Decision intelligence vendor questions — Every vendor demonstrates a dashboard and claims explainability. Four questions establish quickly whether there is a decision layer underneath.
Decision intelligence vendor questions: bu ne anlama geliyor
Evaluations of decision platforms tend to compare interfaces, because that is what a demonstration shows. These questions get past that.
Where does the decision get recorded, and what is in the record?
If the answer is the log file, or “it’s in the ERP”, there is no decision layer. Ask to see one record: inputs at the time, alternatives, the binding constraint, the approver.
What happens when the recommendation is rejected?
If rejection is not captured as data, the system cannot learn from your experts. This is the most common gap and the least often asked about.
Where does the model run, and what does it see?
If it runs outside your boundary, ask exactly which fields leave. Then ask what reasoning becomes impossible without the ones that stay.
Can it take an action, and can that action be reversed?
If the answer is “we produce a recommendation”, the last mile is yours. That may be fine — but the effort of the last mile is usually larger than the platform.
Vendors who have thought about these answer immediately. The pause is the information.
How is uncertainty carried into the decision?
A system that emits a point estimate has discarded the information that would change the action. Ask what the recommendation looks like when the model has four data points rather than four years — if it looks identical, the uncertainty stopped at the model boundary.
The answer you want describes a change in shape: a wider interval producing a more conservative order, a held review, or an explicit flag. Not a confidence percentage displayed next to a number nobody reads.
What does the system do that we could not do in a spreadsheet?
Uncomfortable and worth asking. A surprising amount of decision software is a hosted version of a calculation the planning team already performs.
The defensible answers involve scale, continuity or memory: decisions made across thousands of items every day, monitoring that persists across shifts, or a record that accumulates. If the answer is only “faster”, price it as a spreadsheet.
Who is accountable when it is wrong, and what does your contract say?
Vendors describe shared responsibility. The contract describes something narrower, usually availability rather than decision quality.
That gap is not necessarily wrong — a vendor cannot underwrite your commercial outcomes — but it should be understood rather than discovered. The practical question is what evidence the system will give you when a decision is challenged, since that evidence is what your own accountability will rest on.
How does the system get better, and from what?
A model retrained on outcomes learns what happened. A system that captures rejections and corrections learns from your experts, which is a different and faster loop.
Ask specifically: when a planner overrides, where does that go? If the answer is nowhere, the system will be as good in year three as in year one.
The reference call worth making
Not to the reference the vendor offers. To an organisation two years in, and the question is not whether they are satisfied but what is still manual that they expected to be automated.
Every deployment has such a list. Its contents tell you where this vendor’s last mile actually ends, and that is the part no demonstration shows.
What the pause means
Vendors who have built a decision layer answer these immediately, because each question corresponds to something they had to design.
A pause is not dishonesty. It usually means the question was never forced on them — and the reason it was never forced is that their deployments have not reached the point where it matters. That is worth knowing before yours does.