Automation Geeks Weekly | 17 March 2026

What happened: Microsoft announced Wave 3 of Microsoft 365 Copilot alongside two new products: Agent 365, going generally available on 1 May at $15 per user per month, and a new top-tier bundle called the Frontier Suite at $99 per user. Claude from Anthropic is now also available inside Copilot's main chat, not just as a bolt-on.

What it is: Agent 365 is Microsoft's product for deploying AI agents that can take actions across your Microsoft tools, not just answer questions. Think of it as buying a digital member of staff who lives inside Outlook, Teams, Word, and Excel, and can carry out multi-step work on your behalf.

What it means: The "this is too expensive or complicated for my business" objection is getting harder to make. At $15 per user, AI agents are now priced like a SaaS add-on, not an enterprise IT project. The businesses that start figuring out what they want agents to own will be a year ahead of everyone who waits.

Try this: Write down three tasks someone on your team does every week that follow the same pattern every time. Client intake, report prep, proposal first drafts. Rank them by time cost. The top one is your first agent candidate.

What happened: BuzzFeed reported a full-year loss of $57.3 million for 2025, largely tied to its AI content strategy. The company bet on AI-generated articles and quiz responses. Its audience left. Its stock hit $0.70.

What it is: A cautionary example of AI adoption driven by cost reduction rather than value creation. BuzzFeed used AI to produce more content cheaply, without asking whether cheaper content was actually what their audience wanted.

What it means: The question is never "can AI do this cheaper?" It is "does doing this cheaper make the outcome better or worse for the person receiving it?" For service businesses, the danger zone is using AI to produce client-facing work that looks complete but feels hollow. Your clients hired you because of your judgment and relationship, not your output volume.

Try this: Before you automate any client-facing output, ask: if my best client knew this was AI-generated, would they feel well-served or short-changed? If the honest answer is short-changed, keep a human in that loop. Use AI behind the scenes instead.

What happened: At its GTC conference this week, NVIDIA announced the NVIDIA Agent Toolkit and OpenShell, an open-source runtime that gives autonomous AI agents guardrails: policy-based controls over what they can access, what networks they can touch, and what data they can see.

What it is: OpenShell is infrastructure software that sits underneath AI agents and enforces rules about what they are and are not allowed to do. The analogy is a new hire with a manager who approves their access levels before they can touch anything important.

What it means: One of the biggest blockers to real AI agent adoption has been trust. "What if it does something I did not authorise?" This is the industry's answer. Expect the tools you already use (Salesforce, ServiceNow, and others are already involved) to start offering tighter agent permissions over the next few months. That shifts the conversation from "is it safe?" to "what do I actually want it to do?"

Try this: Map out one workflow you would hand to an AI agent if you trusted it completely. Then note every step where you would want a human checkpoint or an approval gate. That map is your agent design spec, and it is worth having before the tools catch up.