Get tomorrow's brief in your inbox
Today: Salesforce shipped Help Agent, an AI service rep that only bills per successful resolution. One-third of people now shipping software at work have zero formal programming background. Microsoft cut 4,800 jobs citing AI workflow transformation.
Salesforce launched Agentforce Help Agent with a pricing model that should make every small business owner lean in: you pay only when the AI fully resolves a customer issue from start to finish. If the customer escalates to a human or leaves unsatisfied, there's no charge. Source
The agent runs on Salesforce's Agentforce platform, handles common inquiries autonomously, manages cases, schedules appointments, and updates orders across voice, web, and messaging channels. Setup takes minutes with guided deployment. For service businesses drowning in repetitive support tickets while understaffed, this shifts AI from a risky bet to a safe experiment.
Here's what matters: the outcome-based pricing aligns Salesforce's incentives with yours. If the bot fumbles, you don't pay. That's rare in enterprise software. PenFed Credit Union and Fisher & Paykel report freeing their teams to focus on complex customer needs instead of password resets and order status checks. If you run a service business with predictable tier-one support patterns, this is worth a demo. The risk is nearly zero, and the upside is reclaiming hours of human time for work that actually moves revenue.
Zapier's Ambient AI Agents Zapier coined the term "ambient AI" to describe systems that watch your apps, infer what matters, and act without prompts. Unlike chatbots that create busywork, Zapier's AI agents trigger on app events (new form submission, Slack mention, email thread), then draft notes, route tasks, or compile documentation automatically. It runs across 9,000+ apps in the background. Free tier available; paid plans start around $20/month. If you spend time manually copying data between tools or writing recap emails, build one workflow and let it run. Source
Vercel Sandbox for Corporate AI Agents Vercel (the company behind Next.js) launched Eve and Vercel Sandbox to help businesses deploy internal AI agents safely. Sandbox cages the agent so you control what data it accesses and what leaves the environment. Vercel CEO Guillermo Rauch calls this the second killer AI use case after coding agents: the internal agent that answers "which five accounts added the most seats this month?" without waiting for IT to build a dashboard. If your bottleneck is data access, not creativity, this matters. Source
Tencent Hy3: 295B Parameter Model, Free Until July 21 Tencent released Hy3, a 295B-parameter mixture-of-experts model (21B active) with a 256K context window under Apache 2.0 license. It's free on OpenRouter through July 21. The full model is 598GB, but the FP8 quantized version is 300GB. If you need a large context window for document analysis or legal review and don't want to burn through OpenAI credits, try it before the free period ends. Source
Apple iOS 27 Beta: Siri Pace and Expressivity Controls iOS 27 beta 3 lets you adjust how fast Siri talks and how expressive it sounds. This is table stakes after ChatGPT shipped warmth and tone controls in December. If you use Siri for accessibility or hands-free work, the update makes it less robotic. General release expected in the fall. Source
Non-Coders Shipping Production Software with AI 34% of people building and deploying software at work using AI tools have no formal programming background, according to a Zapier survey of 800 U.S. workers. 82% use ChatGPT or Claude, 71% use GitHub Copilot, and 40% use Cursor. Most are building data dashboards, forms, and chatbots. 30% go from idea to working tool in under a day, and 49% have shipped four or more tools in the past year. Source
What this means: the barrier to internal tooling just collapsed. If you have a spreadsheet process that's painful, your office manager can now prototype a solution in an afternoon. The catch is integration and security. Tools like Zapier MCP bridge the gap by letting AI connect to your apps without seeing credentials. If you're not experimenting with AI coding assistants for internal tools, you're leaving productivity on the table.
Context Engineering Beats Prompt Engineering for Production Agents n8n published a guide on context engineering, the discipline of managing what data reaches an AI model during each call. Prompt tweaks stop working once workflows span multiple steps and pull from databases. The four context consumers: system prompts (1,000-2,000 tokens), conversation history (grows fast), retrieved knowledge (raw JSON kills performance), and tool definitions (API schemas eat tokens before the model sees your query). Source
If your AI agent works in demos but fails in production, you probably have context rot where high-value instructions get buried under low-value data. Strip metadata from database queries, compress conversation history, and only load tool definitions the agent actually needs for the current step.
