One supplier, three names: fix the spend picture before AI recommends savings

The finished pilot should give your purchasing manager a spend view they can inspect: supplier identities, purchase categories, original records and unresolved decisions. Our proposed acceptance rule is straightforward: map one quarter’s records without silently merging distinct legal entities or changing the agreed spend total.

This is a reconciliation exercise, not a savings forecast. Keep purchasing decisions and ERP master-data changes outside its scope.

The capability exists; your records still need a test

On September 16, 2026, SAP announced SAP Ariba Spend Analysis and Insights, describing cross-system ingestion, classification against standard or customer taxonomies, and corporate-hierarchy enrichment. The announcement establishes product availability and describes vendor capabilities; it does not demonstrate reconciliation on your exports. SAP announcement

Supplier.io likewise describes legal-entity matching, confidence scores, matched attributes and source documentation. Its stated reporting value explicitly includes cleaner reporting rather than projected savings. These are vendor descriptions, not independently measured outcomes for your business. Supplier.io product page

The strongest counterpoint is therefore practical: you may not need another matching engine. Test existing capabilities against the same acceptance criteria before commissioning additional software. We would begin with the records and decisions the pilot must explain, not a requirement to replace your procurement stack.

1. Freeze the inputs and define the total

Choose one purchasing category over one quarter. Ask the purchasing manager to approve its scope, including which neighbouring or uncategorised records should be inspected for possible inclusion.

For this proposed pilot, preserve the ERP exports and spreadsheets unchanged. Give each imported row a source-file reference and stable row identifier. Agree which files contain transactions and which merely annotate or repeat them; do not automatically add every spreadsheet amount to the ERP total.

Before matching, document the accounting basis: included dates, currencies, tax treatment, credits and exclusions. Reconcile currencies separately unless an explicit conversion basis is approved. Where files overlap, record the decision about which transaction representation counts and retain the excluded copy as evidence.

That produces the baseline the transformed view must reconcile to. Treat scope corrections as documented baseline revisions, never silent cleanup.

2. Build a supplier model, not a cleaned name column

GLEIF distinguishes identity information—such as an official name and registered address—from relationships to direct and ultimate accounting-consolidating parents. That distinction provides a useful reference for separating “same entity” from “related entities.” GLEIF identity and ownership explanation

We would build a bounded shared supplier model connecting source records, aliases, legal entities, supported parent relationships and purchase categories. This is the pilot’s foundational ontology: an explicit description of the business objects and their relationships, not a replacement ERP.

Consider these hypothetical review decisions, not client records:

• Merge aliases: “Supplier A,” “A Ltd” and “A Purchasing” carry the same verified legal identifier and consistent supporting records. Propose mapping them to the same entity; retain every original name.
• Link, do not merge: similar names carry different verified legal identifiers, with evidence of a common parent. Preserve separate entities and offer a separately labelled group view.
• Leave unresolved: names resemble each other, but identifiers are missing or supporting details conflict. Send the proposed match and its evidence to the purchasing manager.

Do not make name similarity an approval rule. Validate identifier availability and registry coverage on the actual sample before relying on either.

3. Make the manager’s decision inspectable

We would build a review queue showing the original rows beside each proposed supplier mapping. Include supporting attributes, conflicting evidence, the proposed relationship and any confidence score. Let the purchasing manager approve, reject or defer the decision, with a recorded reason.

Use the same review discipline for purchase categories. Keep the source category beside the proposed category and the transaction description. Require review where evidence is insufficient; do not force an assignment simply to remove an empty field.

Record the mapping version and reviewer decision so a corrected rule can be applied again without rewriting the source records. In this pilot, publish approved mappings only to the analytical view, not back into the ERP.

The trade-off is deliberate: accept visible unresolved work rather than impose unsupported certainty.

4. Accept the reconciliation before testing recommendations

Set acceptance criteria before running the pilot. Require every included transaction to remain traceable to its source and every amount to appear in the reconciled view, including unresolved supplier and category buckets. Require reconciliation to the baseline under the agreed accounting rules.

Then inspect identity and classification decisions separately. Do not accept a balanced total as proof that the assignments are right. Review proposed merges, disputed classifications and a manager-selected sample of apparently straightforward matches.

Measure reconciliation effort through recorded preparation and review time; unmatched spend through its amount and share of the agreed total; and classification corrections through reviewed assignments changed by the manager. State denominators and keep unreviewed records distinct. Do not translate these measures into promised savings.

For your first step, select the category, name the purchasing reviewer and agree the baseline. Only advance to AI savings recommendations when you can explain where the spend sits, which decisions were approved and what remains unresolved.