Inbound PDF to structured record in forty-seven seconds.
A PDF lands in a shared inbox. Strkr OCRs the pages, classifies the document, extracts the fields, matches to the right record, and routes to the owning team. Humans review exceptions, not every page. The audit log captures every extraction and every routing decision.
Document processing is a four-stage pipeline: ingest, OCR, extract, route. Each stage carries its own quality metrics and exception handling. The pipeline below is the implementation that lands inbound documents on the right record with the right fields populated without a human touching every page.
Ingest
Email inbox, API upload, drag-drop.
Documents arrive via three channels: a configured email inbox (vendor-invoices@yourtenant.strkr.io), direct API upload from a connected system, or a drag-drop on the Strkr record detail page. The ingest layer accepts PDF, scanned image, Word doc, and common attachments. Multi-page documents stay together through the pipeline.
Classify
What kind of document is this.
Strkr AI classifies each document into a type: invoice, PO, contract, DPA, insurance certificate, lease, receipt, report, correspondence. Classification uses the document shape, keywords, and header patterns. Classification confidence displays on each document; low-confidence lands in a human-review queue instead of guessing wrong.
OCR
Readable text from scans and images.
The OCR layer handles native-text PDFs (direct extraction) and scanned-image PDFs (vision-based OCR). Per-page confidence scores surface on the document detail. A low-confidence page triggers either a human review step or a re-scan request. Multi-language support covers 18 languages at ship.
Extract
Fields with confidence scores.
For each classified document type, a Strkr AI extraction template pulls the named fields: invoice number, vendor, line items, amount, due date for an invoice; signer, dates, term, amount for a contract. Each extracted value carries a confidence score and the source page/location. Low-confidence fields flag for review instead of writing silently.
Match
To the correct record on first attempt.
The matcher identifies the correct existing record (account, deal, contact, vendor) using the extracted fields: email domain, company name, invoice number, PO number. High-confidence matches auto-link; low-confidence matches surface a disambiguation step. The matcher never creates a duplicate account by accident.
Route
To the owning team with context.
The routed document lands on the owning team with the extracted fields, the matched record, and the full page images. Billing invoices land on the finance team; vendor contracts land on legal; customer POs land on sales ops. The routing uses the standard Strkr Flow engine, so rules are configurable and visible.
Audit log
Every extraction, every decision.
The audit log captures the full document journey: ingest timestamp, classification result with confidence, OCR per-page confidence, every extracted field with confidence and location, the match decision, the routing destination, every human review action. The audit trail satisfies SOC 2 Type II (in progress 2026-Q4) and HIPAA access logging for teams on the BAA.
Exception queue
Low-confidence lands with a human.
A document that fails classification confidence, matches multiple candidates, or has low-confidence extracted fields lands in the exception queue. A human reviews, confirms or corrects, and the system learns from the correction. Exceptions are the rare path, not the default; teams typically see 7-12 percent exception rate at steady state.
What documents it handles
Six classes most mid-market teams ingest.
Not every document class is equally automatable. The six classes below cover roughly 85 percent of inbound document volume at mid-market B2B companies. Each ships with a pre-built extraction template; teams import and tune the field mappings for their accounting or CRM system.
Invoices
Vendor invoices to AP.
Invoice processing is the most common and most automatable class. Extraction covers invoice number, vendor, bill-to, line items with quantities and amounts, subtotal, tax, total, due date, PO reference. Routing goes to AP for approval and payment scheduling. Three customers reported invoice-processing cycle time down from 11 days average to 3 days after deployment.
Purchase orders
Inbound POs from customers.
POs from customer procurement teams extract PO number, buyer info, line items (SKU, quantity, price, delivery date), total amount, special terms. Routing goes to sales ops for order fulfillment. The extraction catches the common failure mode where a PO sits unprocessed because the sales ops lead did not see the email buried in their inbox.
Signed contracts
Counter-signed PDFs from partners.
Contracts counter-signed outside the Strkr document motion (returned via email, from legacy partners) extract signer, effective date, term, amount, key clauses (termination, auto-renew, SLA tier). Matching links to the right deal record; routing files on the account. Prevents the common failure where a counter-signed MSA sits in a shared drive untracked.
Form processing handles patient intake forms (healthcare), vendor security questionnaires, employment applications, warranty registration forms. Extraction uses the form's structured field positions. Classification distinguishes form types by title and layout. Routing goes to the owning team (clinical, procurement, HR, support).
Insurance certificates
COIs from vendors and partners.
Certificates of insurance (COI) processing extracts policy number, carrier, coverage amounts, effective and expiration dates, named insured. Routing goes to risk management. The expiration date drives an automated reminder flow 30 days before expiry. Classic value: no vendor operates with lapsed insurance because the system flags before it lapses.
Reports
Monthly statements and partner reports.
Inbound reports (monthly bank statements, partner commission statements, carrier settlement reports) extract the summary totals and the line items. The extraction populates downstream reconciliation flows. Classic case: a payment processor settlement report lands as a PDF; the extraction pulls the transaction counts and amounts for reconciliation against the Strkr pipeline data.
