I approved the email. Then I changed the attachment. Which version should the agent send?

Here’s my rule. Show me the final package. Get my decision on that version. Then bring back the receipt.

Claude Duet got me thinking about this. Anthropic says it can prepare work in the background and ask before important actions. That opens up real possibilities for business owners.

We turned that announcement into a case study. This is public-source analysis, not a product test.

Let’s walk through an example. An agent drafts a client email. I approve it. Then I change the recipient and swap the attachment.

Watch those two fields. The person receiving the information changed. The information leaving the business changed. My earlier approval should not silently carry over.

The proposed workflow is simple. Prepare the final action. Show the recipient, message and attachment. Ask for a decision on that exact package. Recheck permission before sending. Then verify what actually happened.

Our case study proposes four checks.

Change the draft or attachment. Does the workflow pause for a fresh decision?

Change the recipient. Does the approval show the final destination?

Revoke permission. Does the action stop and explain why?

Let the send time out. Does the system check the first attempt before trying again?

That last one matters. A timeout does not tell you whether the email was sent.

For the receipt, I want the final recipient, the message identifier and the time. Enough to connect the action I approved with the result we can actually verify.

Those are proposed acceptance checks. We haven’t run them against Duet, and we’re not claiming the product failed them.

That’s the focus behind Whitewater AIOS and the WWAIOS framework. Make the evidence, permissions and decision behind an AI action visible and defensible. Reduce distortion between reality and decision.

Take the four checks in the case study and apply them to one workflow you’re considering. Keep the decision visible. Bring back the receipt.