AI Agents Can Do Real Work Now. The Question Is Which Work.

Every business has a task that runs on a person’s time. Someone pulls numbers from email, a spreadsheet, and a supplier portal each week to build the same report. Someone keys invoices into the accounting system one at a time. The work is not hard. It is repetitive, it happens often, and it quietly ties up a capable person for hours. If you have tried to fix it before, you have probably found that the tools did not quite join up, and the manual steps stayed.

Something changed this year that makes this worth another look. The question is no longer whether AI can do the work. It is which work you should hand it, and who still signs off.

What Changed This Year

AI agents have moved past chat. They can now carry out multi-step work on their own computer, and keep going in the background while you do other things.

The shift that made this practical is quieter than a bigger model. It is the interface. A desktop app can now open a real web browser and use it, clicking through a page the way a person would. Claude Code added a built-in browser in July 2026, and OpenAI’s Codex does the same. That means an agent can reach the web apps and files where your work actually lives, instead of only answering questions in a chat box.

How To Think About The Options

There is a growing list of these tools, and comparing them feature by feature rarely helps. Two questions are enough to place any of them.

First, how much should a person stay in the loop? Some agents wait for you at each step. Others run on their own and report back when they are done.

Second, where should the work happen? Some agents sit on your own computer and work directly with your files. Claude Cowork is one example. Others reach across your connected apps, your email, your drive, and your systems, and do the work there. ChatGPT Work, launched in July 2026, is OpenAI’s version of that. Others again run in the background on a schedule, so a repeated job just happens without anyone starting it.

The tool matters less than your answer to those two questions for the workflow in front of you.

Where This Actually Pays Off

The value is not that you are “using AI”. It is that a task which used to take a person half a day now takes a few minutes to check. The upside is plain: faster repeat work, fewer handoffs, earlier reports, and a process that runs the same way every time.

Here is a simple test. If a workflow is frequent, manual, and time-sensitive, it is a candidate. Accounts payable is a clear example. Invoices arrive constantly, someone keys them in by hand, and a late or wrong payment has a real cost. An agent can read each invoice, prepare the entry, and flag anything unusual. A person then checks the exceptions and approves before anything is posted.

Where It Goes Wrong

Most teams connect a tool quickly, then find the workflow still breaks under real conditions. Connecting a tool is not the same as designing a workflow.

The approval steps, the exceptions, the checks, and the order things happen in all have to match how your business already runs. If they do not, a more capable agent just creates problems faster. Work gets sent on without the right sign-off. Exceptions get handled differently each time. And when a run fails, it is not clear who owns it.

This is exactly where a person still belongs. The agent can carry out the steps. A person still owns the boundaries: where AI may act, what it may produce, and who approves the result before it is trusted.

The Decision In Front Of You

So the question is smaller and more practical than “which AI tool should we buy”. It is which one workflow to improve first, how much a person should stay in the loop, and where a person signs off before the work is trusted.

A tool connected without that design becomes one more thing to supervise, not one less. Name the workflow, name the controls, and today’s very capable agents turn into something your business can rely on.