The person who owns this is not always the person who signs.
Most automation that gets rolled back was not wrong very often. It was wrong in a way nobody caught, on something that mattered, and after that nobody trusted the rest of it. The fix is boring and it works: decide up front which cases a person has to see, make the handoff fast enough that reviewers do not rubber stamp it, and log every decision so you can prove what happened. That design is the difference between automation that survives its first bad week and automation that gets switched off.
Design the review step before you deploy: which cases a person must see, how the handoff works, what gets logged, and the kill switch. The part that decides whether automation survives.
Human in the loop operations design is the work of deciding which cases an automated system is allowed to complete on its own, which must be reviewed by a person, how that review reaches them and in what form, and what is recorded about every decision so the process can be audited and improved.
Human in the loop design sets the boundary between what an automated system decides alone and what a person must approve. It covers the confidence threshold, the review queue and how fast it moves, the escalation path, the audit log, and the kill switch. Thinklytics designs this before a build rather than after, because an automation without it tends to get rolled back the first time it is quietly wrong.
A written boundary: the cases the system completes alone, the cases a person approves, and the reasoning behind where that line sits.
A review step fast enough to be done properly. A queue that takes too long gets approved in bulk, which is worse than no review at all.
An audit trail that records what was decided, by what, on what evidence, so a regulator or an internal reviewer can reconstruct it.
A kill switch with a named owner, and a defined state the process falls back to when it is used.
A person rubber stamping everything. If the reviewer approves 99 of 100 without reading, you have added latency and bought no safety.
A permanent bottleneck. The threshold moves as the evidence accumulates, and it should be reviewed on a schedule rather than left where it was set on day one.
A compliance document written after the build. Designed afterwards, it describes what happened to be built rather than what should have been.
A substitute for accuracy. Review catches the cases the system flags. It does not catch the cases it is confidently wrong about, which is why measurement still matters.
The decision boundary, written down per case type, with the cost of being wrong stated for each.
Review queue design: what the reviewer sees, in what order, and how long a decision should take.
Escalation paths for the cases a first reviewer cannot resolve.
The audit log specification: what is recorded, retained how long, and who can read it.
A kill switch, a named owner, and the fallback state the process returns to.
A sampling plan for the cases that were NOT flagged, because that is the only way to find what the system is confidently wrong about.
A schedule for revisiting the threshold as evidence accumulates.
The queue is too big or the interface is too slow, so reviewing properly is not possible in the time available. The review exists on paper only.
No audit log, or one that records the outcome without the evidence the decision was made on. The second is more common and just as unhelpful.
They are not being shown a boundary, an owner and a kill switch. Those three are usually the whole ask, and they are cheaper to design than to retrofit.
Nobody samples the cases the system did not flag. Review covers what the model doubted, not what it got confidently wrong.
That an automated process deliberately stops and asks a person on certain cases, rather than completing everything it can. The design work is deciding which cases those are, how the person is shown enough context to answer quickly, and what gets recorded. Done well it is a small share of volume. Done badly it is either everything, which saves nothing, or nothing, which is what gets the system switched off.
It comes from the cost of being wrong on that case type, not from a model's confidence score. A misread vendor name is a nuisance; a wrong coverage decision is not. The two should not share a threshold, and setting one number for the whole process is the most common mistake we see.
It adds a step on the cases that go to review, and the aim is to make that a small and shrinking share. In practice the thing that slows a process down is not the review, it is a badly built queue where a reviewer has to open three systems to answer one question. That is a design problem and it is fixable.
What was decided, by what, on what evidence, and who could have intervened. The evidence part is what gets left out and it is the part that matters, because an outcome without its inputs cannot be reviewed, only accepted. Retention and access should be settled at design time rather than negotiated after an incident.
A named person in the business that owns the process, not the team that built the system. If using it requires an engineer, it will not be used at the moment it is needed. The fallback state matters too: what the process does while it is off should be defined before it is ever switched off.
By sampling the cases it did not flag, on a schedule, and reviewing them properly. This is the step teams skip, and it is the only way to find a systematic error, because by definition the system is not raising its hand on those. Without it you know the process is running, not that it is right.
Yes, and it is common. Retrofitting is more work than designing it up front because the logging usually has to change, but it is considerably less work than rebuilding after a rollback. Gartner expects 40% of companies to roll back or shut down autonomous agents by 2027 on governance gaps, and this is the gap.
For a single process, two to four weeks to produce the boundary, the queue design, the logging specification and the sampling plan. It is deliberately quick, because it is meant to happen before a build rather than become a project of its own.
18 to 5 of every 100: Share of claims needing manual review, after the routing was redesigned around what the model was actually confident about. $6.2M in suspicious claims flagged a year. (Mid-Market P&C Carrier)
4.2 days to 6 hours: Prior authorisation review on AI-handled cases, with reviewers seeing only the cases that needed judgement. 18,000 more cases a month, no extra staff. (BlueCross BlueShield Affiliate)
The people who normally own this problem are CISO, CIO, Compliance leader. They are the ones who start the search, and they are not always the person who signs for the work.
A buyer usually describes the situation this way: The pilot works and security won't sign it off.
The spend is normally gated by the CFO or executive sponsor, whose question is: Is the improvement worth the investment? What that person needs to approve it is a business case with stated assumptions and measurable outcomes.