12 platforms have reviewed primary-source claims (the original 35 records plus the added fields below); 4 remain uncovered. Missing values are unknown from reviewed sources.
Can you trace the number back to the page?
Compare the output you can inspect, rather than treating a confidence score as an accuracy result.
Amazon Textract
Page and field provenance
Analyze Lending returns the physical page, detected page number and document class. Extracted fields can carry text, confidence and bounding geometry.SourceMortgage-oriented Analyze Lending; coordinates and confidence do not establish commercial-loan accuracy or cross-file checks.
Missing documents and signatures
The lending summary groups document pages and lists undetected document types and detected or undetected signatures.SourceThis is the API’s recognized document set, not proof that every required commercial-loan document is checked.
Analyze Lending example rate
AWS’s US West (Oregon) example lists $0.07 per supported page through one million pages per month, then $0.055 per page above that.SourceAnalyze Lending mortgage-document API; regional pricing example, not a universal quote or a commercial-loan accuracy result. Dollar symbol is as displayed by AWS.
Which pages the example bills
In AWS’s example, a 200,000-page packet produces 100,000 pages with supported classification and extraction; the bill uses those 100,000 pages.SourceThis is the vendor’s pricing illustration. Confirm how your document mix is classified and billed; packet pages and billable pages are not assumed equal.
Azure AI Document Intelligence
Page and field provenance
Bounding regions identify the page and polygon around an element. PDFs use inches; images generally use pixels.SourceThe cited API v4.0 documentation says bounding regions are not returned for unrendered DOCX, XLSX, PPTX or HTML files.
Extracted versus inferred values
An extracted field can include source content. An inferred field has a value but no content property.SourceA returned value may therefore need a different review path from a value copied from the page. This is an editorial consequence of the documented response types.
Table structure versus meaning
Table cells expose row and column positions. Mapping a layout table to named business fields can require post-processing or a prebuilt or custom model.SourceA detected table alone does not establish borrower-level financial reconciliation.
Docsumo reads business tax returns, financial statements and bank statements (Vendor-authored comparison attributes these document types to Docsumo; does not establish tested recall or coverage.)Source
Cross-doc & review
On the Enterprise plan, it checks them against the application and each other (Vendor-authored comparison claim about cross-document checks; plan-tier detail and actual behavior require validation.)Source
SOC 2 Type 2, ISO 27001, HIPAA, GDPR on every plan; no third-party/shared model training (Vendor security/pricing statement; independent audit report requires vendor access.)Sourceworks with Encompass, MeridianLink, nCino, Blend and any other LOS with an API, through Docsumo's API and webhooks (Vendor-authored comparison claim; no integration implementation or compatibility test performed.)Source
Mortgage lenders: Mortgage processing involves a heavy flow of documentation (Vendor article identifies mortgage lender use; adjacent commercial mid-market fit is not established.)SourceSMB funding and mortgage cash-flow/income underwriting (bank statements, pay stubs, tax forms) + fraud detection (Vendor product homepage; commercial mid-market tax-return/financial-statement accuracy not independently tested.)Source
Cross-doc & review
Ocrolus says it offers human-in-the-loop verification via API and dashboard. Cross-file checks remain unknown from these sources. (Vendor product claim; specific supported LOS list and webhook SLAs require documentation/demo check.)Source
Price / native unit
Unknown from reviewed sources
Security & LOS/API
Unknown from reviewed sourcesAPI and dashboard with LOS integrations and human-in-the-loop verification (Vendor product claim; specific supported LOS list and webhook SLAs require documentation/demo check.)Source
Analyze Lending is a document processing API for mortgage documents (AWS documentation explicitly scopes API to mortgage documents; candidate's commercial-lender loan files are not thereby fully covered.)Source
Cross-doc & review
Unknown from reviewed sources
Price / native unit
Amazon Textract charges you based on the number of pages and images processed (AWS FAQ describes page/image billing basis, not the Analyze Lending rate; current exact rate requires the linked pricing page.)Source
Security & LOS/API
Textract is HIPAA eligible, and compliant with PCI, ISO, and SOC (AWS FAQ vendor-stated service-level assurance; specific report scope and lender applicability not verified.)SourceStartLendingAnalysis operates on a document file located in an Amazon S3 bucket (AWS API documentation specifies S3 input for this operation; integration suitability beyond documented interface not assessed.)Source
