Want this content delivered right to your inbox?

Rebuilding the Rituals Governance Depends On

human in the loop ai

Rebuilding the Rituals Governance Depends On

Every review ritual in your organization was built on a quiet assumption. Work arrives at the speed a person can make it. The weekly design critique, the two-approver rule, the Friday read of the client deck all assume the reviewer can see most of what gets produced. AI broke that assumption. Output now arrives at machine speed, and the review rituals around it still run at human speed. Something has to give, and in most organizations it gives in one of two bad directions. Teams either rubber-stamp, so the ritual survives in name only, or they bottleneck, so the ritual survives and the work stops. Neither is human in the loop AI. Both are human-shaped rituals failing under a load they were never designed to carry. This is the third part of the framework we laid out in Govern AI Like You Hire: give AI the same least-privilege access, the same defined roles, and the same review rigor you give a new hire. Access and roles get most of the attention. Review rigor is where governance actually lives or dies, because it is the part that has to happen every day.

The Two Ways Review Breaks

The bottleneck is the visible failure. Review queues back up, senior people become the constraint, and AI-generated work sits waiting for a human who never has time to read it. We traced that pattern through approval gates in Your Processes Are Human-Shaped. The short version: writing got faster, judging did not. The rubber stamp is the quieter failure, and it is more dangerous because it looks like success. The queue moves. Approvals happen. Nobody is actually checking. Researchers have studied this for decades under the name automation bias, the tendency to over-rely on automated output in place of vigilant checking. A systematic review in the Journal of the American Medical Informatics Association pooled studies of clinical decision support and found that when the system gave wrong advice, it raised the risk of an incorrect decision by 26 percent. The same review notes that higher workloads and more complex tasks are thought to increase that reliance. Read that second finding again. The condition that makes rubber-stamping more likely is exactly the condition AI creates: more volume landing on the same people. How you frame the AI matters too. Researchers at the BCG Henderson Institute studied more than 1,200 managers and found that when AI was framed as an employee, managers identified 18 percent fewer errors, and their sense of individual accountability for errors dropped by 9 percentage points. The full study appeared in Harvard Business Review the same day. That finding sharpens our own framework rather than undercutting it. Governing AI like you hire means borrowing the system: scoped access, a written role, review that matches the stakes. It does not mean treating the agent as a colleague whose work you can trust on reputation. The moment a reviewer starts thinking “the agent has this,” accountability leaks out of the ritual. A polished surface makes it worse, which we covered in Polished Is Not Done.

Human in the Loop Is a Design Choice, Not a Slogan

“Human in the loop” has become a phrase people say to end a conversation about risk. It should start one. Which human? Which loop? Reviewing what, how often, with what authority to stop it? The serious governance frameworks treat oversight as something you design, not something you declare. NIST’s AI Risk Management Framework describes human-AI configurations as spanning “from fully autonomous to fully manual,” and states plainly that some AI systems may not require human oversight while others specifically do. The EU AI Act makes the same point in law for high-risk systems. Article 14 requires that oversight measures be “commensurate with the risks, level of autonomy and context of use.” It also requires that the people doing the overseeing remain aware of the tendency to over-rely on the output, and that they can disregard, override, or stop the system. Commensurate is the key word. Oversight is not one setting. It is a dial, and your job is to set it deliberately for each kind of work. There is a people problem hiding in the phrase, too. In one of our AI ways-of-working masterminds this summer, a participant observed that nobody wants to be “in the loop.” People want to drive. Being in the loop sounds like being a checkpoint the machine has to pass through. I have started describing the shift differently: it is less human in the loop and more human on the loop, someone who owns the work, watches it run, and steps in when it matters. That is not a downgrade. It is the job of a manager.

Match the Review to the Stakes

Here is the design that holds up. Sort AI-assisted work by what happens if it is wrong, then give each tier a different kind of review. Three tiers cover most of what organizations actually do. The first tier is approval. A named human approves before the action happens. Reserve this for work that is high-impact, external, or hard to reverse: anything that moves money, changes access, reaches a customer or regulator, or deletes something. This matches the OWASP guidance on excessive agency, which recommends requiring human approval for high-impact actions before they are taken. Approval is expensive, so the discipline is keeping this tier small. If everything requires approval, nothing really gets approved. The second tier is sampling. The work ships, and a human reviews a defined slice of it on a regular rhythm. This fits work that is reversible, internal, and high-volume: first-draft summaries, ticket triage, routine data cleanup, meeting notes. The sample has to be real. Pick the size and cadence in advance, rotate who reviews, and write down what “wrong” looks like so the reviewer is checking against a standard rather than a feeling. A good sample also over-weights the edge cases, because that is where errors hide. The third tier is audit. Nobody reviews individual outputs. Instead, someone periodically reviews the pattern: what the system did, what it escalated, where humans overrode it, and whether any of that is drifting. NIST suggests that data on how often and why humans overrule AI output in deployed systems may be useful to collect and analyze. Treat that override rate as a vital sign. If it drops to zero, either the system got very good or your reviewers stopped looking. You need to know which.

