Find your biggest AI opportunities in under 30 minutes.Book a consultation
Back[ Coventa ]Case Study

[ SPEND INTELLIGENCE ]LIVE

Hub for Intelligent Purchasing

Spend, understood at the level decisions are made. Hub harmonises historical purchasing against a live canonical product hierarchy, shows how each spend category is evolving against its underlying materials and vendor exposure, and clears the uncategorised bucket that makes every analysis approximate.

HUB // SPEND CATEGORIESCategory × vendor concentration
Hub Spend/Category × vendor concentration
Refreshed hourlyLast sync 12s ago

Spend categorised

96.4%

was 61%

Uncategorised

$2.1M

was $22M

Single-source categories

3

Duplicate materials merged

1,842

Concentration

Vendor AVendor BVendor CVendor DVendor EOtherFastenersBearingsElectricalPackagingMRO servicesUncategorised

Signals surfaced

Fasteners · 90% Vendor A

$4.8M · single-source risk

Packaging · Vendor D

70% · contract expires Q4

Bearings

3 vendors · consolidation saves ~6%

// Problem

The Problem

Every spend analysis starts with a caveat about the uncategorised bucket. Purchase orders arrive with descriptions written by whoever raised them, the product hierarchy was last curated two reorganisations ago, and a material sits under three category codes depending on which plant bought it. So category-level spend is directionally true and operationally useless: a negotiation team cannot tell a supplier what they actually buy, and a cost programme cannot tell whether a category moved or the coding did.

  • A large share of purchase orders sits uncategorised, and every analysis is qualified because of it.
  • The product hierarchy drifts out of date faster than it can be manually curated.
  • The same material is coded differently across sites, so vendor exposure is understated.
  • Category change scenarios cannot be evaluated because the baseline is not trusted.

// Overview

Hub gives procurement a granular and current picture of what the organisation actually buys. Historical purchases are harmonised against a real-time canonical product hierarchy, so category spend reflects the materials underneath it rather than the codes attached to it. Category analysis then shows how spend is evolving against those underlying materials and against current vendor exposure, which is the view a negotiation or consolidation decision needs. What-if scenarios can be run on that base — concentrating on a subset of vendors, for instance — with the risks of the alternative surfaced alongside the opportunity. The uncategorised bucket is attacked directly: generated recommendations propose a match to an existing category or a justified addition to the hierarchy, which a category manager accepts or rejects.

// AI System

Why AI

Categorisation is the whole problem, and it is a language problem. A purchase order line is a short, abbreviated, human-written string; deciding which category it belongs to means understanding what the thing is, not matching a token. That is why rule sets stall at partial coverage and why the uncategorised bucket never empties. A model closes it, but the hierarchy is a governed asset — so recommendations are proposals, and a category manager remains the authority on what the taxonomy says.

// Specs

Specifications

HARMONISATION
Historical purchases against a real-time canonical hierarchy
ANALYSIS
Category evolution by underlying material and vendor exposure
SCENARIOS
What-if evaluation, with risks surfaced alongside opportunity
CATEGORISATION
Recommendations for uncategorised POs, accepted or rejected by a manager
AUTONOMY
Proposes taxonomy changes; the hierarchy stays governed

// Features

Features

  1. 01Historical purchasing harmonised against a canonical product hierarchy that stays current.
  2. 02Spend category evolution analysed against underlying materials rather than coded labels.
  3. 03Current vendor exposure visible per category, so consolidation decisions have a baseline.
  4. 04What-if scenarios evaluated for change opportunities and for the risk they introduce.
  5. 05Uncategorised purchase orders matched to existing categories by generated recommendation.
  6. 06New hierarchy additions proposed with justification, for category-manager approval.

// Architecture

Architecture

SPEND FLOW

Runtime · one item, left to right


  1. 01Purchase Order History
  2. 02Canonical Hierarchy Sync
  3. 03Categorisation + HarmonisationMaterial MasterVendor MasterUncategorised BucketSpend History
  4. 04Category & Exposure Analysis
  5. 05What-If Scenario
  6. 06Sourcing Decision

dashed = the inference step, where the system exercises judgment

System stack

Data in · decisions out

01

Sources

Purchase history in all its inconsistency

ERP purchase orders5 yearsMaterial mastermultiple plantsVendor masterContracts & cataloguesInvoices

02

Ingestion

Descriptions to canonical items

PO line normalisationunits, abbreviations, languagesCanonical hierarchy syncUNSPSC / custom taxonomyVendor deduplicationSpend cube build

03

Ontology

One material, one place

Canonical materialCategory nodeVendorContractSpend fact

04AI

Intelligence

Categorise, harmonise, simulate

Categorisation modelembedding + LLM, confidence-scoredDuplicate material matcherConcentration & exposure analysisWhat-if simulatorconsolidation · risk · savingEval suitecategory precision

05Human

Human control

Taxonomy never rewritten silently

Category manager accepts / rejectsConfidence thresholdsTaxonomy governanceSourcing decision

06

Actions

Written back

Material master updateSourcing eventSpend reports

Observability

Every model call traced; evals run on real cases, not anecdotes.

Governance

Entitlements enforced at retrieval; rules versioned by the organisation.

Write-back

Systems of record are written only through the approval gate.

Category managers accept or reject every proposed match; the taxonomy is never rewritten silently.

// Impact

Impact

Material level
Category resolution, against coded labelsdesign intent
Governed
Hierarchy changes, proposed not applieddesign intent

Interested in Intelligent Purchasing?

Let's see how big your uncategorised bucket really is.

Get in touch
// End of case studyHub for Intelligent Purchasing