Quick Overview
AI-assisted rate-sheet extraction proposes structured rate records from supplier Excel files and PDFs. The difficult part is preserving the commercial meaning behind each value: its lane, unit, currency, validity, surcharge conditions and source location. A pricing reviewer should resolve uncertain or conflicting information before publication. This guide explains how to test extraction quality. For the wider receive-to-quote process, read the freight rate sheet automation workflow.
What AI Extraction Adds To A Rate-Sheet Workflow?
A conventional importer maps predictable columns into known fields. AI-assisted extraction may help interpret changing layouts, labels, tables and footnotes, but a plausible output is not proof that a rate is commercially correct. Compare proposed records with the original supplier file, including conditions outside the main table. Keep the source and reviewer decision attached to the resulting rate version.
This guide focuses on extraction quality and review. For the complete process of maintaining rates after ingestion, see freight rate-sheet automation.
Which file formats should you test?
Support varies by product and file structure. Ask vendors to process representative files rather than accepting a format name as evidence of reliable extraction.
| Source | Challenge to test | Review requirement |
|---|---|---|
| Excel workbook | Multiple sheets, merged cells, formulas and hidden context | Confirm which sheets and values were interpreted |
| CSV file | Encoding, delimiters and limited formatting context | Validate headers, units and missing conditions |
| Text-based PDF | Tables split across pages and footnotes | Connect each condition to the affected rates |
| Scanned PDF | OCR errors and unclear characters | Check critical numbers and identifiers against the image |
| Revised supplier file | Partial amendments and changed validity | Identify exactly which records are superseded |
Receiving a file successfully is not the same as extracting all its commercial meaning.
How To Test An AI Rate-Sheet Extractor
Use representative supplier files from your own workflow: a regular spreadsheet, a revised workbook, a text-based PDF and a scan if scans are part of your intake. Include files with separate surcharge tables, footnotes, mixed currencies, minimum charges and different validity triggers. Record which file types the proposed configuration actually accepts.
| Test | What the reviewer checks | Result to record |
|---|---|---|
| Field extraction | Price, currency, charge basis, lane, equipment and validity against the source | Correct, corrected, unresolved or omitted |
| Source traceability | Whether a proposed value can be tied to a sheet, row, page or document region | Source reference present or missing |
| Conditions | Whether footnotes and exclusions apply to the correct rates | Correct mapping or exception |
| Amendment handling | Whether the new file changes only the affected records | Correctly superseded and unaffected records |
| Publication control | Whether unresolved or unapproved records can reach quoting | Pass or failed control |
| Reviewer effort | Time spent correcting and approving the file | Minutes per document, with file type noted |
Do not compare products using only the number of imported rows. A workflow that imports every row but misses a currency or applies a surcharge to the wrong equipment can still produce unreliable pricing.
Extract values with their source context
A useful output includes more than origin, destination and price. Capture supplier, service, equipment, commodity, currency, charge basis, minimums, validity and restrictions where supplied.
Retain a reference to the source sheet, row, page or region where possible. A reviewer should be able to inspect the evidence behind a proposed value.
Do not treat an empty cell as zero. Blank, unavailable, included and not applicable are different states.
Normalize locations, currencies and units carefully
Match locations to the correct port, airport, inland point or tariff zone. A city name can refer to several facilities. Hold ambiguous matches for review.
Keep original and normalized values available. Recognizing a currency code is different from converting money: conversion requires a separately defined exchange-rate source, date and policy.
Likewise, a charge per container cannot be compared directly with a charge per shipment without checking its applicability. Preserve the equipment, unit and minimum associated with each amount.
Interpret validity and surcharge conditions
A date may relate to booking, departure, cargo receipt or another contract condition. Preserve that meaning when mapping validity fields.
Footnotes may limit a charge to selected lanes or equipment. Confirm whether surcharges are included, separate, conditional or awaiting confirmation.
A supplier amendment may change one component while leaving the rest of an agreement active. Do not replace the complete contract simply because a later file exists.
Use confidence scores as review signals
Some extraction products provide confidence scores; others do not. Ask what the score measures and whether it applies to a field, a row or a document.
A confidence score is not automatically a calibrated probability of commercial correctness. Clear text can still be mapped to the wrong charge or lane.
Set review priorities according to impact. Currency, price, unit, validity and equipment errors can materially change a quote even when the rest of the document is accurate.
Build an exception queue that can be resolved
| Exception | Evidence needed | Reviewer action |
|---|---|---|
| Ambiguous location | Supplier wording and candidate matches | Confirm the intended facility or zone |
| Missing currency | Source context or supplier clarification | Add a verified currency |
| Conflicting validity | Agreement and amendment references | Establish the applicable condition |
| Possible duplicate | Supplier, lane, equipment, charges and dates | Confirm duplicate or retain distinct records |
| Unclear surcharge | Footnote and calculation basis | Clarify inclusion and applicability |
| Low-quality extraction | Original document region | Correct the value or request a usable file |
Assign an owner and resolution status. Keep unresolved records out of published pricing until the required review is complete.
Example: a PDF footnote changes the interpretation
This example is hypothetical.
A tariff lists ocean rates for two container sizes and includes a footnote applying an additional charge to one size only. The extractor identifies the base rates but proposes the extra charge for both.
The reviewer checks the linked source region, corrects applicability and records the reason. The approved version retains the footnote and the structured condition.
This is why extraction accuracy should be assessed at the field-and-condition level, not only by counting successfully imported rows.
Approve publication and preserve recovery options
Require an authorized person to approve extracted contract rates before publication. Keep the imported source, reviewed output, corrections and approval reference.
A rollback should restore an appropriate published version without erasing the record of the change. It should not silently rewrite historical quotes or remove an amendment trail.
Test publication permissions and recovery in the proposed configuration. Confirm which actions are supported directly and which need an associated workflow.
Measure Extraction Quality
Assess a reviewed sample against the original supplier files. Report the file mix and include failed documents in the test results.
| Measure | Formula | Denominator |
|---|---|---|
| Critical-field accuracy | Correct critical field values ÷ critical field values checked × 100 | Reviewed values for price, currency, unit, validity, lane and equipment |
| Condition accuracy | Correctly mapped commercial conditions ÷ conditions checked × 100 | Reviewed inclusions, exclusions, footnotes and restrictions |
| Document processing coverage | Files producing reviewable output ÷ files submitted × 100 | All submitted test files, including failures |
| Source-reference coverage | Proposed values with inspectable source references ÷ proposed values checked × 100 | Values in the reviewed output |
| Unresolved exception rate | Records still requiring clarification ÷ proposed records × 100 | Proposed records in the test set |
| Time to approval | Approval timestamp minus file receipt timestamp | Approved files; show pending files separately |
Separate results for spreadsheets, text PDFs and scanned PDFs. A confidence score, where available, should guide review priority; confirm what it measures before interpreting it as an accuracy percentage.
Review A Representative Rate File With VelocityOS
Bring an actual supplier workbook or tariff, an amendment and a quote that uses the affected lane. Ask the VelocityOS team to show the extracted fields beside their source, how uncertain conditions are assigned to a reviewer and what must happen before a rate becomes available for quoting. Confirm supported formats and controls for your implementation.
FAQs
That depends on the file quality and the configured extraction workflow. Test your own scans and check critical numbers, units and footnotes against the image.
No. Ask what the score represents and inspect commercially important fields and conditions against the supplier source.
An authorized pricing or commercial reviewer should resolve ambiguity and approve the interpretation before publication.