Plausible hallucinations
LLMs invent numeric fields, dates, and party names that look correct on first read. Without verification and citation back to source, no operator will sign off — and no auditor will accept the answer.
Expertise / Document Intelligence
Documents in. Decisions out.
Contracts, claims, forms, invoices, filings — the work that depends on them gets stuck the moment the document is unstructured. We build extraction systems that turn documents into structured, auditable data your business can act on, with citation back to the source.
Built by senior teams from the University of Edinburgh, Amazon, Stanford & the Alan Turing Institute
The gap
Why document AI breaks in regulated environments.
OCR plus rules can’t handle the messy reality of enterprise documents. Vanilla LLM extraction hallucinates field values that look plausible and pass the demo. Production document intelligence requires layout-aware models, schema-constrained outputs, verification agents, and an audit trail — not just a model and a prompt.
Silent failure on novel layouts
A contract template changes; a vendor swaps invoice formats; a regulator updates a filing schema. The system silently degrades. Without classification, routing, and confidence scoring, you don’t catch it until exceptions pile up.
Lost provenance
When extraction errors surface in audit, you need bounding boxes back to the source page — not a model log, not a confidence score. Most systems can’t show their work, and “the model said so” is not a defense.
Governance blind spots
Model versions, prompt drift, PII handling, data residency, lineage — all critical in regulated industries, all routinely missed by generic extraction setups. The compliance team finds out last.
How we build
Extraction engineered for audit.
Six steps that separate production document intelligence from generic OCR with an LLM bolted on.
- 01 · Classify and route. Document type detection routes each page to the right extraction pipeline. One model can’t do everything — and pretending it can is how silent failures start.
- 02 · Parse with layout awareness. Reducto, LlamaParse, Unstructured, or Azure Content Understanding preserve tables, hierarchy, and reading order. Layout matters as much as text in enterprise documents; lose it and you lose the meaning.
- 03 · Extract with VLMs against schema. GPT-4.1 Vision, Claude Sonnet 4.5, Gemini 2.5 Pro, or on-prem Qwen2.5-VL and IBM Granite 4.0 Vision. Schema-constrained outputs with Pydantic or JSON Schema — the model can’t return a shape your downstream system can’t consume.
- 04 · Verify and self-correct. A verification agent cross-checks extracted fields against the source. Confidence scoring routes uncertain extractions to a human review queue. The model proposes; the verifier disposes.
- 05 · Cite back to source. Every extracted field links to a bounding box on the source page. Audit isn’t an afterthought — it’s the output format.
- 06 · Operate and monitor. Drift detection on layouts, accuracy monitoring on critical fields, PII handling enforced at the pipeline boundary. The system tells you when it’s degrading before the operator notices.
Classified → parsed with layout → extracted to schema → verified against source — with provenance to the bounding box at every field
What we ship
Extraction patterns we ship.
Three production patterns — each engineered for a different shape of document work.
01
High-volume operational extraction
Invoices, claims, forms, KYC packets. Straight-through processing on the well-shaped majority — 75–90% on mature use cases — with human review queues on the long tail.
02
Complex document intelligence
Contracts, regulatory filings, clinical notes, loan packages. Multi-page, multi-section reasoning with citation — where the answer depends on the relationship between clause seven and exhibit three.
03
Document agents
Workflows where extraction is one step in a larger agentic process — read, decide, act, update systems of record. The document becomes input to an agent, not the agent’s only job.
The production layer
What we engineer around extraction.
The extractor is one piece. These are the layers that make it trustworthy in regulated production.
01
Schema, verification & confidence
Constrained outputs, cross-checks against the source, calibrated uncertainty. Confidence scores that route to exception queues — not a single number on a dashboard.
02
Audit trail & provenance
Bounding boxes, citations, versioned models, replay capability. Every field traceable to the page it came from; every model version recoverable for re-extraction.
03
Governance & operations
PII handling, model lifecycle, exception queues, accuracy SLAs. The pipeline boundary enforces residency, redaction, and retention — not an afterthought layer bolted on for the audit.
Where it earns its keep
Where document intelligence earns its keep.
Audit-grade extraction pays for itself fastest where documents gate the work — high volume, high stakes, or high scrutiny, and usually all three at once.
