From incoming PDF to operating system, in minutes
A PDF arrives by email. Someone reads it. Someone copies data into the ERP. Someone files the original. Someone looks for a piece of data 6 months later and can’t find it. Manual document processing is where your admin team’s hours go to die.
We build complete pipelines: from document input (email, scan, upload) to output in your system (ERP, CRM, database), with AI extraction, human validation where needed and automatic filing. Document operations, not isolated extraction.
How we build it
1. Document type audit
Before touching anything, we map which documents actually come into your organisation: invoices, contracts, delivery notes, technical sheets, CVs, records, receipts. For each type: monthly volume, extraction complexity, criticality. We design the pipeline for what you have, not for an ideal case.
2. Multi-channel ingestion
Dedicated email, S3 drop-folder, web form, scan-to-server. We connect the channels your team actually uses to receive documents. No one has to change their habits.
3. Extraction + validation
Each document passes through vision models (Claude Vision, GPT-4 Vision) according to its type. Critical fields are validated automatically: VAT balances, the NIF is valid, the total matches. Those that fail validation are flagged for human review with the specific data highlighted.
4. Integration with your systems
The extracted data goes to the ERP (Holded, Dolibarr, SAP), CRM, database or whichever system you use. The originals are archived in S3 / Drive with metadata for future search. Every processed document leaves a complete trail. Auditable, traceable, retrievable.
Types of pipeline we typically build
- AP / supplier invoices: ingestion → extraction → matching against PO → approval → ERP.
- Legal contracts: extraction of dates, amounts, clauses → record in CRM → expiry alerts.
- Product spec sheets: spec extraction → PIM → automated web.
- Delivery notes vs orders: delivery note ingestion → matching against order → discrepancy flagging.
- Document onboarding: new clients/employees with document pack → extraction → validation workflow.
Difference from Document AI
The Document IA product solves the extraction piece: converting a PDF into structured JSON. Document Processing is the complete pipeline: ingestion, extraction, validation, integration, archiving, audit. If you only need to extract data, the first one is enough. If you need documents to stop being handled manually from end to end, this is the way.
What changes when this works
- Your administrative team reclaims 20-40 hours a week previously lost to manual transcription.
- Transcription errors trend towards zero — cross-validation catches them before they reach the system.
- Audits and compliance become trivial: every decision has a traceable trail.
- Documents received at 23h are processed at 23:01 — operations don’t wait for working hours.
When we do NOT recommend this
- If your volume is low (<30 documents/month): the setup does not justify the cost compared to manual processing.
- If each of your documents is unique (bespoke consultancy): there are no patterns to automate.
- If your target system has no API (legacy desktop): integration is not viable.
Privacy
For regulated sectors (banking, health, legal) we work with on-prem vision models (Qwen2.5-VL 72B, InternVL3) on your infrastructure or ours in Europe. Your documents do not go to US APIs. More expensive to operate, but defensible before an auditor.
Stack we use
- Claude Vision / GPT-4 Vision as the primary extractor.
- n8n self-hosted for pipeline orchestration.
- S3 / Google Drive for archiving with metadata.
- Direct API to the ERP / CRM / target system.
We start with a diagnostic session
Before quoting the full setup, we run a 90-minute session. We review with you which documents come in, what is done with them today, and which bottleneck hurts most. We leave with an honest recommendation: where to start, which pipeline delivers ROI fastest, and what to leave as it is.
We don’t charge for that session. If you’re interested, get in touch.