productivity

Reducto

Name: Reducto
Availability: InStock
Author: AI Demos

Structured PDF extraction with nested JSON, citations, and row-preserving tables; invoices worked cleanly, while bank statements still need review.

Visit Reducto

Nested JSONLine itemsCitationsConfidence scores

TL;DR — our verdictUpdated July 2026 · 7 test artifacts

Invoice extraction was strong; bank-statement reconciliation was not fully trustworthy.

Where it wins

You want to upload invoices or bank statements and get nested structured JSON back.
You need citations, bounding boxes, and confidence scores for review workflows.
You can validate and correct bank-statement outputs before downstream use.

Main limitation

You need exact bank-statement transaction counts and summary totals with no post-processing.

Pricing (verified plans)

Standard Pay-as-you-goGrowth Custom pricingEnterprise Custom pricing

Strongest test artifacts

Schema-Driven Nested Extraction →Repeated-Record Table Extraction →Source Citation and Confidence Annotation →

Our take

Reducto did a solid job turning the invoice PDF into nested JSON, preserving all eight line items and attaching citations plus confidence metadata for review. The bank-statement run was less dependable: transaction counts, summary totals, and several transaction-level fields were off or missing. Across both documents, it looks best when the schema is clear and a human can validate the extracted finance data before downstream use.

Screen recording of the invoice extraction workflow, including schema-driven JSON output and review-oriented metadata.

In-Depth Review

Our detailed analysis of Reducto — features, performance, and real-world testing.

AI Demos Team

Expert Reviewer

Verified Review

Feature-by-Feature Breakdown

Schema-Driven Nested Extraction

Mixed

▾

Test Summary

Feature tested: Schema-Driven Nested Extraction

Result: Partial — Mixed

Reducto extracts structured fields from uploaded business PDFs into nested JSON that follows a supplied schema. It was exercised on an invoice and a bank statement, pulling metadata, advertiser and station details, billing and remit blocks, flight dates, line items, and summary values.

pdf

Bank Statement PDF.pdf

↓→

json

reducto-bank-statement-output.json

Loading file...

Parsed bank-statement JSON with metadata, account-holder, branch, and statement-period fields; the transaction section was only partially populated, and the totals did not match the source document.

pdf

Invoice PDF.pdf

↓→

json

reducto-invoice-output.json

Loading file...

Invoice JSON reconstruction that populated the nested schema, including invoice metadata, advertiser details, station fields, billing and remit information, line items, and summary fields.

Bottom Line

Strong when the document layout maps cleanly to the requested schema; the invoice was a clean success, while the bank statement still needed review for transaction completeness and totals.

Repeated-Record Table Extraction

Mixed

▾

Test Summary

Feature tested: Repeated-Record Table Extraction

Result: Partial — Mixed

Reducto preserves tabular repeats as separate rows instead of flattening them. On the invoice it kept eight line items as distinct records, and on the bank statement it returned transaction rows for reconciliation.

pdf

Invoice PDF.pdf

↓→

image

Output artifact for "Repeated-Record Table Extraction" test: Expanded line_items array showing the invoice rows kept separate, including all eight line items., reducto-invoice-extracted-line0items.png

Expanded line_items array showing the invoice rows kept separate, including all eight line items.

pdf

Bank Statement PDF.pdf

↓→

image

Output artifact for "Repeated-Record Table Extraction" test: Expanded transactions array showing individual bank-statement rows; the output still had count and field-completeness issues., reducto-extracted-transactions.png

Expanded transactions array showing individual bank-statement rows; the output still had count and field-completeness issues.

Bottom Line

Excellent row preservation on the invoice; usable but not trustworthy enough on the bank statement without reconciliation.

Source Citation and Confidence Annotation

Strong

▾

Test Summary

Feature tested: Source Citation and Confidence Annotation

Result: Passed — Strong

Reducto attaches field-level source citations, bounding boxes, and confidence signals to extracted values for review when enabled (this is a default feature). The invoice and bank-statement outputs both exposed provenance objects and confidence metadata for auditing.

pdf

Invoice PDF.pdf

↓→

image

Output artifact for "Source Citation and Confidence Annotation" test: Invoice number extraction with citation bbox and granular confidence metadata., reducto-invoice-metadata.png

Invoice number extraction with citation bbox and granular confidence metadata.

pdf

Bank Statement PDF.pdf

↓→

image

Output artifact for "Source Citation and Confidence Annotation" test: Customer_name extraction backed by a text citation with page and bounding-box metadata., reducto-bank-statement-citation-annotated.png

Customer_name extraction backed by a text citation with page and bounding-box metadata.

pdf

Bank Statement PDF.pdf

↓→

image

Output artifact for "Source Citation and Confidence Annotation" test: Transaction text annotated with high confidence and granular parse confidence, showing the review metadata Reducto exposes., reducto-bank-statement-confidence-scores.png

Transaction text annotated with high confidence and granular parse confidence, showing the review metadata Reducto exposes.

Bottom Line

The audit trail is a real strength: both documents attach value-level provenance and confidence metadata, making review and correction practical.

Published plans from the research report

Standard

Pay-as-you-go (first 15,000 credits/month free, then $0.015 per credit)

Includes Parse, Extract, Edit, and Split APIs; supports 30+ file types; up to 5 Studio seats.

Growth

Custom pricing

Volume discounts, higher rate limits, priority support, EU/Australia data residency, zero-retention option, and unlimited Studio seats.

Enterprise

Custom pricing

VPC or on-premises deployment, custom SLAs, SSO/SAML, RBAC, custom pipelines, and dedicated enterprise support.

Pricing and plan notes come from the research report; the Standard plan includes Parse, Extract, Edit, and Split APIs and supports 30+ file types.

✓ Use This If

●You want to upload invoices or bank statements and get nested structured JSON back.

●You need citations, bounding boxes, and confidence scores for review workflows.

●You can validate and correct bank-statement outputs before downstream use.

●You need row-preserving extraction for repeating records like line items or transactions.

✕ Skip This If

●You need exact bank-statement transaction counts and summary totals with no post-processing.

●You need a fully hands-off reconciliation workflow.

●You need every extracted field to be complete on the first pass for a noisy statement.

●You need strict schema-only output with no extra fields outside the requested structure.

productivitypdf-toolstext

The report tested two document types: an invoice PDF and a bank-statement PDF.

It returned structured nested JSON, not plain OCR text. The invoice output was organized into blocks such as invoice metadata, advertiser, station, line items, and summary, while the bank statement output included metadata, account-holder, branch, statement-period, transactions, and summary sections.

The invoice run was strong. It preserved all eight line items as separate records, reconstructed the requested nested structure, and matched the invoice totals in the summary. The report also noted an extra idb_number field that appeared even though it was not in the requested schema.

The bank-statement output was less reliable. The report says the transaction count came out as 53 versus 51 expected, the summary total_transactions was 70 versus 40 correct, and several transaction-level fields such as transaction_id, transaction_type, date, and value_date were missing or inconsistently populated.

Yes. The research shows citation objects with page and bounding-box metadata, plus confidence information at the field or granular level. The invoice screenshots and bank-statement screenshots both expose that review metadata.

Yes. In the invoice run, the report noted that idb_number appeared in the output even though it was not part of the supplied schema.

The report lists a Standard plan that is pay-as-you-go with the first 15,000 credits per month free and then $0.015 per credit. It also lists Growth and Enterprise as custom-priced plans.