Don't Waste a Good Complaint
Turn friction into leverage.
Human in the loop became human in every loop.
The company bought an agent to remove a bottleneck. Now every action it takes waits on an executive approval. Congratulations. The executive is the agent's intern.
Someone who spent months running agents on real work named the constraint plainly:
That was March. The same complaint resurfaced in June and July: one unavailable reviewer stalls the workflow, and repeated approvals become rubber stamps.
Blanket review feels safe, but repetitive approval invites automation bias. People begin accepting the system's answer because it usually looks right. The audit trail proves someone clicked approve. It does not prove they exercised judgment.
Review everything and you may lose speed without gaining the safety you think you bought. Review nothing and eventually an error reaches a customer, a contract, or a regulated process.
The problem is not too much trust or too little trust. It is undifferentiated review.
F2L Score: 82 / 100
F2L means Friction to Leverage. It asks whether a complaint is noise or evidence of an operating problem worth redesigning.
Input | Score | Why |
|---|---|---|
Timeliness of the pain | 28 / 35 | The same complaint appeared repeatedly from March through July. |
Economic cost | 31 / 35 | Executive attention and queue delay recur with every gated action. |
Spread | 23 / 30 | The pattern crosses agent workflows, but broad prevalence data is still limited. |
Total | 82 / 100 | Material, expensive, and ready for governed redesign. |
Under 40 is noise. From 40–70, redesign it. Over 70, it is worth funding and governing this quarter.
Jupiter Peak's guidance on why AI projects fail before production makes the operating point: use automated evaluation, audit logs, exception handling, and human review where judgment actually matters.
The Redesign
Tier the action by consequence, not confidence. One workflow can contain a reversible internal step and an irreversible contractual one. If a vendor says the model will tell you when it is unsure, ask what happens when it is confidently wrong.
Pre-authorize the boundary. Define what the system may access, what it may do, and what evidence it must retain. Access and external action are different decisions.
Run checks first. Validate required fields, sources, policy rules, and expected outputs before a person sees the work. The Reddit operator did this with pull requests and automated checks so the human reviewed a clean diff, not a raw agent session.
Route exceptions. Send the right evidence to one named reviewer with authority and a deadline.
Write the retirement rule when you create the review. For low-consequence work, define the override threshold, sample size, quality limit, and review date that would justify removing the checkpoint. An edit counts as disagreement. Rare, irreversible actions may still warrant review even when overrides are uncommon.
Expand autonomy on evidence, not comfort. Run the agent beside the current process, compare what it would do with what people actually do, and set action-specific promotion thresholds before the test begins. Put a date on the decision so parallel run does not become permanent limbo.
The goal is not maximum autonomy. It is minimum unnecessary management dependency.
What leaders should measure
Override and edit rate. What share of reviews changes the output? An edit is a disagreement, not a pass. Near zero on low-consequence work makes the checkpoint a retirement candidate. A rising rate suggests the boundary may be wrong.
Wait time versus work time. If the agent finishes in seconds and sits in a queue for two days, that ratio is the real automation number.
Escape rate. Track errors that cleared review and reached a customer, record, contract, or regulated process. This is the number that justifies the checkpoints you keep.
Autonomy coverage. Track the share of actions completing without human touch by consequence tier and work type. A healthy average can hide one category that is failing.
These are workflow leading indicators. They map most directly to delivery cycle time and management attention in the Economic Baseline. They do not replace the quarterly Economic Test.
Your turn
Pick one approval your team performs every week. What percentage of the time does the review actually change the output?
Most leaders have never measured it. Reply with the number, or the guess. I'll tell you whether it looks like a control or a queue.
If you're deciding where agents can move faster without adding risk, I'll walk the workflow with you. Thirty minutes, no deck.
One Hard Thing
Ride your mountain bike uphill at a lift-served park.
The chairlift is running. The trail still goes up. Choose your suffering.