Business question
How do AI systems evaluate and recommend vendors?
In stages. An evaluator first forms an opening set for a broad category question, then narrows it as the buyer adds requirements: eliminating vendors that do not visibly satisfy each new requirement, moving the leader when a requirement exposes it, and settling on the vendors that survive the full sequence. The final recommendation reflects the accumulated requirements, not the opening question.
The problem
Treated as a black box, AI recommendation looks arbitrary. Seen as a staged process, it becomes something you can locate yourself within: where you enter, where you are removed, and where you survive.
Why it matters
Understanding the stages is what makes the other questions answerable. Each stage is a place where a company can be won or lost, and each is studied as its own research component.
What the research studies
The components behind this question.
The evidence so far
What we have run, and what we have seen.
- E-045Video conferencing software: When the field survives and keeps its names
- E-043Accounting software: When the whole field is wiped out and replaced
- E-044Accounting software: When the same kill repeats on a second evaluator
- E-041CRM software: When the recommendation set survives but its members do not
- E-034Clinical trial management software: When the recommendation set collapses
- E-040Project management software: When the recommendation set holds
- E-002Healthcare CDP and the HIPAA flip: the program's one causally verified gate
- Observation · Candidate · 2026-07-20The same elimination appeared on two different evaluators
- Observation · Held · 2026-07-20The category leader was removed at the first requirement it could not meet
- Observation · Candidate · 2026-07-20Survival depended on the category, in at least four ways
- Observation · Candidate · 2026-07-20Survival appeared in two modes, and their cause is confounded
What we can conclude
We have observed this staged narrowing across several categories: an opening set that is not the final answer, elimination at specific requirements, the leader exposed at the first requirement it cannot meet, and fields that survive, collapse, or relocate depending on the category. Every observation is held at its evidence level and most rest on a single evaluator.
What remains unknown
How closely this resembles human buying-committee evaluation, and how much of it generalizes across evaluators and over time, are open questions the research program is built to keep testing.
How Upstream Zero helps
From the evidence to your answer.
Upstream Zero studies each stage as a research component and turns that into a per-company diagnosis: where you enter, where you are eliminated, and what must become true to survive. The research is public; the diagnosis is the product.