A coffee machine call came in. Old model: asset says coffee, dispatch sends a coffee tech. Truck rolls. Job closes or it doesn’t. Next call, same logic.
That is not what happened.
Our AI triage agent read the call, then six months of history at that store. Seven other drain issues. Floor drain. HVAC drain. Fountain. Water problems logged as one-offs because nobody had been looking at six months of tickets at machine speed.
A human digging that deep burns an hour or two. The agent did it in seconds.
One call looked like a coffee machine. The pattern looked like a drain.
Our plumbing agent took the signal and recommended a camera down the line. That path costs more than a coffee drain cleaning. It would blow past the customer’s not-to-exceed. That is not a billing detail you discover after the work is done. That is a permission boundary.
Then someone still had to decide.
Stay on the old path and send a coffee tech because the asset field says coffee? Or follow the agents, put a plumber on a camera inspection, and stop for customer approval before you spend past the limit they set?
A leader called the customer’s leader. They walked the scenario. The customer approved the camera.
The camera found a broken pipe. That break was backing up the site. What looked like a string of unrelated service calls was one root cause. It became a project. The project fixed it for real.
The agents found the signal. A human owned the decision. The customer got a lasting fix.
That is human judgment in the AI loop.
The scarce resource
AI can accelerate a process. Compress cycle time. Run continuous improvement at a clock speed that used to be fantasy. That part is real.
What it does not do is carry accountability.
When work gets faster, the bottleneck moves. The calendar piles up where someone still has to say yes or no. That is not a failure of automation. That is what an operating model with integrity looks like.
The job is not to replace the human. The job is to clear the mundane and elevate the judgment that still needs a name on it: customer impact, risk, brand, release.
The recommendation is not the decision.
Same loop, other work
The drain story is facilities. The physics are not.
Software
A defect lands in the queue. Old model: days. Someone picks it up, reproduces it, finds the code, writes the fix, waits for review, waits for a sprint boundary, ships when the calendar allows.
That is not what has to happen now.
An agent can take the ticket, find the bug, pull a branch, write the fix, run the tests, and open a pull request. Supervisory agents can review that work against standards: security, drift, adherence, the things your best engineer looks for. First passes that used to burn human hours now run in minutes.
Then someone still has to decide.
Merge because the agents passed it? Or refuse the ship because the change still carries production risk, customer impact, or a call the agents cannot own?
The agents found the defect and drafted the fix. A human owned the release.
That seat is not permanent. Encode what the human rejects. Train the next model on those rejects. Run the supervisory pass until the misses you still see are a risk you will carry. Then the software cycle can run without a person on every pull request. You choose that. Autonomy is a decision about what is left over, not a slogan about getting humans out.
Same shape as the drain call, until the standard is thick enough to move the human off that gate. Machine speed on the pattern and the draft. Human judgment on the decision that still carries the outcome. And what the human rejects becomes the next standard the agents run before a person ever sees the pull request.
Content
A launch needs assets. Images. Short-form video. Captions. Social drafts. Old model: a queue of designers, editors, and revisions. Days of back-and-forth for work that still may miss the brand.
That is not what has to happen now.
AI can generate the first hundred versions while you are still clarifying the brief. I have watched that up close. My son is using these tools to create launch content and to move faster in real work. Production that used to take a week can land in an afternoon.
Then someone still has to decide.
Publish because it looks finished? Or kill nine out of ten because the voice is off, the image is cheap, the claim is wrong, or it does not look like you?
The model produced the options. A human owned the brand.
Same shape again. Accelerate the mundane. Elevate the judgment. The rejected drafts are not waste. They tighten the brief for the next pass.
Three domains. Same loop. Machine speed on what used to bury people in volume. Human judgment on the decision that still carries the name.
What judgment actually is
Judgment in the AI loop is not a slower version of the model.
In the field, it is the authority to change the trade, stop for customer approval, and enroll a customer before spend crosses the line they set. In software, it is the authority to refuse a merge. In content, it is the authority to refuse a publish. Across all of it, it is the authority to say the recommendation is not good enough.
Judgment is not a staff meeting. Not a dashboard. Not an approve button with no reject muscle. Not treating a recommendation as the decision because the recommendation usually works.
A gate you cannot say no through is not judgment. It is theater with a human face on it.
Encode. Automate. Own. Learn.
The stack is simple. The discipline is not.
Encode
Write down what your best person looks for.
If the standard lives only in someone’s head, the system invents a plausible next step and the organization calls it progress. Encode first.
