Sat through another painful vendor selection meeting last week. Four hours comparing AI platforms using criteria written for buying laptops.

The procurement team had done everything “right”—RFI, RFP, vendor demos, reference checks. They’d scored everything on a spreadsheet with weighted criteria. Very thorough. Completely useless.

The Fundamental Problem

Traditional procurement frameworks assume you know what you’re buying. You specify requirements, vendors respond, you compare responses against requirements.

AI doesn’t work like that. Half the time you don’t know what’s possible until you’ve experimented. The vendor’s demo environment rarely reflects your data complexity. And the “requirements” you wrote six months ago may be obsolete by the time you sign.

I watched a financial services firm spend eight months selecting a document processing AI platform. By the time contracts were signed, two of the four vendors had shipped major updates that completely changed the evaluation. One had pivoted their product entirely.

What’s Actually Going Wrong

Requirements written too early. Procurement wants locked requirements before evaluating vendors. But you often can’t know what to require until you’ve seen what’s available and tested it against your actual data.

A logistics company I work with specified “95% accuracy on invoice extraction” as a requirement. Reasonable-sounding. Except their invoices ranged from simple to genuinely horrible PDFs with handwritten amendments. The 95% requirement was meaningless without specifying which invoices.

Security reviews that don’t match the technology. Standard security questionnaires ask about data centres and encryption at rest. Fair enough. But they rarely probe the questions that actually matter for AI: Where does data go during inference? How are models updated? What happens to the data used for fine-tuning?

Proof of concepts that prove nothing. Vendors run POCs on clean sample data you provide. Everything works beautifully. Then you deploy on real production data with edge cases and the accuracy drops by 30%.

A Better Approach

The enterprises I’ve seen handle this well treat AI procurement as iterative.

Start with a problem, not a solution. Don’t write an RFP for “an AI document processing platform.” Define the business problem: “We spend $2M annually on manual data entry for supplier invoices, with 3% error rates causing payment delays.”

Then explore multiple solution approaches—maybe AI, maybe RPA, maybe process redesign, maybe all three.

Pilot before you procure. Structure initial vendor engagements as paid pilots, not free POCs. Four weeks, real data, real users, real metrics. Yes, it costs money upfront. It costs much less than an 18-month contract with the wrong vendor.

Involve operations from day one. The people who’ll actually use the system know edge cases procurement never will. A call centre manager can tell you in 10 minutes why the chatbot demo won’t work with your actual customer queries.

Procurement’s Legitimate Concerns

This isn’t about bypassing procurement. Their concerns are valid—vendor lock-in, data security, commercial terms, compliance. These all matter.

But the timing and process need adjustment. Commercial negotiation happens after you’ve validated the technology works. Security review needs AI-specific criteria, not repurposed SaaS checklists.

One approach that works: procurement embedded in the pilot team, learning alongside operations and technology. They see what matters firsthand instead of reviewing documents in isolation.

The Vendor’s Perspective

Talked to a few AI vendor sales teams recently. Their frustration mirrors the buyer’s.

They hate responding to RFPs with irrelevant questions. They know their demo data doesn’t represent your reality. They’d rather do shorter, paid pilots than drag out free POCs that consume resources and often lead nowhere.

The good vendors want to prove value quickly and get to commercial discussions. They’re not trying to hide problems—a failed pilot is better than a failed deployment that destroys the relationship.

Where This Leaves Us

Enterprise AI procurement is a mess because we’re applying 20-year-old frameworks to fundamentally different purchases. The solution isn’t abandoning governance, it’s updating it.

Shorter cycles. Real data earlier. Operations involved throughout. Commercial teams learning the technology domain.

It’s harder than ticking boxes on a spreadsheet. It also produces better outcomes.