Procurement has never adopted a technology this quickly. AI has moved from curiosity to budget line in under two years, and research from The Hackett Group now puts deploying AI-enabled technology among procurement’s top priorities, with more than half of procurement organizations reporting some deployment of agentic AI.
Here is the uncomfortable number that sits next to it: in an Efficio survey of CPOs and CFOs at large European companies, fewer than one in five said they had full visibility into their indirect spend.
Put those together and you get the defining problem of procurement AI right now. Organizations are deploying AI on data they can’t see clearly.
The pilot that looks like it worked
The typical pattern goes like this. A vendor or internal team scopes a fixed-fee AI pilot: classify spend, spot savings, maybe draft some sourcing strategies. The demo is impressive. The model produces confident category breakdowns and a list of opportunities.
Then someone checks the numbers.
The same supplier appears under four names. A third of invoices carry no usable category. Contracts live in email attachments and shared drives, disconnected from the invoices they govern. Cost centers were restructured two years ago and nobody remapped the history. The model didn’t fail — it did exactly what it was asked to do with the inputs it was given. The inputs were the problem.
At that point the pilot has two options. Quietly expand into an unplanned data cleanup nobody budgeted for, or declare victory on a demo and stall before production. Most take the second path.
Why this keeps happening
AI tools are sold on what they can do with good data. Very few buyers are told, up front, whether their data is good enough. The incentive runs the wrong way: the fastest way to sell a pilot is to assume the data is fine, and the reckoning comes after the contract is signed.
The underlying issues are rarely exotic:
- No common identifiers. Supplier, contract and invoice records can’t be reliably joined.
- Inconsistent classification. Categories were assigned by whoever entered the invoice, using whatever taxonomy existed at the time.
- Fragmented sources. Spend lives in the ERP, commitments live in contracts, performance lives in someone’s spreadsheet.
- No repeatable refresh. A one-time data extract produces a one-time answer.
None of these are AI problems. They are foundation problems that AI makes visible — expensively.
What to do first
Before funding an AI procurement initiative, answer four questions honestly:
- Can we join spend to suppliers to contracts? If there’s no reliable key across those three, analysis will be guesswork dressed up as insight.
- Do the people doing the work trust the data? Interview the analysts and buyers who handle it daily. They know exactly where matching breaks.
- Is there a repeatable way to refresh it? Scheduled reports or secure file drops are often enough. An API is not automatically better.
- Who owns the decision logic? The model should generate hypotheses. People should decide what counts as savings.
If the answers are weak, the right first project isn’t an AI pilot. It’s a scoped foundation build — identifiers, normalization, governance and a refresh process — priced and planned as its own piece of work. It’s less exciting to announce, but it’s the thing that makes every later AI investment pay off.
The S2V position
We tell clients where their data stands before they spend money on AI, even when the answer costs us the easy pilot. Our engagements start with a readiness decision, and if a foundation build is needed, we scope it separately rather than hiding it inside a fixed fee.
AI will change how procurement works. But it accelerates whatever it’s pointed at — including bad data. Fix the foundation first, and the speed is real.