The pilot is not the hard part
Most large organizations can now build a working model in weeks. The demonstration impresses the steering committee, the business case gets approved, and then the program stops moving. Not because the technology failed, but because nothing downstream of the model was ready to change.
An AI system produces a recommendation. Someone has to be permitted to act on it. That permission is an operating model question, not a technical one, and it is almost never resolved during the pilot.
Three things the model exposes
When intelligence is introduced into an established process, it surfaces problems that were always present but never forced a decision.
- Decision rights are unclear. The model suggests a price, a limit or a schedule. Who is accountable if it is wrong, and who is accountable if it is right and ignored? In most enterprises this has never been written down.
- The process assumes a human pace. Approval chains built around a weekly cycle cannot absorb a recommendation that changes daily. The intelligence arrives faster than the organization can act.
- Ownership sits in the wrong place. The model is owned by a data team, the process by an operations team, the outcome by a business unit. Three owners means no owner.
The constraint is rarely the model. It is the distance between the recommendation and the person allowed to act on it.
What the redesign actually involves
The work is unglamorous and mostly organizational. Define which decisions the system makes alone, which it recommends, and which it only informs. Set the thresholds at which a human is required. Name a single accountable owner for the outcome, not for the model.
Then instrument it. If a recommendation is overridden, that override should be recorded and reviewed. Override patterns are the most useful feedback an enterprise AI program produces, and most organizations discard them entirely.
Sequence matters
Enterprises that succeed with AI tend to redesign one decision process completely before scaling to a second. Enterprises that struggle run twelve pilots in parallel and redesign nothing. The second approach looks faster for about two quarters.
Where to start
Pick one decision that is made often, is currently made on judgment, and has a measurable outcome. Map who makes it today, on what information, and how long it takes. That map is the real design document. The model comes after it.