Agents

Document processing · Computer vision

Document extraction agent

Reads invoices, packing lists, B/Ls and certificates and turns every file into verified data. What it cannot be sure of, it does not guess: it shows it to a person with the evidence alongside.

  • Three blind reads per field, with numbers decided in code
  • Cross-checks between the documents in a file and procedure controls
  • Human review of doubtful cases, with the evidence alongside

PDF · A4 · one page

Read the engineering article: “Every data point, three times” →
Agent datasheet: Document extraction agent

How it works

  1. 1

    Intake

    Email, PDF, Excel or scan; true file type. Kept as the Source of Truth.

  2. 2

    Image & OCR

    Rotation and centring by vote; OCR anchored to the digital layer.

  3. 3

    Classification

    Splits the documents in each PDF; page numbering rules.

  4. 4

    Extraction

    Three blind reads; numbers are decided in code.

  5. 5

    Consistency & validation

    Cross-document checks and procedure controls.

What it delivers

Data with its origin

Every data point lands in a database or file with its provenance (digital or OCR): invoice number, lines, total amount.

Human review queue

Doubtful cases are not guessed: they reach a person with the evidence alongside.

Output to your systems

ERP, CRM, other systems or processes: extract, validate, insert.

Never guess silently

Reads

Three blind reads per field.

Checks

Code decides: arithmetic, check digits, deterministic layer.

Every decision is written down

Verified data, or sent to a person for review with the evidence.

Success story · customs freight forwarder

800–1,500

emails a day in the general inbox

6–8

documents per file

94

automated checks per procedure

3.5–9.3 h

of review saved per day

Triage, classification and routing of every email to its department. Documents: invoice, packing list, B/L, health, catch and SOIVRE certificates. Fisheries, SOIVRE and TRACES controls, with facts already verified in code. Saving: 15–40 files a day × 20 min average × 70%.

Where a person decides

Reviews doubtful cases with the evidence alongside (HITL). The agent reads, the code checks and a person decides whatever cannot be assured.

Other agents

Agent datasheet: Office assistant agent

Conversational interfaces · Enterprise AI Assistant

Office assistant agent

Agent datasheet: Phone service agent

Multi-stakeholder service · Voice agents

Phone service agent

Agent datasheet: Accounting agent

Hybrid apps · Vibe accounting

Accounting agent

Agent datasheet: Process automation agent

Process automation · Hyperautomation

Process automation agent