Get tomorrow's brief in your inbox
Today: Microsoft and OpenAI both launched enterprise AI agents at Build, racing to own business workflows. Google's Gemini Spark revealed how creepy accurate AI assistants can get. Trump signed a voluntary AI pre-release review order after industry pushback.
Microsoft Build 2026 turned into an enterprise AI agent shootout. Microsoft announced Scout, an always-on AI assistant built on OpenClaw that monitors your calendar, email, Teams, and even local traffic to proactively manage your day. OpenAI countered with six new Codex plugins aimed at white-collar work (data analytics, sales, creative production, product design, equity investing, and investment banking). Both companies are betting that AI agents will replace tabs, apps, and repetitive admin work.
Scout is the more aggressive play. It runs in the background, learns your habits, and can call you. Omar Shahine, Microsoft's Scout VP, put it bluntly: "You're going to get a phone call from this assistant. It's a very different type of AI than chat." More than 3,000 Microsoft employees are already using it for expense reports, meeting scheduling, and filling out forms. Scout integrates with Outlook, Teams, and OneDrive, and Microsoft is sandboxing OpenClaw to prevent security disasters (it doesn't have secrets or direct access to Microsoft 365 data).
OpenAI's approach is plugin-based. Each Codex plugin bundles 62 apps and 110 skills tailored to specific roles. The data analytics plugin connects to Snowflake, Databricks Genie, Hex, and Tableau. The sales plugin pulls context from Salesforce, HubSpot, and Outreach to build close plans and prep for customer meetings. The creative production plugin works with Figma, Canva, Shutterstock, and Picsart to turn briefs into campaign assets. Codex also gained a Sites feature that outputs interactive hosted websites instead of just local files, and an Annotations feature for context-specific commands.
If you run a service business, this is the year AI stops being a chatbot and becomes a coworker. Scout and Codex represent two deployment models: Microsoft's persistent assistant that watches everything, and OpenAI's app-integrated copilot that waits for you to summon it. Both require GitHub Copilot subscriptions. Start with whichever fits how you already work, but expect to train it for weeks before it stops feeling like an intern who doesn't know your clients yet.
Microsoft Scout announcement | OpenAI Codex plugins announcement
Freshworks Freddy AI Agent Studio turns service tickets into no-code AI agents. You can build custom agents for employee onboarding, payroll inquiries, or IT support without writing code. The Model Context Protocol Gateway integrates with third-party tools like Notion and ClickUp, so agents can pull context from wherever your docs live. Amerisure's IT analyst used to spend an hour every morning on ticket trends. Now he spends three minutes with Freddy Insights. If you run IT for a small team, this is a fast way to stop answering the same password reset questions manually. Read more
Lindy AI pricing breakdown: Starts at $49.99/month (Plus, two inboxes), $99.99/month (Pro, three inboxes with "computer use" that lets Lindy navigate web apps on your behalf), $199.99/month (Max, five inboxes with 7x usage capacity). Lindy uses a credit system where each task costs 1-3 credits on basic models, 10 credits on advanced models. Overages charge double the standard rate. If you manage back-to-back meetings and want an assistant that drafts replies over iMessage and joins Zoom calls to summarize them, this is built for you. Just watch your credit burn rate. Pricing details
Microsoft Agent Control Specification (ACS): An open-source standard that lets you define policies for what AI agents can and cannot do. Policies check at four interception points (before input, before tool calls, after tool results, before final response). You can allow, block, redact, or require human approval for specific actions. It ships with SDK plugins for LangChain, OpenAI/Anthropic Agents SDKs, AutoGen, CrewAI, Semantic Kernel, and MCP tools. If you're building agents and worried about them calling the wrong API or deleting production data, ACS is the governance layer you need. Details
Slack Activity Hub redesign: Consolidates DMs, mentions, thread replies, and app notifications into a smart inbox. You can create saved views filtered by channel, section, or type, and take bulk actions to clear multiple notifications at once. 80% of pilot users now triage messages from the Activity hub instead of switching tabs. If Slack feels like a firehose, this update is designed to give you inbox zero discipline. Available across all Slack plans now. Announcement
Travelers deployed OpenAI Realtime API countrywide for auto claims. The AI Claim Assistant uses natural conversation to guide customers through first notice of loss, answering policy questions and gathering details. After launching in eight states, Travelers expanded nationwide within two months. 85-90% of customers now complete claim filing through AI without waiting on hold. Patrick Gee, SVP of Auto and Property Claims, said the real-time model's ability to perform under catastrophe-level surges (100,000+ claims in days) is what set it apart. If you have a high-volume call center operation, OpenAI's Realtime API can handle surges without adding headcount. Case study
