AI-assisted product-catalog enrichment

Industry: wholesale distribution. Roles: product-data steward, merchandising manager, category or compliance owner, sales operations specialist, integration engineer. Capabilities: extract, classify, generate. Human control: approve before publication. Benefit: catalog-maintenance capacity and potentially fewer clarification requests. Costs: extraction, normalization, review, integration, and schema and source maintenance.

The decision

Supplier material arrives as spec sheets, price files, and catalog pages in inconsistent layouts. The product-data team wants to spend less time rekeying attributes while keeping the catalog trustworthy for search, ordering, and customer questions. A useful first boundary is a staging assistant that proposes structured attribute values, each linked to its evidence, for a steward to approve. Publication, identity decisions, and anything a customer could rely on for safety or regulatory purposes stay with named owners.

The unit of work is a new or changed SKU record that reaches an approved, attribute-complete entry against the distributor's category schema, or a correct hold that says which supplier data is missing. A record with 14 filled fields and one invented certification is not a success. A record held because the supplier never stated a required value is a correct outcome, not a failure of the tool.

All company names, documents, part numbers, schema rules, budgets, and thresholds in this chapter are original fictional teaching material. No supplier document, catalog record, or measured deployment is represented. The expected records are authored examples, not outputs from a tested model. External sources are cited where they appear and summarized in the research section.

Map the work before automating it

Today a steward finds the governing supplier document, confirms which product and packaging level it describes, reads specifications, converts units, picks permitted values, writes a description, and submits the record. Merchandising may rewrite copy, operations cares about pack quantities, and a category owner approves anything resembling a safety or compliance statement. Measure the whole path, including supplier clarifications and corrections after publication; record waiting time separately from active labor.

Workflow map

Supplier document → identity check → extracted values with spans → deterministic conversion and schema validation → flags → steward review → approved record → staged import → published catalog

An unknown or conflicting identifier leaves the path for identity resolution. A missing required value produces a supplier-data request and a hold. A safety, certification, or hazard attribute goes to its designated owner and is never populated from inference. A conflict between two current supplier documents is escalated to the supplier, not resolved by the model. If the assistant is unavailable, stewards continue with the existing manual process.

Compare against a simpler alternative

Much of this problem is solved without generative models when suppliers send structured data. A distributor can require an attribute template for new items, accept feeds from suppliers that already publish structured product data, and apply rules for units and permitted values on import. Optical character recognition plus templates can handle a supplier whose sheet layout never changes. Test three configurations on the same items: the current manual process, supplier templates with rule-based validation, and templates plus model-based extraction for the unstructured remainder. If most volume comes from a few suppliers willing to send structured files, the second configuration may capture most of the value.

Walk through one item

For your product data owner

The fictional distributor Quillmere Industrial Supply is adding a valve from the fictional supplier Brantwick Fluid Components. The category, schema, document, and values below exist only for this exercise.

Input: supplier spec-sheet excerpt

Brantwick BFC-BV-075 — Full-port ball valve, 3/4 in, FNPT × FNPT

Body: forged brass. Lead-free versions available.

Ball: chrome-plated brass. Seat: PTFE.

Maximum working pressure: 600 psi (non-shock, cold).

Temperature range: −4°F to 300°F.

Weight: 0.9 lb. Sold individually; case of 10.

The last valve you'll ever need — built for a lifetime of trouble-free service.

The sheet mixes units, carries a marketing claim, and leaves the body material ambiguous: "lead-free versions available" does not say whether this part number is one of them. It says nothing about drinking-water suitability or certification.

Input: category attribute schema (fictional)

The schema, not the model, decides what "complete" means. The rule that nominal size is a designation rather than a measured length is a schema decision made by Quillmere's steward for this exercise.

Expected internal staging package

Conversion factors come from NIST's published table and the arithmetic runs in ordinary code; the model only finds spans and proposes schema fields. Keeping the source value beside the normalized one lets a reviewer catch conversion or rounding errors.

Overall disposition: hold with two open items. The record cannot be published as complete because two required attributes lack support.

