New AI capabilities this fall: tools, browsers and connectors, in plain English

The models themselves keep getting better, but that’s not the part I’d pay attention to this year. The bigger change is what they can reach. A model that can only chat is a smart notepad. A model that can look things up in your systems, fill in a form, or file a record is a coworker, and that needs a little more thought.

Here are the three capabilities that matter most for a small business right now, in plain English.

1. Tools: the model calls your software

Tool use means you describe a few actions to the model (“look up a customer,” “check the calendar,” “create a draft invoice”) and it decides when to call them and with what details. Your code runs the action and hands the result back.

This is the workhorse. Almost every useful automation I build is a model plus three or four tools. It’s predictable because the model can only do what the tools allow. If there’s no “delete customer” tool, it can’t delete a customer, no matter what anyone types at it.

What I’d watch: tools that write. Reading a calendar is low risk. Booking on it, emailing a client, or changing an invoice should go through a person first, at least until you have a few weeks of evidence it gets those right. The owner’s inbox is a good example of where I draw that line.

2. Computer use: the model operates a screen

Some software has no API, which means there’s no clean way for a program to talk to it. Think old vendor portals, government sites, some industry-specific tools. Models can now work these the way a person does: look at a screenshot, move the pointer, click, type. Anthropic documents this as a computer use tool for Claude (Anthropic).

It’s genuinely useful for the right job, like pulling a report out of a portal every Monday. BUT it’s slower than a direct connection, it can be thrown off by a layout change, and anything on that screen is something the model can read. I treat it as the option of last resort. If there’s an API or an export, use that first.

3. Connectors: one standard plug

Not long ago, connecting an AI app to your files, your CRM or your calendar meant a custom integration for every pairing. The Model Context Protocol changes that. It’s an open standard for connecting AI applications to outside systems, and its own docs describe it as “a USB-C port for AI applications” (MCP).

For an owner, that mostly means two things:

• More of the apps you already use can plug into AI assistants without custom work.
• When we build something custom for you, like a connector to your job management system, it can work with more than one AI app instead of being locked to one.

What I’d watch: each connector is a new door into your data. Before you turn one on, look at what it can read and what it can change. Read-only is a fine default.

What hasn’t changed

• The model still makes mistakes, especially with numbers and dates. Anything that matters gets checked.
• Garbage in, garbage out. A connector to a messy CRM gives you confident answers about messy data.
• The best first project is still small and boring. One job, measured, as I laid out in the first 30 days.

Agents, briefly

You’ll hear “agents” a lot. Most of the time it just means a model that can use tools in a loop: look something up, decide, act, check, repeat. That’s real and useful. It’s also where costs and mistakes compound, because every loop is another round trip. I wrote about what actually works in AI agents for small business.

Where I’d start

If you’re curious about any of this, start with a read-only connection to one system you use every day, and ask it questions you already know the answers to. You’ll learn quickly how much to trust it.

If you’d like help picking that first connection, send me a note. Please let me know if you have any questions, thanks!