TierWhen it fitsWhat the human does
ApprovalHigh-impact, external, or irreversibleApproves each action before it happens
SamplingReversible, internal, high-volumeReviews a defined slice on a set cadence
AuditMature, well-bounded, low-riskReviews patterns, overrides, and drift periodically

The tiers are not a ranking of how much you care. They are a way of spending a scarce resource, careful human attention, where it changes outcomes. Spread evenly across everything, that attention becomes the rubber stamp. Concentrated where the stakes are, it becomes governance.

human in the loop ai

Let the Ritual Relax as Trust Is Earned

A new hire does not stay on probation forever. The review rigor is highest at the start and loosens as the person proves out. AI work should move between tiers the same way, in both directions. It is the principle behind Govern AI Like You Hire, and review is where it gets enforced. The best public evidence for how this plays out comes from people already doing it. Anthropic analyzed how users of its Claude Code agent oversee it in practice. New users ran with full auto-approval in roughly 20 percent of sessions, and by 750 sessions that rose to over 40 percent. At the same time, experienced users interrupted the agent more often, about 9 percent of turns compared with 5 percent for new users. The authors’ reading is that people shift “from approving individual actions to monitoring what the agent does and intervening when needed,” and that effective oversight means “being in a position to intervene when it matters.” That is the whole arc in one dataset. Experienced users did not stop overseeing. They changed how they oversaw, from gatekeeping every step to watching and stepping in. Make that arc explicit instead of letting it happen by drift. Write promotion criteria for each workflow: how many weeks in a tier, what error rate in the sample, which kinds of mistakes are disqualifying. Write demotion triggers too. A model change, a new data source, a serious miss, or a shift in who the output reaches should move a workflow back up a tier automatically. Trust that can only grow is not trust. It is complacency on a schedule. This is the same logic as the AI driver’s license in the anchor post. Access is earned by demonstrated competence, and it can be suspended.

Review Is a Team Ritual, Not a Checkpoint

Most organizations treat review as a control: a box someone ticks on the way to shipping. Under AI volume, that framing is how rubber-stamping becomes normal. A lone reviewer, a long queue, and no shared standard is the exact setup the automation bias research warns about. Treat review as a ritual instead, something a team does together on purpose, with a facilitator’s attention to who speaks and what gets surfaced. In practice, that looks like a short standing session where the team looks at the week’s sample together. Someone pulls ten outputs. The group reads them, names what is good, names what is off, and updates the written standard when they disagree. That session does three things a solo checkpoint cannot. It calibrates judgment across reviewers, so “acceptable” means the same thing on Tuesday and Friday. It surfaces drift early, because ten people notice a pattern one tired person misses. And it spreads review skill to the people who did not have it yet. It also restores accountability, which is the thing both the BCG study and the automation bias literature say slips away. The JAMIA review found that a person’s own internal sense of being accountable reduced automation bias errors, while externally imposed accountability had mixed results. You cannot mandate internal accountability. You can build a room where people feel it. The ritual starts before review, too. In one of our masterminds, a practitioner described colleagues generating documents and analyses and “chucking them over the wall,” so the review burden grew larger than the build burden. Their standing rule for the team was simple: do not hand something off until you have reviewed it enough to know it is worth handing off. That rule moves the first pass of review to the person with the most context, and it makes the formal ritual smaller and sharper. This is why rebuilding rituals is one of the competencies we see separating teams that make AI stick from teams that stall. It is also the hardest, because people have equated their craft with the ritual. Changing how review works feels like losing the standard. It is not. Love the purpose, not the process. The purpose of review was always to catch what matters before it costs you. The process was just the shape that purpose took when work arrived at human speed.

What to Do This Month

Pick one team with real AI volume. Inventory every kind of AI-assisted output it produces, and for each one ask a single question: what happens if this is wrong and nobody notices for a week? Use the answers to assign a tier. Anything that moves money, touches a customer, or cannot be undone goes to approval. Reversible, internal, high-volume work goes to sampling. Only mature, well-bounded workflows go to audit, and only once they have earned it. Then write three things down for each workflow: who owns the review, what the cadence is, and what would move it up or down a tier. If you cannot name the owner, you do not have human in the loop AI. You have a hope. Finally, put the sampling review on the calendar as a team session, not a solo task. Thirty minutes a week, ten outputs, one shared standard. Track the override rate from day one. The organizations that rebuild review around stakes will scale AI without losing the plot, because their people will know exactly where to look and why. The rest will keep choosing between a queue nobody can clear and a stamp nobody believes. The difference is not how good the models get. It is whether anyone designed the ritual. If your review rituals are buckling under AI volume and you want to redesign them with the people who run them, talk with us. Facilitating teams through exactly that redesign is the work we do.