P
Pagentic

You need more than just Document Intelligence. you need Document Understanding.

The reasoning layer for document AI.

Build specialized extraction agents perfectly tuned into each of your messy real-world PDFs — freight invoices, medical EOBs, policy appendixes — and call them through one signed, multi-tenant API.

Trusted by teams extracting millions of pages a month

Invoice factoring
We're processing thousands of invoices through Pagentic every month. What used to take our team days each week to review and key in by hand now takes minutes — and the accuracy is so good our reconciliation errors have essentially gone to zero.
Operations team, Sunrise Hill
1,000s
invoices / month
days → min
review time
~0
reconciliation errors

Sample document to production API in under a day.

Step 1

Upload sample docs

The system surveys them, drafts an agent (instructions + JSON schema + example), and runs it.

Step 2

Iterate interactively

Review extractions, give feedback, chat with the agent — the next iteration responds. Compare versions, roll back any time.

Step 3

Publish + integrate

One click promotes an iteration to a versioned runtime agent. You POST a document and get structured JSON back — inline, or as a signed webhook to your specific endpoint.

Supporting documents

Reconcile every extraction against your source of truth.

Upload your rate cards, fee schedules, master agreements, or approved-vendor lists alongside your samples. Every extraction is automatically cross-checked — and discrepancies land directly in the structured output with full provenance, ready for your audit trail.

It's the difference between "the model said the rate was $4.12" and "the rate was $4.12, your contract says $4.10, here's the page in each PDF, and we flagged it."

Compliance & audit

Reproducible to the byte.

Every cross-check records the field, the supporting doc, the rule, and the actual values compared. Pin the agent version and any auditor can re-run the exact same check months later and get the exact same result.

Catches silent errors

Pennies you used to miss.

An invoice with a freight rate one cent off your negotiated card. A claim line item with a CPT code that doesn't exist anymore. A vendor name not on your approved list. These used to need a human pair of eyes — now they surface in `_meta.cross_checks[]` with the exact discrepancy.

Works with what you have

No reformatting, no schema.

Drop in the PDFs you already use as your reference. Rate cards, fee tables, ICD-10/CPT lists, master service agreements, vendor rosters — keep them in their original form. The agent reads them the same way a human would.

Real-world cross-checks our customers run

Freight invoices ↔ negotiated rate cardsMedical claims ↔ CPT/ICD-10 fee schedulesSupplier invoices ↔ purchase ordersContract amendments ↔ master agreementsBill-from names ↔ approved-vendor listsInsurance EOBs ↔ benefit schedulesLoan docs ↔ underwriting guidelinesLease abstracts ↔ rent rolls

Approval queue

A human in the loop, when it has to be there.

Flip a single checkbox on any agent and every extraction lands in a tidy review queue. A reviewer eyeballs the result, clicks any value to correct it inline, and approves. The webhook fires with the corrected payload. Reject is silent — no downstream system ever sees an extraction your team didn't bless.

Built for regulated workflows where "the model said it" isn't enough. Built for nervous customers who want to start with the safety net on, then turn it off once they trust the agent.

Edit in place

Spot a wrong number? Fix it before it ships.

Click any value — in Visual or Tree view — and an inline editor swaps in. Type the correction, hit Enter, approve. The webhook fires with the edited payload, not the model's mistake. Numbers stay numbers, totals stay totals, your downstream code never knows the difference.

The model's original output is preserved on the extraction record for audit, so reviewers and compliance teams can always see what the LLM said versus what shipped — and a edited_by_reviewer flag on the webhook payload tells your downstream systems when an extraction had a human touch.

  • Same review session for both Visual + Tree views — switch any time, edits persist.
  • Per-field "edited" markers + a running edit counter so reviewers don't lose track.
  • One-click reset to discard edits and revert to the model's original payload.
Pending review · ext_8f2c…
2 edits · reset
Invoice #
A-1003-29
Date
2026-05-09
Vendoredited
Acme Corp.
was Acrne Corp.
Total Dueediting
$1,235.00
RejectApprove with edits & deliver
Compliance-ready

A signature on every record.

Each approval stamps the reviewer's user id and timestamp onto the extraction. Each rejection records the reviewer's note. Auditors get a clean paper trail without any extra plumbing on your side.

Try-before-trust

Onboard cautiously, scale confidently.

Turn approval on while a new agent is finding its footing. Watch every extraction for a week, build trust, then flip the toggle off when the team is ready. The same agent, the same API, no migration.

Zero silent failures

Nothing reaches your system unreviewed.

Webhooks are held until approval — your downstream code only ever processes records a human signed off on. Rejected extractions stay in your audit trail but never leak into production workflows.

Where teams flip the approval toggle on

Medical claims before payer submissionLarge invoices needing CFO sign-offLegal documents requiring partner reviewFirst two weeks of any new agent rolloutHIPAA / SOC2-mandated review workflowsUnderwriting decisions over a dollar thresholdInsurance EOBs with disputed line itemsAny extraction touching PHI or PII

Built for production document AI.

Not a demo. Real pipelines, real provenance, real money on the line.

Verified provenance

Every extracted field carries page numbers and verbatim source text. You can audit any value back to the original PDF.

Sanity checks

Agents declare invariants — totals reconcile, schema valid, every leaf has provenance. Failures surface up; silent garbage doesn't ship.

Signed webhooks + retries

HMAC-SHA256 signatures, exponential backoff (1m → 12h), dead-letter visibility. Or poll if you prefer.

Inspect every extraction

Every call has a full record — tokens used, exact cost, signed webhook deliveries, errors. Audit any value back to the source PDF, replay dead-letters in one click.

Pinned versions

Pin to a specific agent version so iteration in your workshop never silently changes a customer's output.

Idempotency built in

Pass an idempotency key on every POST. Network glitches stop costing you double extractions.

Pay only for what you extract.

Per-page pricing, quoted up front for your specific document type before any production traffic. We're onboarding customers individually right now — no self-signup, no surprise bills.

Production
Per pagebilled monthly

Pay for what you extract. Per-agent rates priced to your document type — quoted up front, no surprises.

  • Live-mode API keys with per-key rate limits
  • Signed webhooks (HMAC-SHA256) + retry queue with dead-letter visibility
  • Verified provenance on every extracted field
  • Per-tenant retention windows you control
  • Email support, 1 business day

Token usage is reported on every extraction; you only pay the per-page meter, not raw tokens. Volume discounts and dedicated regions available — mention them when you reach out.

Ready to talk?

Pagentic is in private preview. Email us about your document type and we'll get you a sandbox key.