The AI Vetting Your Renewals Runs on Copy-Paste

A founder I work with handed a $30K renewal to her ops manager. He did the smart thing and ran it past Claude. Pasted the quote, asked if the price was fair, got back a clean analysis: market rate, standard terms, reasonable to renew. He approved it. What the model never saw was the second tool they already pay for that covers most of the same features, or the fact that they're using half the seats they license. In isolation the renewal looked fine. It wasn't even close.

She'd paid for a confident answer to a smaller question than the one she thought she'd asked.

Daily AI, zero connection to your spend

A benchmark out this month surveyed 121 procurement teams. The headline number is the gap. Most of them use AI four or more days a week, but almost 90% of that usage runs on general-purpose tools: ChatGPT, Copilot, Gemini, Claude. Only 8% touch AI that's actually wired into the system holding their spend data. Activity is everywhere. Connection is almost nowhere.

For a finance team that's a strange shape. Your people now lean on AI constantly to reason about vendors and contract terms. Every one of those sessions starts from a blank slate and runs on whatever the person remembered to paste in.

A confident answer on partial inputs

The model is good. It'll tell you the quote tracks the market, the auto-renewal clause is normal, the discount is real. All true on the inputs it got. None of it accounts for the contract you signed eight months ago, or the overlapping product, or the usage you're nowhere near. The AI doesn't know what it wasn't told, and it won't flag the hole. It answers with the same confidence either way.

A spreadsheet missing half your data looks obviously broken. An AI answer missing half your data looks finished. That's the trap.

This is worse than no analysis, because it reads like analysis. The renewal gets signed with a paper trail that looks like diligence. Nobody pulled what you already spend with the vendor. The decision feels data-driven and isn't.

Close the gap, or label it

Only 17% of those teams have an AI policy anyone enforces. You don't need a policy binder. You need two habits.

First, for the handful of decisions where AI touches real money, check the answer against your system of record before anyone signs. Renewals over a set dollar threshold, new vendor commitments, anything recurring. Pull what you already spend with that vendor and what overlaps it. That's minutes of work, and it's the exact context the model couldn't see.

Second, make AI-assisted analysis name its inputs. "Reviewed the quote" and "reviewed the quote against our spend history" are different claims. The person approving should know which one landed on their desk.

Claude wasn't wrong. It answered the question it was given, and the question was missing your numbers.