Hybrid intelligence is not AI analysing and a human approving. It is a way of working where either side can change the conclusion along the way, because each is strong at something different. The mark of it is that the result lands somewhere neither would have reached alone.
There are two common ways of talking about AI and strategy.
One says AI cannot think strategically, so humans should lead and AI should be a tool. The other says AI is better at spotting patterns, so AI should do the analysis and the human should validate.
Both are half right, and both end with one party not thinking.
What it is not
Not "AI analyses, the human approves". Here the human is only an assessor. It is not thinking, it is judging a piece of work. And approval is rarely a real test, because whoever approves did not do the work.
Not "the human asks, AI answers". Here the result is bounded by what the human thought to ask. If you do not know the answer in advance, you rarely ask the right question. The tool becomes a bottleneck rather than an extension.
Not "we use AI to save time". Time is the least interesting gain. An hour with real resistance produces a different result from a day of agreement.
What it is
Both sides contribute what they are strong at, and challenge the other exactly where it is weak.
The human has judgement about whether something matters at all. It has the context, meaning what happened last time you tried. It has creativity under constraint, meaning what you do when there is no money. And it knows the social consequences of a decision.
The model has stamina in pattern recognition, systematic handling of how variables connect, and a certain freedom from wanting to please anyone in the room.
In practice the sequence runs: the model proposes from what it can see. The human challenges with what is not in the data. The model adjusts on the new information. The human integrates.
Neither side had the last word alone. That is the whole mark of it.
One exchange, concretely
A manufacturer has to decide whether to digitalise production.
The analysis: the defect rate sits around 12%, and at comparable companies that have digitalised it is considerably lower. Time per unit falls too.
The human: our people are not especially digital. Implementation will be hard.
The analysis: it was hard at the three places I can see too. All three got through it, all three took eight to twelve months, and at all three the staff said afterwards it was better.
The human: we cannot lose production capacity for eight to twelve months. Cash flow will not take it.
The analysis: at two of the three it was done in phases, one line at a time. That doubles the total time, but production keeps running.
The human: then we have to start now to be finished in eighteen months.
The analysis: and what went wrong at one of them was that the hardware replacement was not planned separately. That took three to four months on its own.
The human: so we start with the hardware plan.
Notice what happened. The human had the sense that the defect rate was too high, and the context about cash flow and people. The analysis had the numbers, the patterns from elsewhere, and the detail about hardware. Neither was sufficient.
Alone, the human would either have said it cannot be done, or bought the equipment and hoped. Alone, the analysis would have proposed something that broke the cash flow.
The facilitator's job
This requires somebody to hold the process, and that is a different role from solving the problem.
The facilitator has to ensure four things:
- That the human actually challenges. The question is "what in that analysis do you not believe?", not "do you agree?".
- That the analysis receives the new information. If an objection comes from practice, it goes into the calculation rather than into the minutes.
- That the pace holds. If the same point comes up a third time, nobody is getting wiser.
- That the synthesis gets written down. What did we learn, and what do we do about it?
The poor facilitator decides who is right. The point is that both are right on different dimensions, and the work consists of getting both dimensions in.
How to start
Four meetings are enough to try it.
First meeting, data. The analysis presents what looks most important. Management discusses whether those are the right things, or whether context is missing.
Second meeting, patterns. The analysis proposes what others in similar situations did. Management challenges on where you do not resemble the others. That is the most important question in the whole process.
Third meeting, integration. Management formulates what they believe given both. The analysis works the consequences through.
Fourth meeting, decision. The direction is chosen, with a trade-off in it.
Try holding the third meeting without looking at new data. The purpose is to find out what you actually think, once you have seen both the numbers and the patterns.
How to draw the starting point and the target picture is in AS IS to TO BE, and who should sit in the four meetings is in the leadership room.
Why the model can take the new information in
The weak point in this way of working is step three. The human raises an objection, and if it cannot enter the analysis, it becomes a line in the minutes.
In 360° Sprint each analysis is a node on the board, and the connections determine the order. Correct an assumption in one node and the analyses downstream inherit the correction rather than sitting there with the old one. The version history with diffs means a change can be rolled back field by field if the objection turns out to be wrong.
That is what makes resistance cheap. If an objection cannot change anything without somebody redoing the analysis, it will not be raised next time.
What a model can and cannot contribute at all is covered in what AI agents can actually do.
Hybrid intelligence is not a technology. It is a meeting format where both sides can move the conclusion, and where neither is finished thinking.
How the basis gets separated from the decision itself is covered in decision basis in leadership.