This Week in Tech: What Agentic AI's Big Week Means for Your Service Business

If you blinked over the last week, you missed a lot. Google held its biggest I/O in years, Salesforce and Anthropic pushed AI coworkers deeper into the tools businesses actually run on, and the open-source world quietly shipped models good enough to run your back office for almost nothing. The headlines are written for engineers and investors, but underneath them is a simpler story for anyone who runs a service business: the cost of automating real work just dropped again, and the tools got a lot easier to point at everyday operations. Here is what happened, and what it actually means for the way you book jobs, follow up with leads, and keep customers happy.

At Google I/O on May 19, Google launched Gemini 3.5 Flash and a new general-purpose agent called Gemini Spark. The Flash model is the headline for operators: it delivers intelligence that rivals the big flagship models but at the speed and price of a lightweight model, outperforming last generation's Pro model on coding and agentic benchmarks. Gemini Spark can reason across the apps you have connected, reading information in one place and taking action in another. Why it matters: the previous generation of AI could write you a nice email; this generation can read the incoming inquiry, check your calendar, draft the quote, and tee up the follow-up across connected systems. And because models like Flash are fast and cheap, running that automation on every single lead is now affordable, not a luxury reserved for enterprises. The PointWake read: this is exactly why we start every engagement with an audit before touching automation. Powerful agents are only as good as the workflow you point them at. If your intake process is messy, an agent just makes the mess faster. Map the process first, then let the AI run it.

Two launches the same week tell the same story from the business-software side. Salesforce introduced Agentforce Coworker, an AI teammate embedded directly into the interfaces your team already uses; it can pull CRM context and take actions in place. And Camunda announced ProcessOS, an AI layer that discovers, re-engineers, and continuously optimizes business processes as governed agentic workflows, with human review built in. Why it matters: the big platforms are no longer treating AI as a chatbot bolted on the side. They are embedding agents into the system of record, the CRM, the workflow engine, where the actual work lives. For a service business, that is the difference between an AI that drafts a reply and an AI that updates the deal, schedules the next step, and flags the human only when judgment is needed. The PointWake read: notice that both launches lead with human review and governance. The goal is not to remove people, it is to remove the busywork so people handle the moments that matter. When we implement CRM automation, we design the same way: let the agent handle the repetitive 80 percent, and route the exceptions to a human with full context.

Anthropic had a loud week too. It shipped enhancements to Claude Managed Agents, including self-hosted sandboxes (so tool execution can run on your own or partner compute) and a preview feature that lets agents securely reach internal systems through an outbound-only encrypted gateway. On the business side, the company is reportedly approaching a $19B annual run rate and hired a marquee AI researcher, signals that serious money is betting agents are the next platform. Why it matters: the technical headline here is really about trust and security. Self-hosted execution and locked-down gateways are what let a small business safely connect an AI agent to its real data, invoices, customer records, scheduling, without shipping everything to a black box. That is the unlock that moves automation from fun demo to something you would actually run your business on. The PointWake read: every automation we build has to answer one question first, where does the data live and who can touch it? The industry catching up on secure, governed agent execution is good news for the cautious, real-world deployments service businesses need.

While the big labs grabbed the spotlight, the open-source world shipped quietly capable models. New releases like GLM-5.1 (under a permissive MIT license that allows commercial use) are now strong enough for serious production work, and they can be run locally using tools like Ollama and wired into visual workflow builders like n8n. Why it matters: this is the cost story. You no longer have to pay per-token to a frontier lab for every task. The smart pattern emerging in 2026 is hybrid: run smaller open models locally for routine, privacy-sensitive work (sorting inquiries, drafting standard replies, tagging records), and reserve the premium frontier models for the jobs where quality really matters. For a budget-conscious service business, that can cut the cost of automation dramatically. The PointWake read: most owners assume AI automation means an expensive monthly SaaS bill forever. It increasingly does not. A well-designed system mixes free or near-free local models with paid horsepower only where it earns its keep.

Finally, a policy note worth watching: governments, the U.S. in particular, are pushing to test major AI models before public release, with several large AI companies agreeing to give regulators early access, while other governments move to adopt frontier models inside official institutions. Why it matters: more oversight tends to mean more documentation, transparency, and clearer rules of the road. For a small business that is mostly reassuring, it pushes the industry toward auditable, accountable systems. But it is also a reminder that as you automate customer communication, you are responsible for what your systems say and do. Compliance basics, clear opt-outs on text and email, honest messaging, records of consent, are not optional. The PointWake read: we bake compliance into every outreach workflow from day one. Every automated SMS sequence we build includes clear opt-out language and a handler to honor it, because the AI sent it is never a defense.

Strip away the launch hype and the through-line is consistent: AI agents that take action, not just answer questions, are now fast, cheap, secure enough, and embedded in the tools you already use. The barrier to automating real operational work has never been lower. But cheaper, more powerful tools do not fix a broken process, they accelerate whatever you point them at. So the move this week is not to rush out and buy the newest agent. It is to get honest about where your time actually leaks: the lead that waited three hours for a reply, the quote that sat in a draft, the customer who never got a follow-up. Map those gaps first, then automate them. Your action item this week: pick the single most repetitive task in your business, the one you would hand off first if you could clone yourself, and write down every step it takes, start to finish. That document is the blueprint an AI agent needs. It is also the first thing we would ask for. If you want help turning it into a working automation, that audit-first approach is exactly what we do at PointWake.