An AI-augmented agency starts with one test, applied to every step of a workflow: does this step follow a rule, or does it need judgment? Rule steps can be written down, checked mechanically and handed to a template, checklist, script or AI-assisted tool. Judgment steps stay with people.
What a Rule Step Looks Like
The test sounds simple, and the work is in applying it honestly. Most steps feel like judgment to the person who has always done them. Ask what the right answer would look like on paper. If you can write it, it is a rule step.
A rule step has a right answer you could write on a checklist. In a delivery review, these are the ones I mean:
- Is the template version current?
- Is the data range correct?
- Is the disclaimer present?
- Do the links work and is the labeling consistent?
- Does the deliverable meet the brand spec?
When my two senior reviewers logged every change they requested for three weeks, about 70% of their time was checks like these. The other 30% was strategic judgment. The seniors were doing a checklist's job. The story behind that log is in my LinkedIn article on the reviewer bottleneck.
What a Judgment Step Looks Like
A judgment step depends on context a checklist cannot hold: whether a recommendation fits this client's market, whether an AEO page structure will serve the question buyers actually ask, whether a client's worry needs a call instead of a report. Output from any tool, including an AI tool, also belongs here until a person has judged it.
A Hypothetical Sort
An invented monthly audit delivery, broken into steps and sorted. Pulling the data: rule. Checking the data range and the template version: rule. Checking links and labeling: rule. Deciding which three findings matter most for this client's market: judgment. Choosing how to word a bad result for a nervous client: judgment. Confirming the final deliverable reads right: judgment, because it covers anything a tool produced.
Three of the six steps are rule steps, and they are the ones that usually eat the most review time because they are tedious and easy to miss. That is the argument for automating them first: not that they are unimportant, but that a checklist or script is more reliable at them than a tired senior.
Three Mistakes to Avoid
- Shipping tool output unreviewed. An AI-assisted draft is a rule-step shortcut, not a judgment-step replacement. Someone still judges it before the client sees it.
- Choosing the tool before writing the rule. A tool cannot enforce a standard nobody has written down. Write the rule first, then see what can run it.
- Counting hours freed instead of tracking where they land. If the freed hours drift into rework, you have moved the problem, not fixed it.
How to Run the Test This Week
- Pick one workflow. Reporting, content briefs or audit delivery all work.
- Log for three weeks. Have reviewers record every change they request.
- Sort the log. Rule violations on one side, judgment calls on the other.
- Automate the rule side first. Link checks, data-range checks and template conformance are the first candidates. Pick the tool after the rule is written down.
In my experience, a mid-sized agency that moves repeatable execution into workflows recovers somewhere around 15-25 hours a week per team. Treat that as a rough rule of thumb and measure your own. For the automation side, see AI SEO automation, and to check where freed hours land, see the utilization benchmarks by team structure.