6 min read
I built Protecht Health, a health app, by myself, so I also run its marketing by myself. That is only possible because a handful of AI agents do the repetitive parts. It is also where I learned the most about what AI does not fix, so I want to be specific about both.
What the agents do
I think of them as jobs, not as a clever system. Each one has a narrow brief and a place where it stops.
A drafting agent writes blog posts. It pulls real sources, drafts the post in the house format, makes the header image, and saves a draft in WordPress. A review step sits in the middle: I see the title, the claims and the sources before anything is saved, and I am the only one who can publish.
A social agent turns each post into versions for four platforms. Same idea, four different captions and images, each with a tracked link so I can see which one sent a click.
A radar agent sweeps the social platforms every morning. It finds conversations where a useful reply would help, drafts the reply, and gives me a sheet. It never posts. I read the sheet in one block and approve or kill each draft.
A metrics agent pulls the numbers once a week. Search Console, Google Analytics, App Store Connect and Apple Search Ads land in one place on Monday morning, and a second pass turns them into a short read of what changed.
A task agent keeps the list honest. Everything I owe someone, and everything an agent is waiting on me for, lives in one task manager.
The rule that makes it safe
Agents draft. I send. That is the whole governance model, and it is boring on purpose. Nothing goes out under my name, to a customer, a reporter or a clinic, until I have read it. When an agent can schedule something, it only schedules what I already approved, and it does nothing if I got there first.
That rule costs me time every day. It is also the reason I am comfortable letting the agents run while I am doing something else.
What the numbers said
Here is the part that changed my plan.
In September I did a content review. The blog had been publishing daily, which the agents made easy. Thirty daily posts had brought in 58 search clicks, and about half of those came from a single page. The agents had made volume cheap. Volume was never the problem.
So the plan changed from daily posts to two a week, each aimed at a question people actually search for, each linked to the page that matters most. The agents did not tell me to do that. I read the numbers and decided.
The weekly metrics told a second, less comfortable story. About 58 percent of purchases that people started in the app were failing, in the versions before 2.4. More ad money would only have bought more people who hit that failure. So I paused the plan to buy more installs and fixed where installs already leak, starting with logging the reason each purchase fails.
That is the pattern I keep seeing. AI made the start of every task faster. It did not make anything easier to finish. The finishing is reading the number, deciding what it means, and doing the unglamorous fix.
What it costs
It is not free, and it is not hands-off.
- My review time. Every draft still needs a person who knows the subject. A health topic needs a source check on every claim, and I do it.
- Setup and upkeep. Each agent has a written brief, and the briefs drift when the business changes. Someone has to own them.
- A fast way to do the wrong thing. An agent will happily produce thirty posts that do not matter. It does not know that, so a person has to.
If you run a small team
Three things I would copy:
- Give each agent one job and one stopping point. “Drafts, never sends” beats “handles marketing.”
- Put the review step where the risk is. For me that is any claim about health. For you it might be a price or a contract term.
- Read one set of numbers on a fixed day. An agent that reports is worth more than an agent that acts, because it keeps you honest about whether the work is working.
If you want to see how I would set that up for a team that is not mine, the free one-hour session is the lowest-risk start. It is online, there is no pitch, and you leave knowing whether a paid Workflow Triage makes sense.
Where the numbers come from
- 30 daily posts and 58 search clicks: my own Search Console review, Sep 13, 2026.
- Agent briefs and the review step: my own workflow files for Protecht Health.
Work with me
Something keeps slipping?
A Workflow Triage names the three workflows worth fixing, in order, in one week. $1,500, credited toward a Pilot.