The extraction accuracy model
How Strkr handles the inevitable misreads.
OCR and extraction are never 100 percent accurate on real-world documents. The engine ships with explicit confidence surfacing and exception handling that treats accuracy as a signal, not an assumption. The approach below is what separates deployable automation from the "it works in the demo" trap.
Confidence surfacing
Every field shows how sure the AI is.
Each extracted field displays its confidence score (0-100). Fields above 95 auto-write to the record; fields between 70-95 flag for a quick human confirm; fields below 70 land in the exception queue. The threshold is tenant-configurable. Classic benefit: a $240,000 invoice amount with 73 percent confidence lands on a human, not silently on the ledger.
Source anchoring
Hover the field, see the page location.
Every extracted field anchors to the page and bounding box on the source document. Hovering the field in the record view pops up the source location. Humans reviewing an extraction confirm the source in seconds; no squinting at the full PDF to find where the invoice total came from.
Learning from corrections
Human fixes train the next extraction.
When a human corrects an extracted value in the exception queue, the correction trains the tenant-specific model. The next invoice from the same vendor extracts the field correctly with higher confidence. The learning is per-tenant, not shared across tenants. Specific scenarios: vendor invoices from a particular SaaS where the invoice total sits in an unusual position.
Multi-page tables
Line items across page breaks.
Invoices and POs frequently have line items that span page breaks. The extraction engine handles multi-page tables by tracking column boundaries and continuation markers. Classic failure in naive OCR: a 14-line-item invoice reports 12 lines because two were cut by the page break. Strkr's extraction handles this correctly out of the box.
Rotation and skew
Scans fed sideways still work.
Scanned documents often arrive rotated, skewed, or at low resolution. The ingest layer auto-rotates, de-skews, and upscales where needed before OCR. The source document shows the original and the processed version side by side, which keeps the audit trail honest about what the engine actually read.
Handwriting fallback
Hand-filled forms escalate to review.
Handwritten text (patient intake forms, hand-filled applications) escalates to human review for the handwritten fields by default. Teams on the Pro tier and above can enable the Strkr AI handwriting extraction layer, which gets the common cases but still flags low-confidence writing. The honest posture: handwriting is harder than print; the system does not pretend otherwise.
What this replaces
The AP clerk, the shared mailbox, the three-tool chain.
Teams processing inbound documents manually run a predictable workflow: an AP clerk opens the shared mailbox, downloads PDFs, keys the fields into the ERP, files the PDF in the shared drive. The tools vary; the operational pain does not. The comparison below sticks to what document automation actually removes.
AP clerk time
Twenty minutes per invoice to one.
Manual AP processing averages 20 minutes per invoice: open email, download PDF, enter header fields in the ERP, enter line items, cross-reference the PO, file the PDF. Strkr automation drops the per-invoice human time to 1 minute (reviewing exceptions and confirming high-value items). The 19 minutes per invoice compound across volume.
Shared mailbox
The invoices@ inbox nobody owns.
The classic shared vendor-invoices mailbox has 2,400 unread messages, 90 percent of which are invoices that need processing and 10 percent are irrelevant noise. The mailbox is a documentation debt that grows faster than the team can process it. Strkr's ingest layer processes every email in near-real-time; the mailbox count stays near zero because documents move downstream automatically.
Rossum, Bill.com, Veryfi
Standalones that need glue to the CRM.
Standalone document processing tools (Rossum for AP, Bill.com for invoice capture, Veryfi for receipts) do the OCR and extraction well but sit beside the CRM. The glue layer (middleware zaps or custom integration) is where the operational pain concentrates. Strkr processes inside the record, so extracted fields land on the right account without a sync step.
Microsoft Power Automate
Form Recognizer plus connectors.
Teams on Microsoft 365 sometimes stitch Power Automate plus Azure Form Recognizer for inbound documents. The approach works for a single-document-type workflow but struggles with multi-type classification and the routing to multiple downstream systems. Strkr covers the classification plus multi-system routing in one surface.
Not Zapier
Middleware does not do OCR well.
Zapier has an OCR step via third-party integrations but the quality is uneven and the per-run cost accumulates fast on document volume. We do not recommend a middleware tier for OCR workflows. Strkr's processing lives inside the record and the per-document cost is subscription-flat.
HIPAA posture
BAA covers PHI extraction and audit.
Healthcare teams processing patient intake forms sign under the BAA with per-tenant encryption on the stored PDFs and the extracted PHI fields. The extraction audit log captures every field extracted from a PHI document. SOC 2 Type II attestation is in progress for 2026-Q4. GDPR and CCPA data-subject requests cover extracted document data on the standard SAR workflow.
The real-world numbers
Twelve teams measured document processing impact.
Twelve mid-market B2B customers measured their inbound document processing before and after Strkr deployment. The numbers below are the medians across the sample. Document processing is one of the highest-ROI automation surfaces because the manual labor it removes is specifically the labor most teams prefer not to do.
Invoice cycle time
Eleven days became three.
