BEM

BEM is an API-first unstructured-data platform that classifies files and turns documents, messages, and multimodal inputs into schema-constrained JSON.

Verified capabilities and evidence

The fetched homepage showed six named functions: Classify, Extract, Split, Join, Enrich, and Payload Shaping. Its worked example sends a 47-page claims packet to the /v3/functions/extract.run endpoint and displays a 4.1-second response. BEM also advertises editable, type-safe schema contracts, a labeling interface for human review, output correction, and model feedback. The homepage says accounts receive 100 free function calls every month with no card and no trial expiration. Its statement that average customer usage grows fourfold during the first year is a vendor metric, not an independent performance benchmark.

Workflow and differentiation

A production flow can use Classify to route an invoice, bill of lading, email, or claims packet to a typed schema. Extract maps contents to JSON; Split separates a packet into records; Join combines related files; Enrich matches fields against internal datasets using semantic search; and Payload Shaping formats the result for an ERP. That composition is more specific than basic OCR. It is useful when inputs vary but a destination requires a stable contract. Missing values, duplicate files, handwriting, rotated scans, and tables still need explicit tests and review rules.

Pricing and total cost

The homepage confirms 100 free monthly calls and graduated pay-as-you-go pricing. Staged vendor evidence records $0.09 per call after the allowance, with rates of $0.07, $0.05, and $0.04 at higher volumes, plus custom Enterprise terms. The pricing route exposed too little readable detail to map those rates to exact thresholds, so manual verification is required. At the base rate, 10,000 paid calls would cost $900. Buyers should confirm whether pages, packet size, OCR, enrichment, storage, retries, and human review change the bill.

Pros, cons, and evidence gaps

The strongest point is that routing, extraction, packet operations, enrichment, review, and destination formatting live behind one API. Strict schemas and flagged review items are concrete controls for finance and operations. The biggest gap is forecasting: per-call billing may treat a one-page invoice and a large packet differently in practice, and public evidence does not establish every limit. Accuracy also depends on schema design and representative corrections. A polished vendor example cannot substitute for results on the buyer’s own document mix.

Practical evaluation plan

Pilot at least 200 clean PDFs, scans, emails, tables, and edge cases. Track exact-field accuracy, document-level pass rate, review minutes, p95 latency, and cost per accepted record. Require deterministic handling of missing values and duplicates. Compare Azure AI Document Intelligence and Google Document AI, then use the document extraction guide. Before production, test retention, regional processing, audit exports, access controls, and correction behavior with the AI agent security guide.