Ask an AI to change a blue button on your website to black. It changes the code and tells you it's done.
Okay. But is the button actually black?
With access to the right tools, the AI can run the website, open the page and take a screenshot. If the button is still blue, it can go back and fix it. You don't need to sit there sending the next instruction after every step.
That's a verification loop. Do the work, check the result against what you asked for, fix what didn't work. Then stop.
I think this is one of the reasons coding is such a useful place to learn how to work with AI. There are often concrete ways to check the work. Anthropic's guidance on building agents makes this point too: coding agents can use automated tests as feedback, then try again. It also says human review still matters.
So the question is, how do we do the same for a business?
First, what does done look like?
There are two things I'd focus on.
Number one, be clear about the outcome. "Change the button colour" leaves quite a bit to interpretation. "Change the blue button on the home page to black with white text" gives the AI something specific to do.
But sometimes you know the problem without knowing the best solution. So you might say, "The text on this button is difficult to read. Make it more legible."
Now you're asking the AI to work out how. That's useful, especially in areas where you're not the expert. But you still need a way to judge its answer. In this case, a screenshot helps you inspect the appearance. A contrast checker gives you a more useful check of the colours, and you should still test that the button works.
Number two, give it the tools and information to run those checks. An instruction to check the website doesn't help much if the AI can't open it.
Knowing exactly what you want helps. When you don't, let the AI ask questions before it starts. That's what I covered in Get your grill on.
Stop handing out one instruction at a time
We have armies of very smart AIs available to us, and a lot of the time we use them like this. Give it a task. Get an answer. Give it another task.
Imagine if you had to do that with everyone at work. Every step, every day. You'd spend the whole day handing out instructions.
Ideally, we give people a job description, train them on how we do things and agree on what good work looks like. They get on with it, report back and ask for help when something falls outside what they can handle. At least that's the idea, right?
We don't need to pretend AI is a person to borrow something useful from that arrangement.
Give it a defined job, the information it needs and limits on what it can do. Decide what starts the work, what counts as done and when it must come back to you.
That last part matters. "Keep trying until it works" can turn into a very expensive afternoon.
The check needs evidence
Asking the AI "are you sure?" isn't much of a verification layer.
If it says it updated a customer record, check the record. If it produces a sales total, compare it with the source data. If it says it sent something, check whether the sending system recorded it.
In its guide to evaluating agents, Anthropic distinguishes what an agent says happened from what actually changed. It also describes combining ordinary software checks, AI review and human judgement.
For numbers and fixed rules, use a calculation or a check in software where you can. Another AI can help review things like whether a summary makes sense, but it can make mistakes too. People still need to check whether those reviews are any good.
This is where scale becomes a problem. If AI produces work faster than your team can check it, you can end up with a queue of unchecked work. You haven't necessarily saved anyone time.
A weekly sales report, end to end
Let's take a simple example. Say every Monday someone pulls last week's sales figures together for your team meeting. Here's how I'd set up a first experiment.
- Define the job. Prepare a draft report of last week's sales and compare them with the week before. Agree on what counts as a sale, how refunds are handled and which dates to use. Give it a previous report that you know is correct.
- Give it access. Let it read the relevant sales records and save a draft report. It doesn't need permission to edit orders or issue refunds to do that job.
- Set the trigger. Run it every Monday before the meeting. This needs a tool or workflow that can actually run on a schedule. Writing "every Monday" in an ordinary chat isn't enough on its own.
- Check the result. Use fixed calculations to compare the report's figures with the source records for the agreed dates. Flag missing data and unexpected duplicates. Have the AI link its explanation back to the figures, and avoid inventing reasons for a rise or fall.
- Define when to stop. If the checks pass, save the draft for review. If a check fails, allow one attempt to fix the report, then check again. If it still fails, stop and tell the person responsible what's wrong. It shouldn't change the checking rules to make its answer pass.
- Keep a record. Save the source period, the checks, any corrections and the final report. You need to be able to see what happened without reading an entire chat.
Start by running it against an old week where you already know the answers. Include a messy week too. Missing records, refunds, whatever normally trips people up. Have someone review every report at first, then decide which checks can be automated and which still need a person.
You could build parts of this with a spreadsheet and ordinary automation. That's fine. Use AI where interpreting or explaining the information helps. It doesn't need to do every part of the job.
Then prove it's helping
Before you start, measure how long the report takes today. Afterwards, count the time spent checking and fixing it as well as preparing it. Include the cost of the tools.
There's a useful warning in METR's 2025 study. Sixteen experienced developers completed 246 tasks on projects they knew well. With the early-2025 AI tools, they took 19% longer, even though they thought AI had made them faster.
That was a specific group using specific tools. In a February 2026 update, METR said newer tools likely helped more, but problems with who participated and which tasks they submitted made the size of that improvement uncertain.
My takeaway is simple. Feeling more productive isn't enough. Measure your own result.
And a correct report doesn't prove your business is growing. The next question is whether it helps the team make a useful decision, or frees up time they actually put to good use. Those outcomes take longer to judge than whether a button changed colour.
Pick one recurring job. Write down the outcome, the checks and when the AI should stop. Try it on a small scale, look at what went wrong, adjust and run it again. If it's saving time and the work holds up, build from there.
If you're struggling to define that first job, start with Give your AI a mission. Let it ask you the questions before you let it loose.