An airline buys goods and services across a wide range of operations. Similar purchases can arrive under different names, while similar words can describe very different needs. Embedding Labs addressed that ambiguity in a focused taxonomy and mapping assignment for an airline's procurement team: map commodity descriptions into an existing four-level reference taxonomy and make uncertain decisions visible for review.
01 / The work
The category shapes the view of spend
A purchasing category gives a business a way to compare related activity. If repair work, replacement parts and equipment rental are grouped by a shared word alone, the resulting view can hide distinctions that matter to a category owner. A description, an accounting code and a commodity identifier each provide context; none should silently stand in for the others.
The assignment made the decision explicit. Each proposed mapping needed a place in the hierarchy, an identifiable source item and enough explanation for a specialist to challenge the result. The useful output was a crosswalk that could be reviewed alongside the original purchasing information.
02 / The work
Use AI to prepare a more useful decision artifact
The team developed a retrieval-assisted classifier that consulted the reference taxonomy and returned a structured category proposal with reasoning. Its instructions favored an existing category when the meaning fit, reserving new lowest-level categories for gaps in the reference.
A separate analysis method tested exact codes, normalized descriptions, fuzzy wording and semantic similarity in sequence. The two approaches offered different ways to examine the same classification problem. They produced candidate mappings and review views rather than a single unexplained label.
03 / The work
Make ambiguity a specific conversation
New category proposals, low-confidence suggestions, duplicates and unmapped items were organized into outputs for review. The original identifier and description let a reviewer ask a concrete question: does this item belong here, does the taxonomy lack a category, or is the source record ambiguous?
A confidence score can focus attention, but category meaning requires procurement judgment. Identity coverage and duplicate handling also matter: a plausible category for one representative description does not resolve every record that shares its wording. The work made those issues visible in the mapping artifacts and analytical review.
04 / The work
What the assignment delivered
The assignment delivered taxonomy and mapping artifacts. The practical contribution was a shared classification language and a set of decisions the customer could inspect. The customer imports and uses the delivered mapping.
The case shows how AI and conventional analysis can prepare a structured decision artifact while leaving category acceptance with the people responsible for purchasing.
The system gives our team a consistent way to connect purchasing categories across different taxonomies. It separates straightforward matches from the cases that need specialist judgment, so we can focus our attention where it matters and understand the basis for each proposed classification.
The lesson
A useful procurement mapping preserves the item, explains the category and exposes the decisions that still need specialist judgment.
Go inside the design, integrations and delivery decisions.
Explore the technical study