Debugging AI agent failures: n8n published a three-level framework for finding what went wrong when agents misbehave. Level 1: Tag executions with metadata (trigger type, user ID, model used, outcome) so you can filter to the broken one. Level 2: Trace the decision chain (what data went into context, which tools were called, what parameters were passed, what each tool returned). Level 3: Use external platforms like Arize for deep analysis. Most debugging ends at Level 2 when you discover the agent didn't have a critical piece of data. If you're running agents in production, structured tagging is the step most teams skip and regret later. Read the guide
SBA loan limits doubled: Small businesses can now combine 7(a) and 504 loans for a maximum of $10 million, up from $5 million. Effective July 4. Small manufacturers can access unlimited 504 loans for distinct projects while also tapping $5 million from the 7(a) program. This is the highest financing level in SBA history. If you're capital-intensive (manufacturing, logistics, construction), this expansion opens real runway for production and working capital. You'll need a solid business plan and financial documentation, and combining both loan types adds complexity to repayment schedules. Details
Anthropic confidentially filed for IPO ahead of OpenAI. The company aims to go public as early as this fall, shortly after SpaceX's expected trillion-dollar IPO. Anthropic did not disclose its target valuation. Beating OpenAI to the IPO race matters because it sets pricing expectations for AI companies and signals whether public markets believe AI profit margins are real or hype. For SMBs, this means Anthropic's enterprise API pricing is unlikely to drop anytime soon. CNN report
Florida became the first state to sue OpenAI over child safety risks. The lawsuit targets ChatGPT's alleged failures to protect minors and seeks to label OpenAI as a "public nuisance." Florida Attorney General James Uthmeier said "Sam Altman and ChatGPT have chosen the AI race over the safety and security of our kids." Chatbots are starting to check user ages in response. If you use ChatGPT for customer-facing interactions, expect stricter age verification requirements across all platforms within six months. NPR coverage
Trump signed an executive order for voluntary AI pre-release review. AI companies can submit frontier models to the Commerce Department's Center for AI Standards and Innovation up to 30 days before public release. Google, Microsoft, and xAI agreed to participate. OpenAI and Anthropic were already doing this under Biden. The order explicitly states it's not mandatory licensing or preclearance, but it reflects a shift after Anthropic's Mythos model flagged thousands of vulnerabilities in major operating systems and browsers. For businesses using cutting-edge models, expect slight delays between announcements and public API access. Order details
Microsoft launched Project Solara, an Android-based OS for AI agent gadgets. Two reference designs: a desk concept (Echo Show-like with facial recognition), and a badge concept (wearable with camera and fingerprint scanner for conversation recording and transcription). Microsoft isn't shipping these, but AccuWeather, Best Buy, CVS, and Target are starting pilots. Microsoft picked Android over Windows to run on smaller, lower-power devices while keeping IT management features. If you're planning hardware that runs AI agents (kiosks, wearables, field devices), Solara is the platform to watch. Announcement
Microsoft released MAI-Thinking-1 (reasoning model, 1 trillion parameters with 35B active) and MAI-Code-1-Flash (137B parameters with 5B active, purpose-built for GitHub Copilot). Both trained on clean, commercially licensed data without distillation from third-party models, according to Microsoft. The MAI-Thinking-1 technical paper reveals it's trained on 794 billion web pages after filtering for adult content, piracy domains, and AI-generated content. Simon Willison regrets his initial coverage that misread active parameters as total size. If you care about code model licensing, these still trained on a crawl of the public web. Simon Willison's correction
Webhook automation hack: If you use a tool that isn't in Zapier's directory, check its settings for Developer Settings, API, or Integrations. Look for a field labeled Webhook URL or Endpoint URL. Paste a Zapier webhook URL there, and your Zap will trigger whenever that app sends data. This works for niche tools, custom internal apps, and even QR codes (put the webhook URL in a QR code, scan it to log field inspections automatically). You can store the data in a Google Sheet, send it to Slack, or route it to your CRM. Webhook guide
Process analysis shortcut: Map your lead management flow (how leads are captured, standardized, scored, and nurtured). Look for manual data entry steps. Add a Zap that automatically captures leads and populates your CRM. Process analysis isn't just for manufacturing. Superhuman's customer support team replaced hours of manual copy-pasting between tools with automation, cutting resolution times and freeing agents to focus on customer care. If you're burning five-cup-of-coffee days, your workflows have bottlenecks you can automate. Framework guide
Microsoft Build made it clear: AI agents are moving from experimental to operational. Scout, Codex plugins, Agent Control Specification, and Project Solara all push AI from the chat window into the infrastructure layer. Google's Gemini Spark demo showed how accurate these agents can get (it knew a user's dog's name, kids' ages, wife's dietary preferences, and parents' names by reading email and vet records). That level of context is powerful and creepy. Expect the next six months to focus on security, compliance, and policy controls as businesses deploy agents that touch customer data, financial systems, and production workflows. Watch for mandatory pre-deployment testing requirements, especially in regulated industries.