Category Strategy and Spend Segmentation with AI

Category strategy starts with knowing what you buy, from whom, and under what commercial terms. In most organisations that is genuinely unknown, because spend data is fragmented across systems, supplier names are inconsistent, and classification was done by whoever raised the requisition. AI is very good at this cleanup, and the resulting visibility is often the single biggest analytical improvement a procurement function makes.

It also produces the failure that matters most in this lesson: a clean, confident, well-presented taxonomy that is wrong in ways nobody checks, because it looks so much better than what preceded it.

Classification at Scale

Spend classification maps transactions to a category structure. Done manually it is slow and inconsistent; done with AI it is fast and consistent, which is not the same as correct.

Three error patterns recur and each distorts strategy differently.

Systematic misclassification. An entire supplier or transaction type consistently lands in the wrong category — professional services misallocated between consulting and contingent labour, or maintenance materials split between MRO and capital projects on the basis of description text. Because it is systematic, category totals are wrong in a stable, plausible-looking way, and the error is invisible in any aggregate view.

Long-tail dumping. Ambiguous transactions accumulate in a miscellaneous or unclassified bucket, or worse, get forced into the nearest plausible category to keep the unclassified percentage low. A classifier reporting 97 percent classification is not necessarily better than one reporting 84 percent; it may simply be more willing to guess.

Entity resolution failure. The same supplier appears as several entities because of naming variation, legal entity structure, or acquisition history. Spend fragments across them, each looks small, and none reaches the threshold that would trigger strategic attention. This is the error most likely to hide a genuine concentration risk.

Entity resolution deserves particular care because it cuts both ways. Over-merging is equally damaging: consolidating two genuinely separate legal entities under one parent hides the fact that your commercial relationship, contractual protection, and insolvency exposure differ between them.

Validate Against Decisions, Not Consistency

The usual validation of a spend taxonomy is internal: does everything map somewhere, do the totals reconcile, is the structure coherent. Those checks are necessary and they miss the point.

A taxonomy exists to drive decisions: where to run a competitive event, where to consolidate, where a category manager should be assigned, where a price increase should be challenged. The validation that matters is whether the structure supports those decisions.

Two practical tests. The materiality test: take the ten largest categories and ask whether each is coherent enough that a single sourcing strategy makes sense for it. A category containing genuinely unrelated purchases cannot have one strategy, however tidy it looks. The action test: take five decisions the function needs to make this year and check whether the taxonomy answers them. If deciding whether to consolidate logistics providers requires manual re-cutting of the data, the taxonomy is not aligned to the decision.

Both tests are quick, and both catch structures that pass every internal consistency check while being useless.

Consolidation Recommendations Undervalue Resilience

AI-driven category analysis reliably identifies consolidation opportunities: this category has fourteen suppliers, the top three cover 71 percent of spend, consolidating the remainder would yield an estimated saving.

The saving estimate is usually sound. What the analysis cannot see is what the fragmentation is doing.

Some fragmentation is genuine inefficiency — historic contracts nobody rationalised, maverick spend, duplicate relationships from an acquisition. Some is deliberate and load-bearing: a second source retained for supply security, a regional supplier kept for lead-time reasons, a small specialist supplier whose capability the large suppliers do not have, a supplier retained to preserve competitive tension at renewal.

None of that is visible in spend data. The consolidation model sees fourteen suppliers and a saving; it does not see that supplier eleven is the only one qualified for a safety-critical part, or that the reason there are two logistics providers is a disruption three years ago.

The discipline is to treat a consolidation recommendation as a question rather than an answer: for each supplier the model proposes removing, why does this relationship exist? Where the answer is "nobody knows," consolidation is probably right. Where the answer names a specific capability, risk, or commercial reason, the saving must be weighed against it — and the trade-off should be recorded, because a resilience decision reversed for a saving is exactly the kind of choice that gets scrutinised after a disruption.

What AI Does Not Do Here

Classification, entity resolution, pattern identification, and savings quantification are all genuine strengths. Category strategy itself is not, and the boundary is worth stating plainly.

Deciding whether a category should be competitively tendered or partnered depends on the supply market's structure, your leverage, switching costs, and the relationship's strategic value. Deciding whether to insource depends on capability and capital, not spend patterns. Deciding how much supply security is worth is a risk appetite judgment that belongs to the business. Deciding which supplier relationships to invest in depends on where the market is going.

AI narrows and quantifies the options. The choice among them requires knowledge of the supply market and the business that is not in the transaction data, and a category strategy assembled from analytics output alone will be internally consistent and strategically empty.