A Working AI Workflow Can Still Produce Work You Do Not Want

The automation that taught me this

At Future of Agencies in Cebu, I described a workflow I called Agent Pulse. One agent watched what competitors were publishing. Another turned the research into posts. Another handled publication.

The chain worked. I also told the room that my Instagram had ended up full of AI slop.

That is an uncomfortable but useful distinction. A completed workflow tells you that the steps executed. It does not tell you whether the research was worth sharing, whether the writing reflected your judgment, or whether the audience received something useful.

Check four different things

This review worksheet develops the lesson from that example. Keep the checks separate so you can repair the part that failed.

Give review a place in the workflow

In the same talk, I described using connected inboxes to prepare a report and draft replies. I prefer to read a reply before it goes out. That preference should exist in the workflow as an actual approval step.

For a publishing process, define who reviews the draft and what they check. If the reviewer is unavailable, the item should remain a draft. A missing approval should not silently become permission.

For research, keep the source beside the claim. For writing, keep an example of the intended voice. For an action, keep a record of what was approved. Those details make corrections faster because the reviewer can see how the result was produced.

Turn repeated corrections into instructions

If you repeatedly delete vague openings, make that part of the writing standard. If the system confuses a trial signup with a paying customer, document the distinction and add a test that catches it.

A useful correction explains the principle. “Remove this sentence” fixes one draft. “Do not describe a requested appointment as confirmed until the team confirms it” improves the next draft too.

Save the reviewed example with the reason it passes. That gives the next run something more concrete than “make it better.”

Measure the work after the automation

For a small trial, track how many outputs were accepted, how many needed correction, how long review took and whether any errors reached the audience. Compare that with the previous way of doing the job.

A workflow that creates fifty drafts but leaves you rewriting forty of them may be moving effort rather than reducing it. A smaller number of useful, reviewed outputs may serve the business better.

These are proposed measures for your own trial, not performance figures from Agent Pulse. The purpose is to find out whether your system is helping before you increase the volume.

Your next action: Take five recent outputs from one workflow. Check their sources, usefulness, voice and final action. Fix the recurring problem before scheduling more runs.

The talk behind this guide

These lessons come from my October 2026 Future of Agencies session in Cebu. Read my talk recap for the teaching context. The organizer's 2026 agenda lists the session. The worksheets here expand on that teaching for use in your own business.