Nitin Murali is VP of Supply Chain Excellence at Gallo and founder of Category 2 (cat2.ai).
Imagine this. It’s Monday morning and you’re a supply planner. Over the weekend, the new AI system triaged 340 exceptions and dumped the 52 it couldn’t resolve into your queue. You have until the 10 a.m. S&OE call to clear them. You approve most without reading them, because the system is usually right and there’s no time to find out when it isn’t.
Your manager sees the dashboard: 96% of recommendations accepted, and reports it up as adoption. Her director presents it to the steering committee as proof the AI investment works. Touch-less planning is up, exception volume is down, headcount is flat.
And somewhere above all this sits an SVP who, three months later, is staring at an inventory write-off and asking a question nobody in the chain can answer: Who decided this?
The planner will say she approved what the system recommended. The system did exactly what it was scored on: service level and forecast accuracy. Obsolescence wasn’t in its objective function. That risk belonged to another team, optimizing other numbers. The write-off didn’t happen inside any function. It happened at the seam between them. The money is gone, the decision happened but no human ever made it.
The technology worked exactly as designed. That’s what makes it uncomfortable. The failure is in the org chart: we dropped intelligence into processes built thirty years ago and expected the same accountability to absorb it. It didn’t.
The standard answer is “human in the loop.” That phrase is the problem. It’s a posture, not a job. It told the planner nothing about which judgments were hers.
In the deployments that actually stuck, three roles emerged. Nobody hired for them. Nobody wrote the job descriptions. It’s time we named them.
The Signal Architect
They own what the organization sees: which signals matter, what they mean in context and when they’re trustworthy enough to act on. A dashboard showing inventory levels is data. A signal tells you what changed, why it matters, what to do about it, what happened last time and how confident the system is, so a human can apply judgment. A data point is not a signal, and in the AI era that confusion gets expensive: machines act on whatever you feed them, at scale, without the human who used to fill in the context.
Every operation has people who do this informally. The planner who knows which reports to ignore. The supervisor who can read a production line by the sound of it. A supply chain is a being that can’t tell you what it feels, and these people are the ones who hear it anyway. The signal architect makes that tacit knowledge explicit so machines can carry it.
The Judgment Engineer
They design where humans show up in an AI-augmented workflow and what they decide when they get there. Not whether humans are involved, but where, exactly.
Take materials planning exceptions. An AI can triage hundreds overnight. The judgment engineering question is which ones it resolves alone, which ones it recommends a path for and which ones land on a human’s desk with full context. Those boundaries get drawn around value to the organization, not to a function, because every AI agent inside a silo will optimize its own metric and quietly bleed value at the seams.
Most organizations have no one doing this deliberately. The boundaries get set by accident, by whoever configured the tool, a strange place to leave the most consequential design decision in the system.
The Decision Owner
They are the named human on every consequential decision in an AI-augmented process. A person. Not a committee, not a function, not “the system.”
This sounds bureaucratic. It’s the opposite. When a name is on a decision, the decision gets better. The owner asks harder questions because their judgment is on the line, not the algorithm’s. When something goes wrong, there’s a person who can explain what they knew and why. That’s what makes AI auditable, and what makes adoption stick. It’s the oldest role of the three. AI just makes the absence of one impossible to hide.
If you lead an operation, start here. Pick one decision-heavy process and map who currently owns each judgment inside it. Be honest, because for most processes, the honest answer is nobody. The spreadsheet owns it. The legacy parameter owns it. The person who left two years ago owns it. Then assign the three roles as hats before they become headcount.
The AI talent conversation is fixated on hiring machine learning engineers, and that’s the wrong scarcity. The models are getting cheaper and better on their own. What’s scarce is the design of human judgment around them.
Run That Monday Again
The AI still triaged 340 exceptions overnight. But the 12 in the planner’s queue aren’t leftovers it couldn’t handle. They’re there because a judgment engineer decided, months ago, which decisions carry enough consequence that a human must make them. The other 328 are resolved, logged and auditable. She isn’t cleaning up after the machine. She’s making the decisions the organization chose to keep human.
On the third one, she rejects a buy recommendation. It protects service level, the metric the demand agent is scored on. But she knows about the packaging change coming in the fall, and that inventory would land as next year’s write-off. Her queue exists for exactly this: decisions that cross-functional seams belong to a human accountable for total value, not a model accountable for one number. The signal architect connects the packaging-transition signal to the demand agent, so next Monday, the machine knows what she knew.
Her manager’s dashboard no longer shows acceptance rate. It shows decision quality and override patterns, where humans are still smarter than the system, and where the system has earned more autonomy.
And when the SVP asks who decided, there’s a name, a rationale and a person who can stand behind both.
That’s the difference between deploying AI and absorbing it.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
