We put a human approval step into our automation and nobody uses it. What went wrong?
Approval queues fail for four predictable reasons: the volume exceeds the approver's attention, the item does not carry enough context to decide, approving takes more effort than doing the task by hand, and nothing happens when the backlog grows. Fix them by batching similar items, showing only what changed, making approval a single action, and auto-approving the confident cases while sampling them after the fact.
Volume is the first thing that kills it
A human approval step assumes a human has attention available. If the automation produces more items per day than the approver can meaningfully consider, they will not consider them. They will approve in bulk without reading, which is worse than no approval step because it manufactures a false record of oversight.
The honest arithmetic: estimate how long a careful review of one item takes, multiply by daily volume, and compare that to the time the approver actually has. If the answer exceeds a modest slice of their day, the design is wrong before it ships.
Context, effort and consequence
- Show the decision, not the record. An approver needs to see what is about to change and why, not a full customer record they have to read to find the difference.
- Make approving cheaper than doing it manually. If someone has to open three tabs to verify, they will just do the task themselves and stop using the queue.
- Batch by type. Twenty similar items reviewed together take a fraction of the time of twenty mixed ones, because the reviewer holds one mental model instead of twenty.
- Decide what happens to unapproved items. Expire them, escalate them, or auto-approve them after a window. Leaving it undefined means they sit forever and the queue becomes a graveyard.
Tier by confidence instead of approving everything
Blanket approval on every action is a sign the design has not decided what it trusts. A better structure has three tiers: high-confidence, low-consequence actions execute automatically and are sampled afterward; ambiguous ones go to the queue; high-consequence ones always require a person regardless of confidence.
That last tier does not shrink over time and should not. Writing to a financial record, contacting a customer about a complaint, or anything touching an employment or safety decision stays with a person permanently. AI evaluation supports human judgment in those cases, it does not replace it, which is exactly how we frame rubric-based evaluation and rep coaching.
Measure the queue, not just the automation
Three numbers reveal whether the approval step is real: median time to decision, approval rate, and the share of items approved without being opened. A very high approval rate suggests the queue is not filtering anything and those item types could move to auto-approve with sampling. A low one suggests the automation itself needs work.
Rising time-to-decision is the early warning that the step is being abandoned. Catch it there, before someone quietly turns it off. Designing this properly is a real part of an operational AI rollout, not an afterthought bolted on for comfort.
Topics: human in the loop · approval workflow · UX · governance
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.