I've been in and around procurement transformations for over two decades. The conversations change. The problems don't.
In the early years, it was all about process. Then platforms arrived and we spent a decade arguing about which one to buy. Now everyone's talking about AI.
And almost nobody's talking about the real problem.
The implementation that went wrong in exactly the right way
A few years ago I was brought in to look at a transformation that had stalled. The platform was live. The implementation partner had delivered and moved on. The go-live had been celebrated — presentations made, milestones ticked. And six months later, the outputs were wrong often enough that nobody trusted them.
The diagnosis was quick and unanimous: something must be wrong with the AI.
It wasn't.
When we peeled back the layers, what we found was a spend taxonomy that hadn't been meaningfully updated in seven years. Supplier master data that had been migrated — not cleansed — from the legacy system. Purchase orders full of free-text descriptions that ranged from vague to genuinely baffling. Three cost centre hierarchies that nobody had reconciled before go-live.
The AI was doing exactly what it was built to do. It was processing the data it had been given.
The data it had been given was a mess.
What "clean data" actually means in procurement
The phrase gets thrown around loosely. In procurement it means something specific — and more demanding than it looks in a vendor demo.
It means your spend is classified consistently. Not just categorised, but classified to a level of granularity that allows real comparison and real decisions. It means your supplier master is current, deduplicated, and structured so that Supplier A, Supplier A Ltd, and Supplier A (UK) are understood to be the same entity. It means your cost centre hierarchy reflects how your business actually operates today — not how it was set up when the ERP went live.
Most organisations have none of these things properly in order. Not because they're careless. Because data governance is unglamorous, slow, and nobody's ever got a bonus for cleaning a taxonomy.
So it gets deferred. The platform gets bought. The AI gets switched on. And the outputs are wrong.
Where the problem hides
In every transformation I've been involved in, the data problem surfaces in one of three places. Often all three at once.
The taxonomy gap. Spend classification is inconsistent across teams, geographies, or time. What one business unit calls "IT Services" another logs as "Technology Consulting" and a third puts through as "Professional Fees." The AI sees three categories. You have one supplier doing one job across all three. No model makes good decisions on that.
The supplier master problem. Supplier data gets migrated once — usually at implementation — and then quietly deteriorates. Duplicates accumulate. Suppliers get added informally. Commercial relationships that matter don't appear in the system at all. When the AI is asked to assess supplier risk or performance, it's working from an incomplete picture of who you actually buy from.
The free-text problem. Most purchase orders are written in natural language. "Misc consultancy services Q3" tells you almost nothing. Multiply that across tens of thousands of lines and you have a dataset that looks rich and is analytically thin. AI can help here — but only after the classification layer is working. Using AI to compensate for missing classification is using a sophisticated tool to avoid a problem that a clear taxonomy would have prevented in the first place.
Why it keeps getting skipped
Data quality work isn't glamorous. It doesn't appear in vendor demos. It doesn't sit on implementation Gantt charts as a named workstream. It doesn't get celebrated at go-live.
It also doesn't get funded. Procurement transformations are budgeted for technology acquisition and implementation. The data remediation work — taxonomy design, supplier master cleansing, historical spend reclassification — sits in a grey zone between the business and IT, and tends to fall between both.
There's also an enduring belief that the platform will sort it out. That putting data into a modern system will somehow improve its quality. It won't. A modern system processes bad data faster than a legacy system. That isn't the same thing as making it better.
What good actually looks like
Organisations that get AI right in procurement do one thing differently: they treat data as a design input, not an implementation detail.
That means the data quality work starts before the platform is selected. The taxonomy is agreed and tested before a single purchase order is classified against it. The supplier master is cleansed, structured, and governed before migration. Someone is accountable for data quality as a named, funded responsibility — not a background assumption that the implementation partner is quietly expected to handle.
It also means resisting the pressure to go live before the data is ready. That pressure is real — it comes from timelines, budgets, and stakeholder expectations. The temptation to show something working is powerful.
But showing something working on bad data isn't progress. It's deferral with interest added.
Before your organisation scales its next AI initiative in procurement, ask one honest question: if you stripped out the model and looked only at the data it's operating on — would you trust it to drive a commercial decision?
If the answer is anything other than yes, you're not ready to scale.
Fix the data first. The AI will be there when you are.
