Define the tool’s task and the decision owner

For a catalogue containing many SKUs, distinguish finding candidate categories, examining classification reasoning and authorising data for a declaration. WCO’s 3 March 2022 BACUDA announcement describes recommendations learned from historical declarations. It does not certify a commercial product or a result for this shipment.

CEP Insight proposes the workflow below after reviewing the listed sources on 12 October 2026. Define the destination, tariff edition, product version, required code depth and responsible reviewer. No model, API or customer pilot was operated for this article.

Read research results within their evaluation scope

HSCodeComp, published at ACL in July 2026, is research by Alibaba Group authors. It uses 632 products across 32 categories. Table 1 reports 46.83% ten-digit accuracy for the best listed agent configuration. These are author-reported benchmark results, not independently reproduced by CEP Insight.

The paper’s expert-labelled task and selected model configurations do not establish performance on another catalogue, jurisdiction, edition or current service. USITC distinguishes common six-digit structure from US-specific extensions. Do not label all ten-digit targets as universal HS codes. Benchmark labels are research references, not decisions issued for your goods.

Evaluate the pilot with explicit denominators

CEP Insight recommends an agreed test set with evidence-reviewed reference decisions and deliberately selected difficult cases. Keep new products or held-out families identifiable rather than relying only on familiar repeated descriptions. Record missing input, disputed references and cases the tool cannot resolve. Use the same scope and denominator when comparing tools. The table proposes measures; it reports no results from a CEP Insight trial.

Editorial measures for an AI classification pilot
MeasurePopulation and calculation to defineInterpretation limit
Complete-code accuracyCorrect full destination codes / all cases tested at the stated code depthA prefix match is not a complete-code match.
Candidate-list coverageCases with the reference in the stated top-k list / all tested casesA listed candidate still requires selection and reasoning.
Coverage and abstentionAnswered, incomplete and escalated cases / all tested casesReport accuracy on accepted outputs and coverage separately.
Review and reworkReviewer time, corrected outputs and repeat queries for the same casesMachine response time alone does not measure operating effort.
Evidence and scope checksCases with verified citations, edition and product match / all tested casesA plausible explanation or working link does not prove applicability.

Proposed measures from CEP Insight. Agree reference quality, sample composition, code depth and treatment of disputed cases before testing; no observed scores or acceptance thresholds are claimed.

Separate candidate codes from approved product records

Store product facts and code decisions as linked records. A supplier code, an AI candidate and an approved destination record may coexist with different owners and scopes. Do not silently replace one with another. For batch imports or API exchange, agree what each state means and which state may be passed to the declaration partner.

Retain SKU and revision, destination, tariff edition, candidate list, source passages, review reasoning, reviewer, decision date and superseded record reference. The state model below is editorial guidance, not a legally prescribed workflow.

Editorial states for a product classification record
StateRecord to retainCondition before the next step
Facts incompleteKnown facts, missing fields and supplier/evidence ownerObtain the evidence needed to examine the classification.
Candidate generatedTool/version, input, candidate codes and source referencesReview against actual goods and applicable legal provisions.
Under reviewCandidate comparison, disputed facts and reviewerResolve material questions or assign escalation.
Approved for defined useReviewer, reasoning, destination, edition, product version and dateUse only within the recorded scope and agreed responsibilities.
Reopened or supersededChange trigger and links to earlier and replacement decisionsComplete the new review and preserve the earlier history.

This is a proposed record model. Internal approval is not a customs ruling, guarantee of compliance or confirmation of release.

Make review and escalation visible in batch work

Route incomplete descriptions back to the product-information owner. Ask the reviewer to resolve conflicting candidate headings, unsupported explanations or disagreement with a relevant legal note. A score supplied by a model needs evaluation against observed errors; a high score alone is not review evidence.

For recurring uncertainty, define who decides whether further technical evidence or a jurisdiction-specific ruling is needed. Preserve cases that cannot be decided and their next action. A larger batch does not justify treating unknown fields as confirmed.

Check the service’s evidence and data conditions

Ask the provider which jurisdictions, code levels and tariff editions are covered; how legal sources are updated; and whether outputs link to the actual passages used. Request the evaluation population, error examples, review effort and limits of any reported accuracy. Compare the scope actually demonstrated.

Before connecting a catalogue, agree which fields are sent, who can access them, retention, permitted reuse, export of decision records and service-change notification. Include the human review and exception workload when comparing cost. These are proposed selection questions, not ratings or endorsements of a vendor.

Reassess changes without overwriting the history

USITC notes that editions contain change records. CEP Insight recommends triggers for product composition or function changes, destination changes, tariff updates, relevant decisions and observed errors. Reopen affected records for review and retain the earlier reasoning and effective period.

Link the approved record to the invoice and declaration version actually used. Classification remains separate from origin, customs value, taxes, restrictions and release. Consult the companion classification, ruling-preparation and document-reconciliation articles for those decisions.

Conclusions

An AI recommendation becomes operationally useful when its scope, evidence, uncertainty and reviewer remain traceable. Evaluate the full workflow on a defined catalogue before extending it to further products or markets.

Sources & further reading

Industry knowledge and interpretation; carrier terms and local requirements vary.

Sources reviewed:

  1. WCO — BACUDA neural-network model for HS recommendations; 3 March 2022; historical project announcement
  2. Lan et al., Alibaba Group — HSCodeComp; ACL 2026, July; corporate-authored conference research; methodology and Table 1
  3. USITC — Tariff classification FAQ; legal hierarchy, country-specific digits and edition changes; US guidance, undated

Related reading

Enquiries related to this article

This article can provide background for an industry, research or cooperation enquiry. Review and edit the brief before contacting the team.

Prepare a brief with this reading

Contact CEP Insight

Contact on WhatsApphi@cepinsight.comContact on Telegram