Lead, Not Loop
Why "human in the loop" is the wrong metaphor for the AI era — and what should replace it.
Walk into any AI governance discussion in 2026 and you will hear the same three words within five minutes: human in the loop. It is the phrase regulators reach for, the phrase consultants sell, the phrase boards repeat to reassure themselves that someone, somewhere, is still in charge.
It sounds responsible. It is, in fact, the wrong way to think about how leaders should engage with AI — and the cost of getting the metaphor wrong is already showing up in the data.
How the loop took over
“Human in the loop” was not coined for AI. It comes from control-systems engineering, where it described an operator monitoring an automated process and intervening when something drifted out of tolerance. The human’s job was binary: approve, reject, or adjust. The system did the work; the human was the safety catch.
When AI ethics discussions needed a vocabulary for oversight, they borrowed it wholesale. And the borrowed metaphor brought its assumptions with it. In a loop, the system runs the process. The human enters when summoned, evaluates a discrete decision, and exits. Authorship belongs to the machine. The human is the exception handler.
This was a useful pattern when AI did narrow, predictable tasks. It is a dangerous pattern now.
What the loop hides
Recent adoption research keeps surfacing the same problem. MIT’s Project NANDA, the Politecnico di Milano’s Osservatorio Artificial Intelligence, Deloitte’s State of Generative AI in the Enterprise, the Swiss AI Report — different methodologies, same finding. Organizations that frame AI adoption as “putting humans in the loop” are producing three predictable failures:
The reactive posture. When the human’s role is to approve or reject AI output, the human stops generating. Direction-setting atrophies. After a few months, the team can edit but no longer originate. The strategic muscle weakens precisely as the strategic stakes rise.
The exception trap. A reviewer scanning AI output for errors is doing pattern-matching at high speed. They catch obvious mistakes. They miss the systemic ones — the cumulative drift, the embedded assumptions, the things that are individually fine and collectively wrong. The loop optimizes for local correctness and is structurally blind to global incoherence.
The authorship void. When no one feels they wrote it, no one defends it. Content produced this way performs measurably worse — lower engagement, lower trust scores, faster discounting by audiences who have learned to recognize the texture of unowned writing. Customer experience built on the same pattern produces interactions that feel competent and forgettable.
None of this is an argument against using AI. It is an argument against the metaphor we are using to govern it.
Human in the lead
A different model: the human is not in the loop. The human is in the lead.
In the lead, the human owns the direction, the standard, and the outcome. AI is the most capable instrument they have ever held — but it is an instrument, not a co-author with a vote. The shift is not cosmetic. It changes three things:
Framing shifts from reactive to directive. The first question stops being “Is this output acceptable?” and becomes “What outcome am I producing, and what is the shortest path to it?” The AI is recruited to the path. It does not propose the destination.
Ownership shifts from exceptions to outcomes. The human is not accountable for catching errors. The human is accountable for what was produced, full stop. This is a higher bar, and it is the only bar that holds up under scrutiny — legal, commercial, or reputational.
Cadence shifts from spot-checks to direction-setting. Instead of reviewing each piece of AI output, the leader invests in defining the brief, the constraints, the evidence base, and the standard of done. The review burden drops. The quality ceiling rises.
This is not a softer model. It is a more demanding one. It requires leaders who can articulate what they actually want — which most organizations have spent two decades de-skilling themselves out of.
Why this matters now
The Swiss and Italian markets I work in show the divergence clearly. A growing share of mid-to-large enterprises have deployed generative AI in at least one workflow. A much smaller share can articulate, in a single sentence, what they are trying to accomplish that the AI is helping with. The gap between deployment and direction is widening.
Every quarter that gap stays open, the reactive posture compounds. Teams produce more, originate less, and accumulate an organizational dependency they did not consciously choose.
The companies that will compound advantage in the AI era are not the ones with the best models. They are the ones whose leaders stayed in the lead — who used AI to amplify a direction they were already setting, rather than to substitute for direction they no longer had time to set.
What this blog will do
This is the opening of a series. Over the coming months I will write — evidence-led, citing the actual research — on three intersecting territories:
Content in the AI era. What changes when production cost drops to near zero. What remains scarce. What audiences actually reward.
Customer experience as the new front line. Where AI helps, where it hurts, and why the firms winning on CX are the ones treating AI as a CX tool rather than a CX strategy.
Adoption and governance. What the research is showing about which patterns work, which fail, and why so many AI programs stall at proof-of-concept.
Each piece will be grounded in primary research — academic, market, or operator — and will state its sources. The promise of this blog is not that I will be right. The promise is that I will show my evidence and let you check the work.
The first specific question, in the next post: what does the evidence actually say about content performance when AI is in the production loop — and what changes when the human moves from the loop to the lead?
Stay close.
— Stefano