
Six months of RFP. Four partner pitches. An assessment deck to cap it off. A year gone, a procurement file three inches thick, and at no point in the process did anyone watch the software run on a problem that belonged to you.
The process is not irrational — it was built for a world where software was a licence, capability was documented, and the risk you were managing was contractual. AI inverts all three. The value sits in a workflow rather than a licence, the output is probabilistic rather than specified, and the risk you are actually managing is whether anyone will use it in six months.
What an RFP actually selects for
It selects for the capacity to answer an RFP. That favours suppliers with a bid team — which correlates with size, and correlates only loosely with whether they can put a working capability inside your operations. The firms with the deepest delivery scar tissue in a narrow problem are frequently the ones least equipped to produce two hundred pages on a deadline.
You are not buying a document. You are buying whether a specific team can make one workflow work in your building.
The criteria that actually predict delivery
Six, each with a way to test it rather than a way to ask about it.
- Depth in your specific problem, not your industry. "We serve financial services" is positioning. "We have done permissions-aware retrieval across a document estate like yours four times" is evidence. Test: ask them to describe the third-hardest problem they hit last time.
- Production deployments, not pilots. Ask what proportion of their engagements are still running unassisted a year later. The hesitation is the answer.
- Handover versus permanent embedding. Some partners are structured so that you never stop needing them. Test: ask what their handover looks like and who owns the runbook afterwards.
- Governance posture at your scale. Not whether they have a policy — whether they have been through a review at your risk tier and can describe what got blocked.
- Independence from any single model or cloud. Ask what happens commercially and technically if you switch model providers in eighteen months.
- Who exactly staffs your engagement. The people in the pitch are frequently not the people who arrive. Test: ask for the names and the calendar.
The alternative process, in five steps
- Name the problem — one workflow, one number, one owner.
- Shortlist on fit rather than on category presence. One or two, not four.
- Watch one work live, on a scenario shaped like yours, with your operators asking the questions.
- Run a short working session in person, with the department and the partner in the same room, and produce something the team can run.
- Decide on the evidence you gathered, not on the deck you were sent.
What vetting should cover — and why most directories vet nothing
The uncomfortable fact about most vendor directories is that placement is bought. The ranking reflects a marketing budget, and the listing text was written by the vendor. That is not a scandal; it is simply a different product from the one buyers assume they are reading.
Vetting worth the name covers whether the deployments are real and referenceable, whether the delivery team exists, whether the governance answers survive contact with a legal reviewer, and whether the vendor will demonstrate in public without controlling the room. The last one is the cheapest filter available and almost nobody applies it.
Where to start
If you know your problem, the shortest path is to have the match made for you and watch it work. If you would rather look first, browse by category — but read every listing knowing which of them were earned and which were bought.
Keep reading
See who's already vetted for this.
Every partner in the directory has completed A³ review — enterprise references, capability, domain fit and delivery readiness. Live Proven is earned in an A³ session.