I spent eight years at SEB, and in that time I do not remember a single initiative that failed because the model was not good enough. Not one. They failed for reasons that would bore you at a dinner party, which is exactly why nobody writes about them.
Root cause, almost every time: nobody owned the data contract, and nobody had decided who was allowed to act on the output.
The data contract
A data contract is the boring agreement that says this field means this thing, it arrives at this time, and if it changes you tell me first. It is not a technology. You cannot buy it.
When I built the self-service platform that took data onboarding from weeks to hours, the engineering was maybe a third of it. The rest was getting teams to agree on what they owed each other. Once that was written down, everything downstream got faster, including things nobody had planned for.
Skip that step and you get the pattern everyone recognises. The pilot works on an extract that somebody prepared by hand. It goes to production. The upstream team changes a column in good faith because nobody told them anyone was depending on it. The model quietly gets worse for six weeks before a human notices.
Decision rights
The second one is harder because it is political. A model produces a number. Somebody now has to be allowed to act on that number, and somebody has to be accountable when it is wrong.
If you have not answered that before you build, you will end up with what I think of as advisory AI. The system makes a recommendation, a human reviews every single one, and the human is measured on errors rather than throughput. So they check everything. The efficiency you modelled never arrives, and the conclusion in the steering committee is that the AI underperformed.
It did not underperform. It was never allowed to perform.
What I would ask before funding anything
Four questions. If a team cannot answer them in a meeting, the money is early rather than wrong.
- Which decision does this change, and who makes it today ?
- Who owns the data it depends on, and have they agreed to keep it stable ?
- What is this person allowed to stop doing when it works ?
- How do we find out it has broken, without waiting for a complaint ?
The fourth one catches more failures than the other three combined, and it is the one that gets cut for time.
The uncomfortable version
Most transformation programmes are structured as a technology rollout because a technology rollout is legible. It has a vendor, a budget line and an end date. The actual work is redistributing who gets to decide things, which has none of those, and which no steering committee has ever enjoyed.
I am not arguing that the technology does not matter. I have spent my career on the technology. I am saying that after eight years inside one of these, I would rather inherit a mediocre model with clear ownership than an excellent one with none.