Nythrex

Example project · Document AI · Back office

Document AI for a distributor: from inbox to ERP, with a human where it counts.

“Tekton Supply” (a fictional company) distributes building materials in Ukraine. Four operators retype data from supplier invoices, waybills and acceptance acts into the ERP every day. Documents arrive as PDFs, phone photos and scans; formats change with every supplier. Here is how we would automate it without losing control.

By Nythrex EngineeringUpdated 3 min read

Illustrative scenario. This is an illustrative scenario that shows how Nythrex approaches this type of project. It is not a specific client case study: names, figures and timelines are examples.

  1. 01

    Inbox & scans

    Email, upload, mobile photos

  2. 02

    Classify

    Invoice · waybill · act · other

  3. 03

    Extract

    OCR + LLM → structured JSON

  4. 04

    Validate

    Totals · VAT · PO match

  5. 05

    Review exceptions

    Operator confirms or fixes

  6. 06

    Post to ERP

    Via API, with audit trail

Starting point

~6,000

documents per month

200+

suppliers with their own formats

4

operators retyping data

1

ERP (BAS) as the system of record

Example figures for an illustrative scenario.

The pain wasn’t only time. Typos in amounts and item codes caused reconciliation problems at month-end, and documents that arrived as photos from drivers were often skipped until someone chased them. A previous template-based OCR tool broke every time a supplier changed its invoice layout.

Approach

  1. 1

    Sample and label

    We would take a few hundred real documents across the top suppliers and document types, including bad photos, and label the fields that matter: supplier tax ID, document number and date, line items, quantities, prices, VAT, totals.

  2. 2

    Extraction without templates

    OCR for scans and photos, then an LLM that returns strict JSON against a schema. No per-supplier templates — new layouts work on day one, and the evaluation set tells us how well.

  3. 3

    Validation with rules, not trust

    Line totals must add up, VAT must match rates, the supplier must exist, and the document must match an open purchase order within tolerances. Anything that fails goes to review.

  4. 4

    A review screen built for speed

    Operators see the document and the extracted fields side by side, with the uncertain ones highlighted and the source region outlined. Confirm with one key; fix a field inline.

  5. 5

    ERP integration

    Validated documents are posted to BAS through its integration interface, with the original file attached and an audit record of who confirmed what.

Architecture

Inputs

  • Shared mailbox
  • Web upload
  • Driver mobile photos
  • Supplier EDI where available

Processing

  • Queue
  • Classification
  • OCR
  • LLM extraction (JSON schema)
  • Confidence scoring

Validation

  • Arithmetic & VAT
  • Supplier registry
  • PO matching
  • Duplicate detection

People

  • Exception review UI
  • Keyboard-first
  • Source highlighting

Systems (client-owned)

  • BAS ERP
  • Document archive
  • Audit log
  • Monitoring

What we would measure

MetricWhy it matters
Field-level accuracy per document typeShows where extraction is reliable and where review must stay
Straight-through rateShare of documents posted without human edits — the real efficiency number
Review time per exceptionThe review screen’s job is to make the human part fast
Errors reaching the ERPMust go down, not up — the safety metric
Cost per documentOCR and model usage at your volume, tracked monthly

Discovery & labelling

Document sample · field schema

3

Extraction PoC

Accuracy by type

3

Validation & review UI

Rules · operator screen

5

ERP integration

Posting · audit

4

Parallel run

Old and new side by side

3
0815 weeks
Illustrative plan. A parallel run before switching over is non-negotiable for financial documents.

Lessons that transfer

  • Photos from phones are a different problem from scans. Budget for image preprocessing and a clear “retake photo” path.
  • Purchase-order matching catches more errors than extraction confidence does. Business rules are your best guardrail.
  • Measure straight-through rate, not accuracy alone. 98% field accuracy can still mean most documents need a touch.
  • Keep the original and the lineage. Auditors will ask where a number came from.

Frequently asked questions

Keep reading

GuideDocument AI: extract data from anythingHow modern document AI works: OCR plus LLM extraction, business-rule validation, human review for exceptions and integration with your ERP or CRM.GuideAI integration with BAS / 1CConnecting AI to BAS and 1C-based ERPs: integration options, document recognition, reconciliation help, natural-language reports and safe write-back.IndustryAI for logisticsPractical AI for logistics and distribution: document processing, shipment status automation, exception handling, customer communication and dispatch support.ServiceAI integrationAdd AI to the systems you run: classification, summaries, drafts, extraction and routing inside your CRM, ERP and helpdesk — securely, in your cloud.Free toolAI automation ROI calculatorCalculate hours freed, net monthly saving and payback period for an AI automation — using your own task volumes, costs and a conservative automation share.

Want a second opinion on your project?

Tell us what you’re building and where you’re stuck. We’ll reply within one business day with the most practical next step — even if that step isn’t us.

Start a project