Senthil Info

automating business processes, honestly

Where Humans Stay in the Loop

Some decisions should not be made without a person, and the reasons differ. Knowing which reason applies matters, because each implies a different kind of human involvement — and one of them is frequently implemented as a rubber stamp that provides no protection at all.

The five categories

One. Legally or contractually required. Certain approvals, sign-offs and regulated decisions must have an accountable individual. Non-negotiable, and the requirement is usually specific about who rather than about how. For a product-side reference related to workforce operations, Monitask also covers online timesheets.

Two. Consequential and irreversible. Payments, deletions, terminations, anything that cannot be undone. The automation can prepare it; a person releases it. For broader background and an independent point of comparison, see World Economic Forum.

Three. Judgement without a stateable rule. If the rule cannot be written, it does not exist yet — it lives in someone's assessment. Automating the wrapper around it adds a handoff and removes nothing.

Four. Cases where being wrong is rare and expensive. The dangerous quadrant: errors that surface only at audit, when the automation has been repeating them for months.

Five. Anything affecting a person's standing. Employment, credit, benefits, access. Beyond any regulatory requirement, these need a nameable human who can explain the decision to the person affected.

The rubber stamp problem

The failure mode that matters most, because it looks like compliance.

A human approval step where the human approves everything is not oversight. It is a delay with a name attached, and it provides the appearance of control without the substance.

It happens predictably. The approver sees a hundred items a day, ninety-nine are fine, the interface has an approve button, and the volume makes real examination impossible. Within weeks the step is reflexive.

Three things distinguish real oversight from a stamp:

The approver has time. A rate that permits actual examination, which means the automation should be routing fewer items to them, not all of them.

They have the information to decide. The screen shows what the decision needs, not just the outcome to confirm.

And rejection actually happens. If nothing has been rejected in six months, either the automation is perfect or the step is decorative. Measure the rejection rate — it is the only evidence that the control is real.

Designing the loop properly

Route by risk, not by volume. Everything above a threshold, everything in an unusual category, a random sample of the rest. A person examining 5% carefully is worth more than one glancing at 100%.

Include a random sample deliberately. Without it, the reviewer only ever sees flagged cases and loses any sense of what normal looks like — the same degradation that affects exception handlers.

Make rejection easy and logged with a reason. The reason codes tell you where the automation is weak, and they are the feedback route that keeps it improving.

And give the reviewer authority to stop the whole thing. Someone who can only reject individual items cannot respond to a systematic fault.

The cost of the loop

Worth stating honestly, because it changes the business case.

A human step caps throughput at whatever that person can process, which is frequently below the automation's capacity and is the real limit on the saving.

And it must be staffed on holidays and sick days, or the queue backs up. A single named approver is a single point of failure that nobody costs in.

Both belong in the calculation. The saving applies to the automated portion, and a mandatory review step means the human portion is larger than the exception rate alone suggests.

The short version