Freight Invoice Extractor

A freight invoice extractor turns carrier bills, whether PDF scans, emailed PDFs, or EDI 210 / 410 payloads, into structured line items a clerk can map to shipment and rate records. The model reads charge codes, weight, miles, accessorials, fuel, and totals, then proposes a field-level mapping instead of forcing someone to rekey every row.

Speed is the primary outcome. Extraction finishes in seconds so the clerk starts on exceptions, not transcription. Each proposed field carries a source span (page, region, or EDI segment) and a confidence score. Empty fields stay empty when the document does not support a value. The clerk still confirms the mapping before anything posts to accounts payable or the TMS.

Carrier invoices are dense and inconsistent. The same accessorial may appear as “DET,” “detention,” or a free-text note. Rate sheets and contracts define what should have been billed; the invoice shows what the carrier claims. Matching those two views requires more than character recognition.

EDI helps when partners send clean 210s, but many lanes still arrive as PDFs with tables, multi-page summaries, and stamped adjustments. Template OCR breaks when layout drifts. An LLM-based extractor treats the invoice as a document plus a schema: expected fields for this carrier or mode, plus optional links to shipment ID, PRO, BOL, and contracted rates.

That design keeps the pipeline honest. If a fuel surcharge line is missing or illegible, the system leaves the field blank and flags low confidence rather than inventing a number. Downstream processes, including automated three-way match and invoice anomaly detection, then operate on explicit gaps instead of silent guesses.

Typical flow:

1. Ingest PDF or EDI and normalize pages or segments.
2. Extract candidate fields and line items with provenance and confidence.
3. Resolve shipment keys (PRO, BOL, container, order) against TMS or warehouse records.
4. Propose rate-record matches for base rate, fuel, and named accessorials.
5. Present a confirmation UI where the clerk accepts, edits, or rejects each mapping.

Line items matter more than a single total. A clerk who only sees “invoice amount $4,812.40” cannot audit detention versus layover. Structured lines let rate audit and finance see the same breakdown the carrier used.

Confidence should be field-local. A high-confidence invoice number next to a low-confidence detention amount should not average into a misleading “good” score. Source spans make review fast: the clerk clicks the field and sees the exact PDF crop or EDI element that produced it.

When booking intake still arrives as email or unstructured attachments, a related pattern is the unstructured booking intake parser. Keep those stages separate so invoice confirmation stays focused on billed charges, not order creation.

Several vendors cover parts of this problem with different centers of gravity.

Cass AI sits close to freight audit and payment. Teams already on Cass often want extraction that feeds audit rules and carrier dispute workflows rather than a generic AP inbox.

Billtrust is stronger on broader accounts receivable and invoice presentment for B2B billing. Logistics teams evaluating Billtrust usually ask how well freight-specific charge codes and multi-stop invoices survive into structured AP data.

UiPath approaches the problem as document understanding plus orchestration. It fits when extraction is one step in a larger RPA path: pull email, extract, write to ERP, open a human task for low confidence. The risk is treating freight semantics as just another form template unless the schema and rate mapping are modeled explicitly.

ABBYY has long strength in OCR and intelligent document processing. It can deliver solid layout and table capture; LLM or classification layers still need to map captured text to shipment and contract rate records with empty-field discipline.

None of these replace clerk confirmation for high-dollar or disputed freight. They compress the time from “invoice received” to “mapping ready for human review.” Choose by whether you need freight-audit depth, general AP automation, orchestration, or classic IDP accuracy as the foundation.

Empty fields stay empty. That rule prevents false rate matches and poisoned anomaly scores. Prefer a null with a reason code (“not found,” “illegible,” “ambiguous”) over a default of zero.

Cite every field. Source span plus confidence turns review into verification, not re-reading. Store those citations with the audit trail so disputes can show what the system saw.

Keep the clerk in the loop for mapping confirmation, especially when shipment resolution is ambiguous (duplicate PROs, split loads, corrected BOLs). Automation should draft; posting should wait for an explicit accept where policy requires it.

Separate extraction errors from commercial disputes. Bad OCR of a detention window is an extraction defect. Correct extraction of an unauthorized accessorial is a billing discrepancy for rate audit and carrier ops. Mixing those buckets hides root cause.

Measure speed as time-to-confirmed-mapping, not time-to-first-parse. A fast wrong extract that the clerk redoes from scratch is not a win. Track edit rate by field, override rate by carrier, and share of invoices that clear without manual keying.

Start with a closed schema per mode or carrier family: truckload, LTL, parcel, ocean, air. Expand fields only when production invoices show demand. Over-wide schemas increase empty noise and dilute confidence calibration.

Train or evaluate on your real mix: multi-page PDFs, scanned faxes, EDI with sparse accessorial segments, and invoices that reference multiple stops. Score extraction at line-item and field level, not only invoice total.

Wire outputs to shipment and rate systems of record before you expand volume. Extraction without a reliable join key becomes a second inbox. Once joins are stable, anomaly detection and three-way match become much more useful because they compare structured claims to contracted expectations.

Roll out by carrier volume and dispute cost. High-volume, stable layouts give early speed gains. Complex or adversarial invoices stay on higher human review thresholds until confidence and citation quality earn trust.

The durable pattern is narrow: parse PDF and EDI into structured lines with provenance, propose mappings to shipment and rate records, leave unknowns blank, and let the clerk confirm. That is how freight AP confirmation gets faster without pretending the model owns the final commercial decision.