Expected customer-facing description draft

3/4 in full-port ball valve with FNPT threaded ends and PTFE seat. Rated to 600 psi (4,137 kPa) non-shock cold working pressure; temperature range −4°F to 300°F (−20.0 °C to 148.9 °C). Sold individually; case of 10.

The draft uses only approved attributes. It omits body material and certification until they are resolved, and it drops the marketing line. Merchandising can edit tone; it should not add facts.

Reject a plausible but unsupported record

Lead-free brass body, certified for drinking water. Size: 19.05 mm. Pressure: 600 kPa. The most reliable valve on the market.

This record guesses the body material, asserts an unsupported certification, converts a nominal designation into a length against the schema rule, attaches the wrong unit to the pressure, and repeats a superlative as fact. A buyer choosing a valve for a drinking-water line on the strength of it could be harmed. Reject it even if every other field is correct.

Change one input at a time

• A second current document conflicts. Brantwick's price file lists BFC-BV-075 at 400 psi. Expected behavior: show both values with their sources and dates, hold the pressure attribute, and ask the supplier which is current. Do not average, pick the more conservative value silently, or prefer the newer file without a versioning rule.
• The spec sheet is a phone photo of a printed page. Expected behavior: extract what is legible, mark low-legibility values for review, and request a clean document for any required attribute that cannot be read with certainty. A confident value from a blurred digit is a major failure.
• The seat is listed as "TFE". Expected behavior: flag the value as outside the permitted vocabulary and propose a mapping for the steward to approve or reject. The model should not decide that two material names are equivalent; an approved synonym table should.

Design the data, access, and system layer

For your technical lead

Start with one product category and a defined set of authorized supplier sources. For each value, store the supplier, document identifier and version, effective date where stated, the source span, the extracted text, the normalized value, the conversion rule applied, and who approved it. Keep extraction, derivation, and approval as separate records so a later audit can tell which step failed.

Identity first. Attributes attached to the wrong item are worse than missing attributes. GS1's GTIN Management Standard assigns a unique GTIN at every packaging level above the base unit and requires a new GTIN when the number of items in a case changes. GS1 GTIN Management Standard, rule 8 A proposed control follows: if a supplier document pairs a known case-level identifier with a case quantity that differs from the item master, flag it for identity review rather than updating the quantity. Whether a given supplier follows GS1 allocation rules is a fact to check, not an assumption.

Deterministic where possible. Unit conversion, rounding, required-field checks, numeric ranges, and permitted-value lists belong in code with tests. The model proposes; validation decides whether a proposal is well formed. A value that fails validation goes back to the steward, not to the model for another guess.

Safety and compliance attributes are owned, not extracted. Certifications, hazard classifications, and regulatory statements should come only from designated documents, approved by the category or compliance owner. GS1 requires a new GTIN when a certification mark is added to or removed from packaging, a sign of how consequential such claims are in product data. GS1 GTIN Management Standard, rule 5 This chapter makes no claim about which certifications apply to any category or jurisdiction.

Supplier documents are untrusted input. A spec sheet can contain text that reads like an instruction, the prompt-injection risk. The extraction step should have no write access to the catalog, and its output should pass through the same validation and approval as any other proposal.

Rights and retention. Confirm supplier agreements permit reuse of their content in the catalog and in a model provider's processing. Never use competitors' catalog text as a source. Record which documents may be retained for evaluation.

Publication controls. Stage imports, approve records individually at first, sample batches before bulk publication, and keep a tested rollback that restores the previous approved record. Description edits must not overwrite approved specification fields.

Choose models against this task

For your technical lead

Test a complete configuration: parser or OCR, model identifier, prompt and schema revision, vocabulary tables, conversion code, and review interface. A parser change alone can move accuracy on scanned sheets; see document extraction. Photographs of the product itself, such as packaging images used to confirm a pack count, are a computer vision task with its own evaluation.

Microsoft's Document Intelligence documentation, which is vendor guidance for one product, describes confidence as an estimated probability that an extracted result is correct, notes that not all fields return a confidence score, and suggests human review where accuracy is critical. Accuracy and confidence documentation Treat any confidence signal as something to calibrate against your own labeled items before using it to route review.