Pretrained Bank statement, Pay slip, W2, Lending splitter/classifier + Form Parser & Custom extractor (Google Cloud pricing documentation lists processors. Built-in commercial-loan cross-file checks are unknown from this source; ask whether the intended checks need a separate workflow.)Source
Cross-doc & review
Unknown from reviewed sources
Price / native unit
Bank statement $0.75/classified doc; W2/pay slip $0.30/classified doc; lending splitter $0.05/classified doc; Custom/Form $30/1k pages through 1M, $20/1k above; custom hosting $0.05/hour; OCR $1.50/1k pages (first 1k free; $0.60/1k above 5M) (Specialized parsers count 1 count = up to 10 pages (Invoice/ID) or per classified document; custom processor hosting adds $0.05/hr; 1-yr/3-yr Flexible Savings Plans apply to select general processors.)Source
v4.0 GA prebuilt Bank statement, US tax (W-2/1098/1099/1040), US mortgage (1003-1008), Contract, Layout, Custom models (Microsoft Learn API v4.0 overview lists models. Built-in commercial-loan cross-file reconciliation is unknown from this source; ask whether the intended checks need a separate workflow.)Source
Cross-doc & review
Unknown from reviewed sources
Price / native unit
F0: 500 pages/month, first 2 pages/document, max 4 MB; Microsoft Q&A approximate guidance: ~$1.50/1k Read, ~$10/1k Layout/Prebuilt, ~$30/1k Custom. Region, currency and agreement vary; confirm calculator rates. (Rates from Microsoft Q&A official pricing guidance (azure.microsoft.com pricing page timed out); verify regional currency and agreement rates in Azure Pricing Calculator.)Source
Extracts PDFs, invoices, receipts, and any document into Excel/CSV (24h free reset) (Vendor pricing page; no lending-specific cross-document reconciliation workflow claimed here.)Source
Cross-doc & review
Unknown from reviewed sources
Price / native unit
Free 50 pages (no expiry); Standard $29/mo for 100 pages/mo, 1 seat; Scale $7,000/yr upfront for 42,000 pages/yr, up to 10 users; Enterprise from $35k/yr on card vs $30k/yr in FAQ (Simultaneous page discrepancy retained: Enterprise card shows $35,000/yr while FAQ states $30,000/yr. Scale credits billed annually upfront.)Source
Security & LOS/API
SOC 2 Type II, HIPAA (BAA on Enterprise), no training on customer data (Vendor compliance claim; custom BAA requires Enterprise tier.)Source
150+ pre-trained Skills covering bank statements, tax forms, lending documents, handwriting, checkboxes (Vendor product overview; public pricing is not published and cross-document lending reconciliation is not detailed here.)Source
Cross-doc & review
Unknown from reviewed sources
Price / native unit
Unknown from reviewed sources
Security & LOS/API
SOC 2-certified ABBYY Cloud (US/EU/AU) or on-prem / Azure private cloud (Docker/K8s) (Vendor deployment statement; SOC 2 report scope and enterprise quote remain unverified.)SourceVantage REST API + RPA/BPM connectors (Power Automate, UiPath, Blue Prism, Automation Anywhere) (Vendor product overview; native LOS connectors not listed on this page.)Source
FedRAMP High (Palantir), on-prem, air-gapped, private tenant (AWS/GCP/Azure), or SaaS (Vendor deployment and compliance claim; check whether FedRAMP High applies to commercial SaaS vs Palantir deployment.)Source
Confidence routing with attached evidence, cross-field business rules, and page/row audit trail (Vendor homepage overview (includes MortgageCheckai); pricing is not published and lender accuracy is untested.)Source
Price / native unit
Unknown from reviewed sources
Security & LOS/API
SOC 2 Type II, RBAC, complete audit trail, human review controls (Vendor homepage claim; report scope and deployment options require vendor verification.)Source
Emails, documents, images, spreadsheets and scans in unstructured packets; commercial-loan fit untested. (Vendor platform overview; public pricing, native LOS connectors, and commercial loan testing not established here.)Source
Cross-doc & review
Packet-aware cross-document validations and multi-step business logic across unstructured packets (Vendor platform overview; public pricing, native LOS connectors, and commercial loan testing not established here.)Source
Price / native unit
Unknown from reviewed sources
Security & LOS/API
Customer-controlled storage/processing (own cloud storage), 24/7 monitoring, committed SLAs (Vendor platform overview; specific SOC 2/ISO/HIPAA certifications not detailed on homepage.)Source
Starter $44/mo billed annually ($529/yr, 6k credits, 1 seat); Pro $1,393/yr, unlimited seats; credits expire at billing-period end; 1 page = 1 credit (1.5 with confidence); +20% overage (Unused credits expire at end of billing period; confidence scores add 0.5 credit/page; OCR API focused rather than full lending LOS workflow.)Source
Security & LOS/API
Unknown from reviewed sources
Blend
Status
Digital origination platform; no primary IDP documentation reviewed in this pass.
Will it work with your old loan system when things get busy?
The vendors list documents, review features and connections to loan systems. We haven’t tested those connections or measured throughput, errors or reviewer time. We also haven’t tested accuracy on commercial loan files.