When six months of site history shows a drainage pattern behind a coffee machine call, the path is plumbing, not coffee tech. When the plumbing agent recommends a camera past the customer’s not-to-exceed, the next step is a human call for approval before work. In software, write what “good enough to merge” means before agents start reviewing other agents. In content, write brand voice, image quality, and what never ships before you generate a hundred drafts.
You are not documenting for the binder. You are encoding judgment so the routine can move without pretending every signal is a decision.
Automate
Then let the machine run the first passes at machine speed.
Triage. Site-history correlation. Plumbing recommendations. Draft production, tests, supervisory review. Image generation, short-form drafts, first-cut edits.
That is the work that used to consume the day so a strong operator never got to the exception. Automation is not the end of the loop. It is how you clear the mundane so judgment has somewhere to stand.
Own
A human owns the exceptions and the consequential authorizations.
Trade path when triage surfaces a site pattern the old model would have missed. Approval when the camera inspection crosses the customer’s not-to-exceed. A leader-to-leader call when trust and spend are on the line. Merge and release when the change still carries liability. Publish when the brand is on the line.
The agents can read history in seconds. They can recommend a camera. They can draft code and content. They cannot enroll the customer, carry the contract, own the release, or own the brand.
The dangerous stamp is not the sloppy draft everyone can see. It is the approval that follows a recommendation that has been right for weeks. The confidence is earned. The reading stops anyway.
If you are not reading, you are not judging. If you cannot reject, you are not owning.
Learn
Every judgment that changes the outcome should change the next pass.
The plumbing path updates the routing standard. The customer approval on the camera becomes the playbook for the next NTE breach. The broken pipe that explained seven one-off drain calls becomes a pattern the agents look for earlier. The failed code review becomes a gate the agents run before a human ever sees the pull request. The rejected draft tightens the brief for the next batch.
Continuous improvement is not a poster. It is learning that alters how the loop runs tomorrow.
Encode. Automate. Own. Learn.
The job
Go back to the drain call.
The agents saw six months of tickets in seconds. A human changed the trade, called the customer, and stopped before the spend crossed the line. Seven one-off calls became one broken pipe. That save became the rule for the next one.
Run that call often enough and the shape changes. The agent’s recommendation is as good as the person on the phone. Sometimes better. The human hears it and nods before the history is even on the screen.
The nod feels like trust. It is attention leaving.
Pilots met this with autopilot long before these models. When the machine is usually right, people stop scanning the instruments. New pilots do it. Veterans do it. More hours in the seat do not cure it. The system earns the trust, and the eyes come off the panel.
You will see the same thing in a review queue. Someone has been right to agree for a month. The approve click gets faster. The history, the diff, the photos stay closed. A person’s name is still on the workflow. The judgment is gone.
Two moves follow from that. Take the person off the work the model has already matched. Keep a person on the work that still needs a name, and make that work hard to wave through.
Take them off the matched work
Write down what a good reviewer used to reject. Let the machine refuse anything that breaks that written standard. Then delete the signature on every ordinary pass.
Leave the chair empty. An approval agent sitting there is the same stamp with a new face. If a model can say yes, that yes is the standard already running. It checks the work. It does not own the outcome.
This is how the software cycle leaves the human behind, once you accept the misses that remain. Same choice on a routine drain path, a routine edit, a routine draft. The standard holds, so the person goes.
Keep them where a name is still required
Some calls are still a person’s. Enrolling the customer. Crossing the spend line that customer set. Carrying the contract. Publishing in a voice the written rules do not yet cover. Releasing a change the standard has not seen.
Those stay. The queue they sit in has to stay hard, or the nod comes back.
Send three kinds of work, and stop sending the rest:
- Cases the model is unsure about.
- Cases where two reviews of the same work disagree.
- A sample of the cases it was sure about, so the easy yeses do not disappear.
A queue made only of easy agrees teaches people to stop reading.
Then reread real approvals with the recommendation hidden. A second person, or the same person on a later day, looks at the ticket, the diff, or the photos alone. If that read keeps finding nothing, the model has earned the gate and the person should leave this one too. If it finds misses the first person waved through, the gate is a stamp. Fix the queue. Do not add more volume to it.
Watch what people open, not how often they agree. Agreeing more can mean the model got better. It can also mean the human stopped looking. One fast yes means nothing. A month of faster yeses, with the record opened less, means the scan is dying.
One case is different. A person still has to see every item, because you cannot yet tell the routine from the exception. Make them decide from the evidence before the recommendation appears. They record a lean. Then they see what the model said. A written explanation after the fact will not produce that read. People dislike the extra step. Keep this queue short, or they will find a way around it.
Accelerate the mundane. Elevate the human. Remove the human where the standard already holds.
Who owns the judgment on the workflow that matters most, and what would tell you they are still using it?