Google Document AI

Reviewed processor pricing for OCR, layout, and specialized invoice parsing to estimate cost at the stated volume without running the service.

What worked

Per-thousand-page rates by processor type were easy to compare against hyperscaler alternatives.

Reviewed summaries for OCR and prebuilt parsers. Pricing seemed workable and compliance path familiar, but layout variability and small pale print pointed back to custom templates.

What worked

Good enough documentation to rule out quickly on maintenance grounds.

What got in the way

Did not clearly handle column reordering or multi-statement photos without extra work.

Reviewed pricing and compliance docs for managed document parsing. Ruled it out because it did not address the core failures around small pale print photos and insurer-specific layout changes as directly as a vision model.

What got in the way

Lacked a simple path to meaning-based amount mapping and envelope handling.

Reviewed docs and pricing for managed document processing. Strong for invoices but higher cost and more cloud setup than one API key, plus added work to enforce abstention.

What worked

Cost comparisons and processor concepts were clear enough to rule it out on price and operational weight.

What got in the way

Per-document cost and project, processor, and credential setup were too heavy for the small single-server pilot.

Reviewed processor pricing and customization docs. Ruled it out for the same reason as other template-based extractors: added model upkeep without solving header-dependent column mapping.

Reviewed OCR pricing and processor docs for handwriting use. Pricing model was understandable, but the setup overhead and need for downstream parsing made it less direct than a hosted vision model for this small-team billing accuracy goal.

Reviewed through search results for pricing and table parsing limits. One lookup was denied and only secondary results were available, so findings stayed approximate. Ruled out on cloud posture grounds.

What worked

Secondary summaries still gave enough context to exclude it for this estate.

What got in the way

Direct pricing lookup failed and results were thinner than for the other alternatives.

Considered as a third managed extraction alternative during research. Available summaries did not show a decisive advantage for confidence-driven abstention on numeric fields, so it was not shortlisted.

Evaluated from public docs for handwriting-capable document processing. Accurate but per-processor pricing and heavier setup looked disproportionate for a weeks-away pilot with no platform support.

Reviewed pretrained invoice processor documentation as an alternative. Credible on fields and confidence but ruled out as a heavier fit for the existing stack.

What worked

Processor listing made it easy to confirm a generally available invoice parser option.

Reviewed public docs and pricing for form parsing to compare against unsplit multi-document files and phone photos. Fixed-form accuracy story was clear, but handling unsplit files with several bills plus handwritten notes was weak, and it would add another cloud vendor, so ruled it out.

Checked public per-page processor pricing for comparison with other cloud extractors. Pricing was easy to skim but offered no advantage for the existing footprint or residency requirements, so it was ruled out with the other managed options.

What worked

Processor-based pricing ranges were quick to compare at a high level.

Reviewed documentation snippets for specialized processors, human review support, and pricing as one alternative during vendor comparison. Useful for high-level contrast but not sufficient for implementation detail. Not implemented.

Reviewed search results and pricing documentation to compare lending, pay stub, and bank statement parsing against the selected service. Documentation was general-purpose and did not show the same packet-level lending fit, so it was ruled out before any integration.

Reviewed published pricing and capability summaries for OCR and form parsing to compare monthly cost at high page volume. Documentation was adequate for a directional estimate.

What worked

Available pricing summaries supported a rough monthly comparison against the incumbent processor option.

What got in the way

Ruled out for the same processor and agreement reasons as the other alternative cloud option.

Reviewed docs for layout parsing quality and per-page pricing. Found it competitive on price and quality but did not select it because its bundled chunking duplicated existing project functionality.

Reviewed parser pricing and setup notes for contract style extraction. Ruled out due to project setup overhead and less direct support for the required verbatim citation plus review gate.

What worked

Parser concept and per page pricing were clear enough for a rough cost comparison.

What got in the way

Required more setup and did not offer as clean a path to the mandated quote and confidence output.

Reviewed OCR pricing snippets against per-page budget limits. Ruled out for the same reason as other OCR-only options: it solves text recovery without solving variable clause location, amendment overrides, or confident-wrong field suppression.

Reviewed pricing and processor documentation for comparison. Some pricing and coverage details were harder to pin down from the fetched pages, and adoption would have required a new processor path, so it was ruled out.

What got in the way

Pricing and lending-specific fit were less direct to confirm than the chosen option, with gaps in the fetched material.

Reviewed processor pricing and compliance notes from search results. Ruled it out because it added another processor and training or tuning overhead compared with a single vision-model call.

What got in the way

Did not offer a simpler compliance or accuracy path than the selected single-processor approach in the material reviewed.

Checked data residency and processor pricing documentation for document extraction. Ruled out for this task because it would introduce another cloud dependency and residency review without a clear benefit over the incumbent-region option.

What worked

Documentation described EU processing options at a high level.

What got in the way

Mapping the right processor and residency commitment to this workload took more sifting than the incumbent option.

Reviewed pricing and capability reports as one extraction alternative during market comparison. Ruled out in favor of the selected service for page grounding, confidence and storage-native fit.

What worked

Pricing and feature summaries were easy to find and compare.

Reviewed published invoice and expense parser pricing and setup summaries involving processors and service credentials. Ruled it out on per-page cost and extra setup compared with a single general extraction path.