Document AI you can start using in a minute and still defend in an audit.
Free with your own model key. Templates, workflows, confidence gates, human review, batch and API. Every extraction is logged: which model ran it, where, and under whose key. When that has to be your own hardware, it can be.
Parsing a document is the easy part. Everything after it is the work.
Most document AI is an API that just hands you a JSON. What you actually need is: a definition of what “correct” looks like, a way to catch the 4% that’s wrong, a person to look at it, a path into your system of record, and a way to explain any of it six months later.
Proof Perimeter does that part. It works the same way whether the model is OpenAI’s, yours, or one running inside your own network.
Set the rules once
Define the fields, the types, the validation rules and the failure conditions once. A passport template knows an MRZ checksum should pass. A bank statement template knows the balances should reconcile. The model isn't guessing what you wanted. You've told it.
A confidence gate in the middle
Classify → extract → validate → route. Set a threshold per field, not per document. High-confidence fields go straight through; anything under the line lands in a review queue with the source region highlighted. Straight-through processing, with a record of why each field went through.
Built into the workflow
A reviewer sees the field, the bounding box, the model's confidence and the provenance record side by side. Corrections are saved. They become evaluation data, and on the enterprise tier, fine-tuning data.
Interactive, batch, or headless
Drag a file in. Drop 40,000 into a batch. Or never open the app at all and POST to the API with a webhook on the other side. Same templates, same workflows, same audit trail in all three.
Export where you need it
JSON, CSV, XLSX, or straight into a database. The schema stays the same across runs, so your downstream code doesn't break when a model version changes.
Bring your own model. It works better here.
You can call OpenAI or Anthropic yourself. Most teams who do end up with a prompt nobody wants to touch, a JSON parser wrapped in three try/catch blocks, no evaluation set, and a bill that grows with the length of the document rather than what’s in it.
Field-scoped prompts, schema-constrained output and per-document-type instruction sets, tuned against a benchmark suite, and measured. Fewer retries, less malformed JSON, higher first-pass field accuracy.
Page routing, region cropping and layout-aware chunking mean a 40-page loan file doesn't become 40 pages of tokens. Cost per document falls because the token count falls.
Structured, high-confidence fields go to a small model; genuinely ambiguous ones escalate. Per-field routing, not per-document.
Repeated layouts hit cache. Batch jobs use the provider's batch pricing. Rate limits are handled once, correctly, instead of once per team.
Run a new model version against your own labelled documents before you trust it. Watch accuracy per field, per document type, over time. Most teams don't build one. Then someone asks how accurate it is.
Every document comes with a record of where it was processed.
Ask where your document AI runs and you’ll get an architecture diagram. Ask where one specific KYC pack was processed, on which model version, under whose key — and usually nobody knows.
Your own key, hosted
The provider, the region, your key. We hold nothing after the run.
Every tier has the audit trail. What you’re paying for is what it says.
Pricing
The full workspace, running on your model provider.
- Unlimited documents, templates and workflows
- Bring your own key: OpenAI, Anthropic, Google, Bedrock, or any compatible endpoint
- Interactive upload, batch and API access
- Human-in-the-loop review and all export formats
- Full provenance log on every document
- Community support
Proof Perimeter's proprietary document processing models — hosted or on-premises.
- Proof Perimeter's proprietary, tuned extraction models — no API key required
- Deploy hosted, in a private cloud, or fully on-premises; air-gappable when required
- Fine-tuning on your corrections and your worst documents
- Custom workflows, legacy core connectors, system-of-record integration
- SSO, RBAC, maker-checker review, admin controls
- Governance dashboard: residency map, exportable conformity pack
- Deployment support and an SLA
Who uses Proof Perimeter
You have a Textract call, a prompt, and a growing folder of edge cases. Start free with your own key.
You automated the easy documents two years ago. The hard ones — the faxes, the bilingual IDs, the handwriting — still go through people. Confidence gates and review queues move them.
You're accountable for decisions made on documents you can't trace. Every extraction here carries a record of the model that produced it.
In the free tier: your key, your provider, your region, nothing retained. On-prem: nothing leaves. Both are in the record.
You can test it yourself: accuracy per field, per document type, on your own worst files, before you commit to anything.
Common questions
Is it really free?+
Yes, with your own LLM API key. The full workspace — templates, workflows, batch, API, review, export, provenance — with no document limit. You pay your model provider directly, at their rates. We don't charge anything. If you'd rather not manage a key, the Enterprise tier runs on Proof Perimeter's own proprietary document processing models — no API key required.
Do my documents pass through your servers?+
On the hosted tiers, yes — that's what makes it a web app. Your files are processed and then dropped; we don't retain document content or extracted values beyond the run, and the provenance record states exactly which provider and region handled the inference. If your regulator, your risk function, or your own judgement says that isn't good enough for a given workload, that's what the enterprise deployment is for: the same product, running inside your network, with no egress at all.
Why is it better than calling the API myself?+
Field-scoped prompts and schema-constrained output tuned against a benchmark set; layout-aware chunking so you don't pay for tokens the model doesn't need; per-field model routing; caching, batching and retry logic you'd otherwise build twice. And a way to test model changes, which almost nobody builds for themselves. The result is fewer tokens per document, fewer retries, and a first-pass accuracy number you can show someone.
Which models can I use?+
OpenAI, Anthropic, Google, Amazon Bedrock, and any OpenAI-compatible endpoint, with your own key, on the Individual tier. On Enterprise, Proof Perimeter's own proprietary extraction models — hosted or running locally on your hardware, optionally fine-tuned on your documents.
Do I need to be technical?+
To upload a document, define a template and export a spreadsheet: no. To wire the API into a pipeline: about an afternoon.
How accurate is it?+
That depends on your documents. Anyone who gives you a single number is guessing. Bring your worst files — the faxes, the handwriting, the bilingual IDs — and see the numbers before you commit to anything.