✓
Insurance claims processing
FNOL, medical reports, repair estimates, supporting documentation. Modern IDP cuts processing costs 60–80% and turnaround times 70–90% on well-defined claim types.
✓
KYC, AML & onboarding
ID documents, proofs of address, bank statements, beneficial-owner declarations. BFSI is 32.7% of the IDP market in 2026 — the maturity is real, the integration work isn’t.
✓
Contract review & obligation extraction
Legal, procurement, vendor management. Multi-clause reasoning with citation to the controlling language — for renewal triggers, indemnity exposure, and obligation tracking.
✓
Accounts payable & invoice processing
Line-item matching to POs and receipts. Straight-through processing rates of 75–90% on mature use cases — the long tail goes to a human queue, with the model’s work attached.
✓
Regulatory filings ingestion
10-Ks, drug submissions, ESG disclosures, prudential returns. Audit-grade extraction with provenance — because the regulator will ask.
✓
Loan & mortgage document packages
Multi-document, multi-format, multi-page. Coordinated extraction with cross-document verification — the appraisal must match the title must match the application.
FAQ
Frequently asked.
Is OCR dead in 2026?
As a system, yes. As a step, no. OCR is now preprocessing inside a layout-aware parsing stage — the extraction engine is a multimodal vision-language model. Pure template OCR plus rules is legacy; running it as a production extraction system in 2026 is a maintenance burden, not a strategy.
When do we use a VLM versus a specialist extraction model?
Usually a VLM, with specialists where they earn their place. Frontier VLMs (GPT-4.1 Vision, Claude Sonnet 4.5, Gemini 2.5 Pro) handle most enterprise documents better than yesterday’s specialists. Specialist models still win on high-volume, narrow tasks where latency and unit cost matter — invoices at scale, ID document verification, signature detection. We test before we recommend.
How accurate is “production accurate” — and how do we measure it?
It depends on the field, the document, and the cost of being wrong. We measure per-field accuracy, per-document straight-through processing rate, and per-pipeline exception rate — against your operational baseline. A 99% field accuracy on a critical invoice field is not the same product as a 95% accuracy on a contract clause, and we engineer to the SLA you actually need.
How do we handle PII and data sovereignty?
At the pipeline boundary, not in the prompt. PII detection, redaction, and residency are enforced before the document reaches the model — and policy is versioned alongside the extractor. For EU-only, sector-regulated, or air-gapped workloads, we deploy on-prem with Qwen2.5-VL, IBM Granite 4.0 Vision, or equivalent open-weight VLMs.
Do we own the document intelligence system you build with us?
Yes. The code, the schemas, the prompts, the verification agents, the eval harness, the integrations — all yours. We design every engagement so your team can operate, extend, and retrain independently. Enablement is built into the project, not sold as an upsell.
Do you build with Reducto, Azure Document Intelligence, Hyperscience, or open-source?
Whichever earns its keep on your documents and your constraints. Reducto and LlamaParse for layout-aware parsing where accuracy is paramount; Azure Content Understanding and Google Document AI where the cloud commitment is already made; Hyperscience and Rossum where managed accuracy on narrow tasks matters; open-source where sovereignty wins. We benchmark on your data before we commit.
Where to start
Where to start.
Bring us a document workflow that should be automated.
Claims, invoices, contracts, KYC, regulatory filings, loan packages. We audit your current process, sample your documents against modern extraction approaches, and deliver a target architecture with accuracy estimates, integration plan, governance mapping, and ROI sizing.
What you get: a production-readiness assessment scored against twelve criteria; a benchmark of two to three modern extraction approaches on a sample of your documents; a target architecture for your environment — cloud, on-prem, or sovereign; a staged delivery plan with timelines, effort estimates, and accuracy targets; and one workshop with your operations and engineering leads. Led by a senior consultant — fixed scope, fixed fee.
Book a Document Intelligence Review →Start the conversation
Ready to turn documents into decisions your team can defend?
A 30-minute conversation with a senior consultant. Bring a document workflow you’ve been trying to automate, or an extraction pilot that won’t pass audit. We’ll tell you whether it’s ready to ship, what the gaps are, and what a Document Intelligence Review would surface.
Book a Document Intelligence Review →