Compare configurations on correct values per required attribute, correct holds, unsupported additions, identity errors, reviewer minutes, latency, and cost per approved record. Use the model-selection guide to structure the comparison. Check each candidate's current specifications at the provider's own documentation on the day you test, and record that date. No model has been benchmarked on this fictional catalog task.

Work through the economics

For your finance lead

The monthly example below uses invented assumptions. An update is one new or changed SKU record carried to approval or a correct hold. Assisted effort includes review, rejected proposals, and correction work measured over the same window as the baseline.

That gives 480 assisted updates: 800 × 75% × 80%. Four minutes released per update is 480 × 4 ÷ 60 = 32 hours a month, worth $1,280 at $40/hour, or $580 after the recurring budget, before setup.

An illustrative eight-minute baseline: two minutes to locate the document and confirm identity, three to read and key attributes, one for conversions and permitted values, one for the description, and one for a final check and submission. An illustrative four-minute assisted path: one minute for identity confirmation, two for checking proposed values against their spans, half a minute for the description, and half a minute to resolve flags and submit. These allocations show where the saving is supposed to come from; they are not measurements.

Make continuing work visible

Split the hypothetical $700 monthly budget into $250 for software and usage, $200 for schema and vocabulary upkeep, $150 for quality sampling, and $100 for integration monitoring. The split is invented, not a quote. Upkeep absorbed by existing staff is consumed capacity, not cash, and review already inside the assisted minutes is not deducted again; the economics chapter explains both.

Stress-test review time and adoption

At base volume and value, the recurring budget needs 700 ÷ (480 × $40 ÷ 60) ≈ 2.19 minutes released per assisted update just to break even on capacity value, before setup. The margin is thin: halving adoption alone turns the base case negative. A low-volume category may not justify a dedicated assistant on labor alone.

Explain how capacity becomes money

If every released hour replaced paid work, such as a contracted data-entry service the distributor would otherwise buy, the illustrative net cash benefit would be $1,280 − $700 = $580/month, and the $8,000 setup would pay back in about 13.8 months, assuming immediate steady state. If only half the capacity value can be realized as avoided spending, the result is $1,280 × 50% − $700 = −$60/month: the assistant costs more cash than it saves.

Released time may still clear a backlog of incomplete records; evaluate that as a quality outcome, not a second cash benefit from the same hours. Better attributes might also improve search, reduce order errors, or lower returns. Those are revenue and loss-avoidance hypotheses, unmonetized here because nothing measures them; test them separately if they matter to the decision.

The economics chapter separates benefit location from accounting classification. Finance should map costs under applicable standards and company policy rather than assuming a label from the use of AI.

Define failure severity before the trial

For whoever runs the pilot

These are proposed pilot categories for owners to adapt and agree before seeing results.

Track errors caught in review separately from published errors; careful stewards can otherwise hide a weak assistant.

Build the evaluation set

For whoever runs the pilot

As an initial coverage plan, assemble 150 items from authorized historical updates or constructed examples: 65 routine items with complete documents, 20 with a missing required attribute, 15 with conflicting current documents, 15 with unit or nominal-designation traps, 10 scanned or low-quality documents, 10 similar variants or packaging-level differences, 10 with safety or certification language present or implied, and 5 with instruction-like text embedded in the document. These counts are illustrative, not a statistically justified sample. Keep this challenge set separate from a sample reflecting the real mix of categories and suppliers.

Label a reference record for each item before testing, using only documents available at the time of the original update. Split by supplier and product family so near-identical sheets do not appear in both development and evaluation. Reserve the held-out set before tuning prompts or vocabulary tables.

Record for each item: item ID, category, supplier, document versions, configuration, proposed values, spans, flags, reference values, identity correctness, per-attribute correctness, unsupported additions, severity, reviewer minutes, cost, and final disposition. Report with denominators: correct values among populated required attributes, correct holds among items needing one, unsupported claims among items with safety language, identity errors among all items.

