In early 2026 Google DeepMind published a framework called “Intelligent AI Delegation”, and I would put it on every decision-maker’s desk if I could. The core idea fits into one sentence: we no longer delegate the process to the AI. We delegate the outcome.
In practice at SlopeLift that looks like this: we build Agentic Workflows that take over whole pieces of campaign management autonomously. Budget shifts between channels, automatic creative rotation based on performance signals, reporting that assembles itself. For someone who, like me, spent hours patching VLOOKUP formulas in 2008, that sometimes feels like science fiction. And still: the moment when the decision really counts stays human.
The contextual knowledge no model has
I like to give a concrete example when people ask why we put so much weight on “Human-on-the-Loop” at SlopeLift. There is a client whose ROAS drops every year in calendar week 12. Reliably, reproducibly, every time. An agent sees that data point and would react: budget down, adjust bids, maybe pause the campaign. All logical, all wrong. Because the drop has nothing to do with the campaign. It is payday week for that audience, and buying behaviour normalises by itself afterwards. That is contextual knowledge that appears in no data set and that no model in the world can replicate.
In-the-Loop means: the human does the work, the AI assists. On-the-Loop means: the AI does the work, the human sets the goals and steps in when context is needed.
Or: the 30-second conversation with a client in which you notice that the business strategy is turning 180 degrees right now. No briefing document captured it, no dashboard shows it. But that conversation changes every campaign decision of the coming weeks.
That is the difference I see between “Human-in-the-Loop” and “Human-on-the-Loop”. In-the-Loop means: the human does the work, the AI assists. On-the-Loop means: the AI does the work, the human sets the goals and steps in when context is needed.
We are in the middle of that transition at SlopeLift, and I do not want to pretend it is comfortable. It requires people to let go of what they were good at for years. At the same time they have to develop capabilities for which there is no course and no prompt: judgement, client understanding, strategic thinking under uncertainty.
The gap between promise and reality
The 56% figure from the PwC study did not surprise me, because I know the same mistakes from my own experience. Where AI worked for us immediately: on the tasks nobody wanted to do anyway. Data cleaning, first drafts, quick research sprints. There the time gained was measurable, and there was no discussion.
Where we fooled ourselves: everywhere we put AI as a turbo on top of an existing process without asking whether the process itself was any good. A bad process plus AI produces more of the bad, only faster and in greater quantity. That sounds obvious when you say it out loud. In practice you often only recognise it once the pile of output is already sitting there.
The difference between the 56% without ROI and the 44% with ROI probably does not lie in the technology. It lies in the willingness to allow the uncomfortable question: is the process any good at all, or are we currently accelerating a standstill?
Three things that will not become commodities
After 18 years in this industry and a year as CAIO I can say one thing with certainty: three things will rise massively in value, and none of them is a tool.
Data. But not simply “more” data — the kind of data that contains insights your competitors have no access to. Data is what makes AI intelligent in the first place. Without it, the best model is blind.
Talent. And by that I mean real, deep talent, not the wave of people passing around prompt recipes and tool hacks. Anyone who wants to build differentiated solutions needs minds that can supply the context no model replicates.
Leadership. The most underrated factor. Leaders who understand where the technology is moving and place the right bets. That is an advantage no tool and no model can produce.
The technology becomes a commodity. These three things will not. And that is, essentially, the contemporary version of what we meant in 2014 when we called SlopeLift a “workshop”: the efficiency that technology delivers is what makes craft properly sharp. The time that agents give back flows into the questions for which no prompt exists.
Back to Part 2 | Continue with Part 4: The barbarians are not coming