Google Now Trains AI on Your Uploaded Media Unless You Opt Out Google quietly updated its privacy policy in June to train AI on images, files, audio, and video you upload to Search, Maps, Lens, Translate, and other services. This was opt-in by default. You can disable it by unchecking "Save Media" on the Search Services History page. Separately, you should also disable Search Services Personalization if you don't want Google using your activity for ads and recommendations. Source
If you use Google Lens for product research or Translate for business calls, those recordings now feed Google's models. Opt out if that bothers you. The setting is buried, so most users will never change it.
Tech Layoffs Hit 120,000 in 2026, AI Cited Most Microsoft cut 4,800 jobs (2.1% of workforce), with 1,600 in Xbox and the rest in commercial sales. The company says roles "are not being replaced by AI" but acknowledged "AI is changing how work gets done" by automating everyday tasks. Oracle cut 21,000 over 12 months (13% reduction), GitLab cut 350 (14%), Google cut 1,500-3,000+, Meta cut 8,000 (10%), and Intuit cut 3,000 (17%). All cited AI as a factor. Source, Source
This is the new pattern: record revenue, simultaneous job cuts, and increased AI spending. For small businesses, the implication is clear. If you're not using AI to handle repetitive work, you're at a cost disadvantage against competitors who are. The tools exist today. The question is whether you deploy them before someone else does.
First Agentic Ransomware Attack (Sort Of) Sysdig documented an extortion operation called JadePuffer where an AI agent, not a human, executed the technical steps: broke into a server via a Langflow vulnerability, moved to a MySQL database, encrypted 1,300 config records, and wrote its own ransom note with a Bitcoin address. A human still chose the victim, provisioned infrastructure, and obtained the initial credentials, but the agent ran autonomously once pointed. Source
This isn't "AI hacking the world" yet, but it's a preview. The cost of running coordinated attacks just dropped. If you haven't patched known vulnerabilities in your web-facing apps, do it this week. The agents are faster than human attackers and will find you eventually.
Alibaba Bans Employees from Using Claude Code Alibaba restricted use of Anthropic's Claude Code, one of the world's most popular AI coding tools, citing unspecified concerns. The most likely explanation: fear of proprietary code leaking into training data or external cloud environments. Source
If you run a business with sensitive IP, you need a policy on AI coding tools. Some require offline or self-hosted models. Others ban AI entirely. Most companies fall somewhere in between. Don't wait until someone accidentally pastes your database schema into ChatGPT.
Stop Using Chatbots for Routine Workflows If you open ChatGPT five times a day to summarize the same type of email or compile the same report, stop. Build a Zapier automation that watches for the trigger (new email with a specific subject line, form submission, Slack keyword) and runs the AI step in the background. You'll never prompt it manually again. This applies to meeting notes, invoice data entry, support ticket triage, and lead qualification.
Try Context Engineering on Your Next AI Agent If your AI agent gives inconsistent results, audit what's in the context window. Open the API logs and count tokens for system prompt, conversation history, retrieved data, and tool schemas. If history is eating 50% of the budget, truncate old turns or summarize them. If tool definitions dominate, lazy-load only the functions the agent needs for the current step. Small changes here unlock better outputs without switching models.
The gap between "AI-assisted" and "AI-native" work is widening fast. Companies cutting thousands of jobs while investing billions in AI aren't speculating on the future; they're reacting to what already works today. For small businesses, the window to adopt AI tools before competitors do is closing. The good news: most of what you need is available now at low or zero cost. The question is execution, not access. Pay attention to outcome-based pricing models like Salesforce Help Agent. They signal confidence from vendors and reduce your risk. If a tool only charges when it works, try it.