Document intelligence
Invoices, contracts, delivery notes, forms and reports arrive by the thousand. We build AI that reads them, extracts the data, checks it against your systems and hands a person anything it is unsure of.
What you get
- An extraction schema agreed with your experts, field by field
- An evaluation set of your own documents, with accuracy measured per field
- Every field checked against your ERP and master data
- A review queue where a person decides every uncertain case
- An audit trail per document: the quote and page behind each value
- Cost per document, measured and within budget
Who it is for
Your teams key in data from documents every day: supplier invoices, delivery notes, contracts, claims, certificates, application forms. The volume grows faster than the team, and errors surface later, in a payment, an audit or a customer complaint. Off-the-shelf capture tools read the easy layouts and leave the rest to your staff.
What we build
- Extraction that can show its source. Every value carries the quote and page it came from, and a value the system cannot quote is discarded.
- Checks against your systems. Extracted fields are matched to orders, goods receipts, contracts and master data, so a mismatch is caught before it is posted.
- A review queue for people. Uncertain documents arrive with the open fields highlighted and the source beside them, and a person decides.
- The pipeline around it. Intake from email, upload or scanner, classification by document type, connectors to your ERP or document system, and monitoring of accuracy and cost per document.
The simulated invoice agent on our homepage shows the pattern: it matches each invoice to its order and goods receipt, and hands a person the one whose IBAN differs from the supplier record.
How an engagement runs
| Stage | What happens | Duration |
|---|---|---|
| 01 · Ground the case | Which document types to start with, the fields that matter, the metric, and a verdict on samples | 3-5 weeks |
| 02 · Prove the value | An evaluation set of your own documents frozen first; the prototype measured against it | 3-6 weeks |
| 03 · Engineer the system | Intake, extraction, checks, review queue and connectors for the pilot’s document types | 8-16 weeks |
| 04 · Pilot in one place | One team works with the system on live documents; every correction logged | 6-12 weeks |
| 05 · Scale what works | More document types, teams or sites, with the evaluation set extended | 3-6 months |
| 06 · Run and evolve | Operation; every model change reruns the set; quarterly review of accuracy and cost per document | Ongoing |
Each stage is contracted separately, and the price is quoted on request once its scope is fixed.
Built for Switzerland
Documents often carry personal data, so the hosting option is settled in stage 01 and an impact assessment is scoped where the risk is high. Names and identifiers are replaced with aliases before a model call wherever the design allows, and an independent check fails the run if one leaks. Retention follows your own rules: we keep the audit trail you need and nothing beyond it. The rules, with dates and sources, are on the regulatory reference.
Proof from the lab
| Build | Label | What it shows |
|---|---|---|
| A tiered document intelligence pipeline | Working prototype | Contract intelligence that sends clause extraction and redlining to a stronger model and summaries to a cheaper one, with budget caps per tenant |
| Tenderix | Live | Quote-or-discard extraction from tender documents in German, French and Italian, with OCR for scans and a local model so the text never leaves the machine |
| Agent orchestration for the back office | Working prototype | Accounts-payable invoices processed by role-bounded agents, with duplicate detection, fraud signals and escalation to a person by confidence |
| Anonymise before inference | Internal build | Personal data replaced with aliases before any model call, with an independent check that fails the run if a real value leaks |
Built on
Questions
Which documents can it read?
PDFs, scans, photos and emails. Languages, tables, stamps and handwriting are tested on your own samples in stage 02 before we promise anything about them.
What happens when the AI is not sure?
A value without a quote from the document is discarded, and any document below the confidence threshold you set goes to a person with the open fields highlighted.
Can it post to our ERP on its own?
Only what you allow. Reading and preparing come first. Posting is enabled one document type at a time, and payments, bank details and master data changes always pass a human approval.
Where are our documents processed?
Where you decide in stage 01: Swiss or EU cloud, your own infrastructure, on premises, or a local model on your own hardware. Personal data is replaced with aliases before a model call wherever the design allows.
How do you know it is accurate?
An evaluation set built from your own documents, with the correct value for each field agreed with your experts before the build. Every model change reruns it, and every correction a person makes in operation is added to it.
Next step
Drowning in documents?
Start with the AI Opportunity Assessment: in 1-2 weeks we look at your document flow, the systems it feeds and a first sample, and tell you which documents are worth automating first. The price is on request.
See the AI Opportunity Assessment