Archive 003: Automate the Mailroom
The Question
A workflow can use several AI tools and still depend heavily on a person.
One system does the research. Another checks the claims. Something fails. A person notices the problem, carries it back to the right place, explains what needs to change, waits for the correction, and then sends it through review again.
From the outside, that can look automated.
But someone is still doing the routing.
So when is the human providing judgment — and when are they just carrying work from one step to another?
What We Tried
We looked at one of our own multi-step AI workflows.
Different systems had different jobs.
One produced the work.
Another verified it.
If something failed, the problem went back to the system responsible for that part.
The correction came back through review.
Then the work moved forward.
That sounds organized.
But there was still a hidden job in the middle:
A person had to notice every failure and manually move the work to the next place.
The workflow had specialized roles.
It had review.
It had correction.
What it did not have was reliable routing.
What Happened
Once we separated those things, the problem became easier to see.
Some parts of the workflow genuinely needed a person.
A person may need to decide whether two sources actually disagree.
A person may need to judge whether a correction changes the meaning of the work.
A person may need to decide whether something is good enough to publish, send to a client, or use in a real decision.
Those are judgment calls.
But other steps were much simpler.
A failed claim did not need a person to decide which system should fix it.
The workflow already knew who created that section.
A corrected section did not need a person to decide whether it should be checked again.
That was already part of the process.
The person was not adding judgment.
They were acting as the switchboard.
The Unexpected Part
That distinction changed the way we thought about automation.
The goal was not to remove the human from the workflow.
The goal was to remove the human from the parts of the workflow that did not require human judgment.
The rule became pretty simple.
If the problem is obvious, send it back to be fixed.
If the fix needs to be checked again, check it again.
If something is unclear, disputed, or actually needs a decision, bring in a person.
People should handle the decisions. The workflow should handle the traffic.
The person stays involved where they are useful.
They stop carrying routine correction traffic between steps.
The person should decide what needs judgment. They should not have to act like the workflow’s mailroom.
Why It Matters
AI workflows are often described as automated because several systems are connected together.
That can be misleading.
If a person still has to:
watch every step,
notice every failure,
decide where the failure belongs,
move the work,
restart the next step,
and keep track of what has already been checked,
then much of the workflow is still manual.
The AI may be doing the individual tasks.
The person is still running the process.
That does not automatically make the workflow bad.
Sometimes human oversight is exactly what keeps the work reliable.
But it is worth asking whether the person is there because judgment is needed or because the system does not know how to route its own work.
Those are very different problems.
Takeaway
A good AI workflow does not remove people from every step.
It removes unnecessary human handling.
Ask:
Where does the workflow genuinely need judgment?
And then:
Where is a person only moving work from one place to another?
Keep the human where decisions matter.
Automate the mailroom.
