The AI Agent Race Just Got Real for Small Business

On July 21, OpenAI launched a small business program built around something called ChatGPT Work. Two weeks earlier, Anthropic pulled its own agent, Claude Cowork, off the desktop and put it on the web and on phones.

Neither of these is a chatbot update. Both companies are placing a bet on the same idea: the useful version of AI isn't the one that answers your question well. It's the one that goes and does the thing.

That's a bigger shift than it sounds like.

ChatGPT Work launched July 9 alongside GPT-5.6, and OpenAI says it's already got 10 million people using ChatGPT Work and Codex combined. The small business program layered on top of it includes training sessions, in-person AI academies in a handful of US cities, and a set of partner tools built specifically for owners who aren't going to read a technical manual before Monday morning.

Claude Cowork took a different route to the same place. It started as a desktop-only tool back in January, and Anthropic just expanded it to run in a browser or on a phone, with sessions that keep working even after you close the laptop. Anthropic pulled the numbers from 1.2 million anonymized Cowork sessions across more than 600,000 organizations, and found that over 90 percent of that usage had nothing to do with writing code. People were compiling research, drafting documents, building reports, and working through the kind of recurring operational tasks that pile up in every small business and never quite get prioritized.

Here's where I want to slow down, because the headline version of this story is "AI agents are here," and that's not quite it.

What's actually here is a specific category of task these tools handle well: work with a clear input, a clear output, and no real judgment call in the middle. Pull last week's numbers into a summary. Draft five follow-up emails from a list of names and a template. Watch three competitors' pricing pages and flag changes. That's assembly work. An agent that can hold context across a few hours and touch your calendar, your files, and your inbox is genuinely good at assembly work now, in a way it wasn't a year ago.

What these tools are not good at yet is the judgment sitting on either side of that assembly. Deciding which competitor's move actually matters. Deciding whether that draft email is the right tone for a client who's already annoyed. Deciding whether to make the hire the report says you can afford. Nobody outside Anthropic and OpenAI has run either of these products long enough at scale to tell you different, and neither have I.

So here's the test worth running this week, not next quarter. Pick one task you or someone on your team does on a recurring basis that's mostly assembly and not much judgment. A weekly summary. A first draft of a client update. A scan of five sources for anything relevant to your industry. Hand that one task to ChatGPT Work or Claude Cowork for two weeks. Keep every output. Don't send anything unreviewed. At the end of two weeks, look at what came back and ask whether it saved you real time or just moved the work from "doing it" to "checking it."

That distinction matters more than which company's agent is better. Both are racing to look less like a chat window and more like a colleague, which is presumably why one of them is literally named Cowork. It's a good bit of naming. It doesn't mean the thing has judgment yet.

There's a tension worth naming honestly. When a task gets easier to hand off, the pull is to hand off more of it, not to get better at knowing which parts you should still be doing yourself. A business owner who delegates the assembly work but stops checking the assembled thing hasn't actually saved time. They've just moved the mistake further downstream, to a place where it's harder to catch.

The interesting question this week isn't whether ChatGPT Work or Claude Cowork is the better agent. It's whether, six months from now, the owners using either one will still notice when the thing they handed off comes back wrong.