For an initial offline gate, a team might require zero observed unsupported safety or certification claims, zero identity errors, a correct hold on every item missing a required value, and at least 97% of populated required attributes matching the reference on routine items. These are illustrative management thresholds. A clean result on a small set bounds the true failure rate only loosely; pilot design shows how loosely.

Run a supervised pilot and make the decision

Run offline replay first with nothing published, and fix schema, vocabulary, and identity problems before live use. Then pilot one category with every record approved individually in staging; test bounded batch approval with sampling only afterwards. Where practical, assign comparable updates to assisted and baseline handling, balanced by supplier and category. Measure all assigned updates, including fallbacks and rejected proposals, and follow each record for a fixed window to catch later corrections.

The product-data steward owns attribute approval; merchandising owns descriptions; the category or compliance owner approves safety attributes; operations confirms pack and unit fields; engineering owns validation, imports, and rollback; finance reviews realization assumptions. Before expanding:

• Schema, required fields, permitted values, and severity rules were agreed in advance.
• Identity checks cover packaging levels and passed the agreed tests.
• Safety and certification attributes cannot be populated without an owner-approved source.
• Held-out items include missing values, conflicts, scans, variants, and embedded instructions.
• Human effort includes review, flags, rejected proposals, and later corrections.
• Results beat supplier templates with rule-based validation on the same items.
• Vocabulary and schema upkeep has an owner and a budget.
• Any cash benefit names a specific avoidable expense.
• Rollback was tested on a real staged import.

People and skills

In this proposed design, stewards spend less time typing and more time judging evidence, resolving flags, and maintaining the schema and vocabulary that make validation possible. Category knowledge matters more: a steward must recognize a technically valid but implausible value. In this design, a hold for a missing required value produces a supplier request, so supplier management matters too. The actual shift in effort remains unmeasured.

Who you need

This pilot needs the roles below, named by responsibility rather than job title. Some can come from outside for the build, but whatever the business owns after go-live, including the schema, the item master, evaluation, and approval, stays in-house.

Staffing an AI pilot sets out the method behind this table, and Choosing and working with an AI partner covers bringing in outside help.

When to defer

Defer if supplier data rights are unclear, if item identity and packaging levels are unreliable in the item master, or if the schema itself is not agreed. Defer for categories where most suppliers already provide structured data that meets the schema, and for low-volume categories where the recurring budget cannot be covered. The simpler alternative is a supplier attribute template, rule-based import validation, and a short list of priority attributes maintained by hand.

What external research establishes

For your analyst

GS1 defines the GTIN as an identifier a company can use to uniquely identify its trade items, meaning products or services priced, ordered, or invoiced at any point in the supply chain; its allocation rules for packaging levels and certification marks are cited above. GS1 GTIN GS1 describes its Global Data Synchronisation Network (GDSN) as a way for trading partners to share product master data by subscribing to a certified data pool. GS1 GDSN Its Global Data Model defines a consistent set of foundational product attributes in four layers: global core, global category, regional category, and country or local. GS1 Global Data Model These are standards-body descriptions of their own programs. GS1 frames the Global Data Model around brand owners and retailers; its fit for a particular B2B category, and whether a distributor's suppliers participate, needs category-specific review.

NIST publishes the conversion factors used in the worked case; it marks the inch-to-meter factor as exact, while its pound-to-kilogram and psi-to-pascal factors are given to seven significant figures. NIST Guide to the SI, Appendix B.8

None of these sources establishes the accuracy of AI-assisted extraction, reviewer effort, or effects on search, orders, or returns. No such measurement was found or performed for this chapter.

Evidence still needed and related reading

Obtain an attributable operator account of the current enrichment process, a permissioned set of supplier documents with reference records, category-specific attribute and compliance requirements, measured reviewer time, and correction rates after publication. No operator has been contacted and no catalog outcome verified for this chapter.

Continue with wholesale distribution, purchase-order intake, document extraction, sales operations leaders, human review, and the economics worksheet. When comparing products, use build, buy, or configure and vendor diligence.

Companion templates

Prepare your pilot with the pilot kit, evaluation scorecard, and evidence checklist. The companion workbook provides blank economics inputs and case records; worked chapter figures remain illustrative examples.