Choose the check that creates the most rework
An invoice arrives by email, the packing list has another revision and the purchase order uses different internal item codes. Teams spend time reconciling the same shipment across spreadsheets and messages. A first pilot can organize this set, extract fields and prepare a discrepancy queue with sources and owners.
Start with one supplier, product family and transport mode. AI prepares document analysis; customs treatment, tariff classification and declaration decisions remain with accountable professionals. The expected outcome is a reviewable package with less manual searching and fewer late corrections.
Define the checklist for the actual operation
Brazil’s Receita Federal lists documents such as transport documents and commercial invoices, with packing lists and other evidence depending on the operation. Its official page also describes conditions and exceptions. Use it as a reference for a specialist to define the applicable checklist rather than treating a generic list as a universal requirement.
For the pilot, record document type, expected issuer, version, purchase-order and shipment links, fields required for checking and the owner responsible for missing items. Missing documents must appear as pending; AI must not create substitutes that appear issued by a supplier or carrier.
Extract with source, unit and version
For each field, preserve the original value, normalized value, document, page and revision. The data contract can include supplier, buyer, purchase-order reference, currency, item, quantity, unit, price, package count and weights. Define an approved mapping between supplier codes and internal master data.
Do not mix boxes with units, gross with net weight or USD with BRL values. A decimal separator or OCR error can change a total. Route unreadable fields for review; do not fill numbers based on textual plausibility. A new revision should retain its predecessor and identify the current version.
Check a partial shipment using explicit rules
Synthetic example: a purchase order specifies 120 units at USD 20, totaling USD 2,400. The invoice states 100 units and USD 2,000; the packing list shows five boxes of twenty units. Invoice and packing list agree on 100 units, but twenty remain against the order. The conclusion depends on confirming whether partial shipment was authorized.
The agent should show the difference and evidence, not correct the invoice to USD 2,400 or declare a physical shortage before receipt. Totals, unit conversions and tolerances should use approved deterministic rules. Comparing documents identifies a pending issue; it does not establish the actual contents of the cargo.
Keep the DI or Duimp route under review
Siscomex maintains a DI phase-out schedule with special situations, unavailable operations, a flowchart and version history. The page consulted on October 28, 2026 directs users to check rules for the operation and notes that dates may be revised. Do not hardcode one migration date into an agent for all imports.
Record the source and version the responsible professional used to choose the process route. An agent can flag potentially outdated guidance and prepare questions for a specialist. This blueprint does not assume native Zild integration with Portal Único, automatic declaration submission or model access to a digital certificate.
Coordinate documents, logistics and finance
Every discrepancy needs an owner: procurement confirms the order and terms; the supplier corrects documents; logistics tracks shipment and receipt; the customs owner validates the package; finance checks obligations and due dates. Record decisions, evidence and deadlines without hiding pending issues behind a summary.
Separate expected arrival, confirmed arrival, customs clearance and warehouse receipt. In finance, distinguish commercial invoices, payables, payments and import costs. Taxes, applicable exchange rates and customs value require their own calculation and validation; a generated summary does not replace these records. Give each partner only documents authorized for their role.
Evaluate extraction and decisions separately
Build an anonymized sample with different suppliers, poor scans, revisions, differing units and partial shipments. Have the team label correct fields and actual discrepancies before testing. Separate files used to tune the agent from final-evaluation files.
Measure critical-field accuracy, relevant discrepancies found, false alerts, review time and cost per approved package. OCR with strong average accuracy can still miss currency or quantity. Do not count an ignored critical issue as success because the summary reads well. Compare the same operation mix and record evaluated volume.
Prepare a document pilot with Zild
The document API accepts files and processes them asynchronously; retain the upload identifier and correlate completion before using output. Deployment can combine Admin for access and connectors, Chat for queries and Flow for pending work. ERP and logistics integrations need confirmed read and write capabilities.
Bring an anonymized purchase order, invoice, packing list and transport set with an already-resolved discrepancy to the assessment. Start read-only, produce a review package and then evaluate approved system actions. The goal is to connect documents and owners to a traceable decision before expanding suppliers and operations.
Official sources consulted
Connect one operational exception
Gather this information to scope the work with Zild. During the conversation, we assess integrations, owners and how to measure outcomes.
- A shipment or delivery with a resolved exception
- Events and systems showing progress
- The owner and deadline for the next action
Describe the process in the form; samples and documents can be prepared later without confidential data.
The form opens with this guide’s title. You can edit everything before sending.