AI Weekly: AI Learns to Use Your Computer — March 23-29, 2026

This was the week AI stopped just talking and started clicking. Anthropic gave Claude the ability to open apps and drive a browser on your own machine, and Google shipped a voice model fast enough to hold a real conversation. A benchmark launch the same week was a useful reality check on how far these agents still have to go.

The Deep Dive: AI Learns to Use Your Computer

Claude can now open apps and fill spreadsheets

On March 23, Anthropic added Computer Use to Cowork and Claude Code. Claude can open applications, navigate a browser and fill in spreadsheets directly on your computer, rather than just telling you how to do it. At launch it was macOS only.

This is a real shift. Most AI tools give you instructions and leave the clicking to you. An agent that does the clicking can take on the repetitive multi-app admin that eats a small team's day, like copying data between a website and a spreadsheet.

Google's voice model gets fast enough to feel natural

On March 26, Google released Gemini 3.1 Flash Live, a real-time audio and vision model with roughly 960 millisecond response time. That is quick enough that a spoken back-and-forth stops feeling like a walkie-talkie and starts feeling like a phone call.

Cohere also open-sourced Cohere Transcribe the same day, a free transcription model that topped the accuracy leaderboard and runs on a normal GPU. Good, cheap voice tools are arriving fast.

Why this matters:

• Agents that control your computer can absorb the boring admin work no one wants to do, not just advise on it.
• Fast, cheap voice AI makes natural phone assistants and support lines realistic for small businesses.
• The same week, ARC-AGI-3 launched and top models scored under 0.4% while humans solved every task, a blunt reminder these agents still fail at anything genuinely novel.

What This Means for Australian SMEs

Computer-use agents are the first version of a genuinely useful office assistant, but treat this as a supervised trial, not a set-and-forget tool. Give it a low-stakes, repetitive job first, like reconciling a simple report, and watch what it does.

The voice news matters if you handle a lot of calls. A natural-sounding assistant that books appointments or answers common questions is now within reach for a small operator, not just a call centre. But the ARC-AGI-3 results are the caveat worth remembering: these tools are brilliant at repetitive, well-defined work and hopeless at problems they have never seen.

The practical takeaway for this week:

• Pick one repetitive admin task and trial a computer-use agent on it under close supervision before trusting it with anything that matters.
• If phone calls are a bottleneck, start scoping a voice assistant for your most common, most predictable enquiries.
• Keep a human in the loop for anything novel or high-stakes. The benchmarks say these agents are not ready to think for themselves.

Knowing which jobs are safe to hand an agent, and which need a person, is exactly what we help clients sort out at IntelliAgent. If you want help running a sensible trial before you commit, get in touch.