Median vendor invoice cycle time (receipt to AP-ready) dropped from 11 days to 3 days. The savings came from skipping the manual data entry step and routing to the owning approver immediately. Days-payable-outstanding (DPO) improved too because vendors got paid on their actual terms instead of a 10-day lag of internal processing.
Extraction accuracy
Ninety-six percent at steady state.
After two weeks of correction-based learning, extraction accuracy on invoice fields settled at 96 percent across the sample. The 4 percent miss rate is almost entirely low-confidence fields that correctly flagged for review, not silent wrong writes. The high-confidence auto-writes maintained near-100 percent accuracy on the fields they covered.
Exception rate
From 42 percent to 9 percent in six weeks.
Week one exception rate (documents requiring human review) averaged 42 percent as the model learned the tenant-specific vendor patterns. By week six, the exception rate had settled at 9 percent, which is the steady-state figure. The exceptions that remain are typically novel vendors or unusual document layouts; the model learns each one once.
Hours reclaimed
Thirty hours a week per AP team.
The AP team of 3-4 people reclaimed 30 hours a week in aggregate across the sample. The time redirected to vendor relationship management, discount capture, and month-end close acceleration instead of manual data entry. Month-end close shortened by 1.5 days on average because the invoice backlog cleared in near-real-time instead of accumulating.
Compliance capture
COI expiry misses went to zero.
Pre-automation, 11 percent of vendor COIs expired before the risk team noticed. Post-automation, the extraction captured the expiration date and the automated reminder triggered 30 days before expiry. COI expiry misses dropped to zero. One customer avoided a vendor-liability incident that would have cost north of $200k in uninsured exposure.
Audit readiness
Quarterly compliance review from days to hours.
The audit trail capturing every extraction and every routing decision cut quarterly compliance review time from 3 days (sampling-based manual audit) to 2 hours (query-based audit against the full trail). Compliance teams reinvest the time into actual risk analysis instead of chasing paper. The audit log is tamper-evident with hash linking across entries.
Watch an invoice process from inbox to AP-ready in forty-seven seconds.
Bring a sample PDF; we will walk through classification, OCR, extraction, matching, and routing on screen.
What file formats and sizes does the ingest layer support?
Native PDF (text-selectable), scanned PDF (image-based), JPG, PNG, TIFF, Word doc, and common email attachments including nested attachments. Individual files up to 100 MB, multi-page documents up to 500 pages. Beyond the limits, the Pro tier supports higher thresholds; the Enterprise tier is effectively unmetered for typical mid-market document volume. Password-protected PDFs flag for the ingesting user to provide the password; the system never guesses passwords.
How does the model handle a new vendor the system has never seen?
A first-time vendor invoice processes with the base extraction model. The confidence scores on fields are typically slightly lower on a new vendor compared to a vendor the model has seen before. Low-confidence fields flag for human review. Each human correction updates the per-tenant model, so by the third or fourth invoice from the same vendor, extraction confidence lands in the auto-write band. The learning is per-vendor and per-tenant; patterns from vendor A at tenant X do not affect vendor B at tenant X.
What happens if we need to extract a field type the templates do not cover?
The admin UI lets you add custom extraction fields to any document type. A custom field defines its name, data type, and typical location (header, body, footer, line-item). The model starts with heuristics and learns from the first few corrections. Classic examples: a vendor-specific PO number format, a custom contract clause (right-of-first-refusal section), a vertical-specific form field (prior auth number on a healthcare form). Custom fields are first-class alongside the pre-built ones.
Can we require two humans to confirm high-value extractions?
Yes. The configuration surface supports dual-review requirements by document type and extraction value threshold. An invoice over $50k can require two reviewers to confirm the extracted total before auto-write. The dual-review workflow uses the standard approval primitives, so the pending reviews surface on the task list of the configured reviewers. The audit log captures both review events and the final confirmed value.
How does this interact with our existing accounting system?
The extracted invoice fields write to Strkr first and then sync to the accounting system via the configured connector (QuickBooks Online, Xero, NetSuite, Sage Intacct, custom). The sync layer uses the native API of the accounting tool, not a middleware zap. The Strkr record remains the source of truth for the extracted data; the accounting system receives a clean structured payload instead of a manual data entry. For teams that want to run the accounting system as the source of truth, the architecture is bidirectional and configurable.
What about documents that contain mixed content, like an invoice plus a packing slip in one PDF?
The engine supports multi-document PDFs. Classification runs page by page and groups contiguous pages of the same document type together. A 10-page PDF with pages 1-4 being an invoice and pages 5-10 being a packing slip extracts as two separate documents linked by the source PDF. Each document routes independently to its owning team. The source PDF remains available on both routed documents for context. The pattern is common in vendor packets and is handled without manual splitting.
No sales call, no migration consultant, no four-month implementation. Enter your card, get 14 days of the full Pro tier, cancel any time before day 14 with zero charge. Spin up a workspace, import your CSV, and have something useful before lunch.
We use cookies.
Essential cookies keep the site working. If you accept, we also enable Google Analytics
so we can see which pages help and which don't. Reject to opt out entirely. Details in our
Privacy Policy.
