Cristiano Amon says agents will become "the new app" and that apps will change fundamentally. Instead of navigating through an app to find a banking transaction, you ask an agent and it retrieves the details instantly.
Amon is bullish on smart glasses specifically, predicting shipments could reach "hundreds of millions" in a couple years and rival smartphones in scale. The implication for businesses: the way customers interact with your services may shift from app interfaces to voice/visual agent interactions. If you're building digital products, start thinking about how an agent would access your functionality.
Source: CNBC on Qualcomm
India's Sarvam Becomes AI Unicorn at $1.5B
Sarvam raised $234 million (with $150M from HCLTech) at a $1.5 billion valuation to build full-stack AI for India: models, inference infrastructure, enterprise applications. The company's models are designed for Indian languages and deployed across banking, insurance, government, and defense.
The investment reflects growing concerns over AI sovereignty after the U.S. government's Anthropic ban. Countries want control over critical AI technologies. Sarvam's platform handles 2 million conversations daily, processes 10 million API calls daily, and transcribes 500,000 hours of audio monthly. The strategic lesson: as frontier AI access becomes conditional, regional AI providers will gain traction.
Source: TechCrunch on Sarvam
Nvidia Raising $20B in First Debt Sale Since AI Boom
Nvidia is raising at least $20 billion in debt - its first bond sale since ChatGPT launched. The company generated $49 billion in free cash flow last quarter and plans to return roughly 50% to shareholders this year. This is not a sign of weakness. It's financial engineering: borrow cheap, buy back stock, increase earnings per share.
The broader pattern: AI-linked companies are tapping capital markets aggressively. Alphabet announced $85B in equity offerings after securing $55B in debt since November. Amazon raised $54B in debt earlier this year. The build-out continues, and the capital is flowing.
Source: CNBC on Nvidia debt
A Satellite Used a Vision-Language Model to Find Things on Its Own
For the first time, an Earth observation satellite identified areas of interest in response to natural language queries without human analysts. YAM-9, built by Loft Orbital, ran Google DeepMind's Gemma 3 vision-language model on a Nvidia Jetson Orin GPU in orbit. Researchers asked it to classify where natural environments meet human development or identify infrastructure around railway hubs. It worked.
The near-term value: satellites can triage data on orbit instead of flooding analysts with raw imagery. The longer-term implication: always-on monitoring where you tell a satellite to "watch this border and alert me when something suspicious happens." This is infrastructure, not science fiction.
Source: TechCrunch on satellite AI
If you run a service business: Build an AI triage layer for customer support. Have AI read the inquiry, pull relevant information from your knowledge base, and draft a response. Human reviews and sends. Start with your 5 most common support questions. That's 40% of your ticket volume handled in 2 minutes instead of 15.
If you manage sales: Map your lead-to-first-contact workflow. Every manual step is a place leads get dropped. Automate form submission to CRM record creation, CRM record to lead enrichment, enrichment to rep assignment, assignment to first email. Use Zapier, Make, or n8n. The tools are cheap. The revenue loss from dropped leads is not.
This week showed AI is now a geopolitical tool, not just a business tool. The Anthropic ban signals that access to frontier models is conditional and can be revoked without warning. Expect more countries and companies to invest in sovereign AI capabilities. Meanwhile, the M&A activity (Salesforce/Fin, NewCore's $66M raise) shows the agent economy is real and consolidating fast. Smart money is betting customer service, sales, and support will be majority-automated within 24 months. Position accordingly.