There is one distinction I like to weld into slides for client meetings — occasionally onto the back of a beer mat too — because it decides more about whether AI pays off than any choice of tool. The distinction between Human-in-the-Loop and Human-on-the-Loop.
Two modes, one fundamental decision
With In-the-Loop, the human does the work and the AI assists: you write the email, the model polishes it. You build the campaign, the model suggests adjustments. The human stays inside the process step by step, the AI only hands over tools.
With On-the-Loop the centre of gravity flips completely: the AI takes over execution, the human stands above it — setting the goals and stepping in exactly when judgement is required.
The degree of automation can be just as high in both. What matters is where the human stands in that picture: in the middle of it, taking notes, or above it, with a hand on the ripcord.
For two decades I optimised performance marketing campaigns by hand, adjusting bids by device and time of day. Today, at SlopeLift, we build agentic workflows that shift budgets between channels, rotate creatives on performance signals and assemble reports on their own. And the moment that really counts still belongs to a human.
The context no model has
We have a client whose return on ad spend collapses every year in calendar week twelve. Reliably, reproducibly, every time. An agent sees that data point and reacts exactly as it was built to: budget down, adjust bids, pause the campaign if in doubt. All logical. All wrong.
The drop has nothing to do with the campaign. It is payday week in the target group, and buying behaviour normalises by itself a week later. That knowledge sits with a human who has been watching the account for three years. No dataset records it.
Or the thirty-second conversation with a client in which it becomes clear that their business strategy has just turned 180 degrees. No briefing document catches that, no dashboard displays it, and yet it changes every campaign decision for the next month. The famous “reading the room”. Human-on-the-Loop exists for exactly those moments. The agent drives the machine. The human owns the goals and the judgement the machine cannot get anywhere near.
Why this is what decides whether AI pays off
The PwC Global CEO Survey (January 2026, Davos, 4,454 respondents from 95 countries) found that 56 % of CEOs see no measurable return on their AI investments. That does not surprise me, because I know the mistake from the inside.
Where AI worked for us immediately: tasks nobody wanted anyway — cleaning up data, first drafts, quick research. Where we fooled ourselves: everywhere we strapped AI onto a process without first asking whether the process was any good at all.
A bad process plus AI produces more bad output, faster.
Human-on-the-Loop is the discipline that prevents this. It forces an explicit decision: which tasks move to the AI, which stay with a human, and where the human has to be able to intervene. It sounds like a process question. In practice it reshapes the work more deeply than any tool rollout, because it forces you to name what may truly belong to a human only.
What stays human
After a year as Chief AI Officer I am fairly certain what is in short supply — and not one bit of it is a tool:
- The goals — an agent optimises for whatever you point it at. Pointing it well is a strategic act.
- The context — payday week, the strategy that turned, the things that are in no dataset.
- The judgement — to overrule the logical-but-wrong move, in the moment you recognise it.
A Delegation Matrix — recording task by task what the agent decides alone, what needs human sign-off and what stays entirely off limits — is the most useful tool most teams are still missing. The matrix has to be on paper before the agent goes live.
The transition is uncomfortable. It asks people to let go of work they have done well for years, and to build skills for which no prompt exists: judgement, client understanding, strategic thinking under uncertainty.
The teams that take this step reclaim the hours automation hands them and put them into the questions no model can answer. The teams that skip it automate their bad processes at scale — and wonder why the return never arrives.
The question I would put to anyone rolling out AI this year is narrow and uncomfortable: in your most important workflow — who stands on the loop, and does that person have the ability to overrule the machine at any time, when the data looks perfectly logical and is still completely off?
You can read the full guest commentary on horizont.at (digital subscription).