Story

The Automation I Had to Turn Off

By Felix Maru · November 9, 2026 · 6 min read

The HR admin dropped the request on a Tuesday morning like it was nothing: "The leave approval process is a disaster. Three people have to sign off, and requests sit in inboxes for a week. Can you automate it?"

This was 2019 at a large international NGO. The process was genuinely painful. A staff member filed a leave request. It went to their manager. If the manager didn't respond in a reasonable time, the request drifted. HR had no visibility. Staff kept pinging. Managers missed things because they were traveling or just swamped. Felix had a solution in his head before the conversation was over.

He built it over the next few days. Leave request submitted via a form connected to a spreadsheet. A Zapier workflow picked it up, routed it to the direct manager via email with an approve or decline link. If the manager took no action after 48 hours, the request escalated to HR. If HR took no action after another 24 hours, the system auto-approved it. Everyone got notified at every stage. The process that had taken a week now resolved in under three days.

HR loved it. The team loved it. Felix felt good about it.

For about four months, it worked perfectly.

The Program Review

Twice a year, the NGO ran a multi-country program review: a tense, week-long exercise where field teams flew in, presented their numbers, defended their decisions, and planned the next period. These reviews mattered. They influenced funding decisions, staffing calls, everything. Senior leadership was present. Country directors were present. The energy was high and the schedule was packed.

The review for one of the larger sub-regions was set for a Monday. The Friday before it, the country director for that region got pulled into an emergency trip to one of the field sites. She left Friday afternoon, unreachable until Sunday evening at the earliest.

She had four direct reports with leave requests pending in the system. The 48-hour escalation timer had already been running since Wednesday.

Felix didn't know any of this. The automation didn't either. It just knew the rules: 48 hours with no response from the manager, escalate to HR. Another 24 hours with no response from HR, auto-approve.

HR was also stretched thin that week, preparing materials for the review itself.

By Sunday morning, all four leave requests had been auto-approved.

Monday

Two of those four staff members held critical roles in the review. One was the data analyst who owned the program's outcome tracking spreadsheet. The other was the admin coordinator responsible for logistics during the review week itself.

They both showed up Monday, slightly confused. They had their leave approved. They had made plans. But nobody had intended for them to actually be away during review week. The country director had planned to deny those requests as soon as she got back. HR had assumed someone else was handling it. The automation had done exactly what it was designed to do.

The data analyst stayed. She felt she couldn't leave mid-crisis. The admin coordinator tried to defer her leave by a week and mostly succeeded, but not before two hours of back-and-forth that nobody needed on a review Monday. The country director was mortified. HR was apologetic but clear: the system had technically functioned correctly.

Felix got called in that afternoon. Not to be blamed, exactly. More to explain what had happened. He pulled up the logs. The workflow had run perfectly. Every trigger had fired on schedule. Every notification had been sent. The approvals were entirely legitimate by the system's own rules.

That was the problem.

What the Automation Couldn't See

The automation knew when a manager had responded. It did not know why a manager hadn't responded. Those are very different things.

A manager who is traveling and unreachable is not the same as a manager who forgot. A leave request submitted two days before a critical all-hands event is not the same as a request submitted in a quiet week. The system had no way to hold context. It just had timers and rules.

What HR had been doing manually, before the automation existed, was looking at the calendar before approving anything. If a leave request came in and there was something significant happening that week, they would hold it and check with the manager. It was informal. It was slow. But it was judgment, and judgment was what the auto-approve tier had replaced.

The fix was not complicated. Felix turned off the auto-approve tier entirely. The revised flow went like this: leave request submitted, routes to manager, 48 hours with no response sends a reminder and also flags HR. HR then makes the decision manually with the manager's context in mind, not automatically. The system's job became to get the right person the information they needed to decide, not to decide for them.

The speed difference between the old version and the new one was about 24 hours. That was the cost of adding the human back in.

It was absolutely worth it.

The Rule I Took Away

There is a category of decision that looks simple until it is wrong. Leave approvals looked simple. They are mostly routine. The edge cases are rare. But the edge cases carry real consequences, and those consequences require someone to be thinking about context, not just checking a box.

The mistake I made was treating "no manager response" as a neutral signal that the system could resolve on its own. In practice, "no response" can mean anything. The manager is traveling. The manager is ill. The manager is overwhelmed. The manager doesn't know something the approval would affect. Those are all different situations and they deserve different handling.

Auto-approval felt like efficiency. What it actually was: the removal of the last person who might have caught something.

I still believe deeply in automation. The routing, the reminders, the notifications, the escalation paths: all of it is valuable. What I rebuilt was automation that informs humans rather than automation that replaces human judgment at the point where judgment is actually needed.

Some steps in a workflow exist to be touched by a person. Not because a person is faster. Not because the tool isn't capable. But because the step genuinely requires someone to look at the full picture and make a call. Automation's job is to make sure the right person gets to that step without wading through ten minutes of admin first.

That is a different job than making the decision for them. The difference matters enormously when it goes wrong.

If something in your stack is making decisions automatically and you don't fully understand what "no input" means in every possible context: that is worth looking at. The question isn't whether the automation is working. It's whether the thing it's deciding is actually safe to decide without a human in the loop.

Often the answer is yes. Sometimes it isn't. The hard part is being honest about which one you're building.

Want to think through where your approval workflows need a human checkpoint? Drop me a line.

Share 𝕏 in ✉

Comments