EUR 482k
Documented pilot
234 invoices · 3,472 lines · 99.1% reconciled · 48 regression tests
← All case studies
AI & agentic workflows In progress
Controlled invoice reconciliation pipeline
The purchase side of an end-to-end supply transformation: controlled invoice extraction with layered stock-determination logic, staged for Business Central behind a sign-off gate — estimated to release roughly 450 hours a year once posting goes live.
391
Staged run
4,119 lines · staged for Business Central behind the sign-off gate
EUR 827k
Reconciled, ready to post
100%
Conservation
funds fully accounted · zero leak · line, header and value parity
≈450 hrs/yr
Est. manual effort released
estimate — ~18 min saved on each of ~1,500 invoices a year
3,799 Item lines
stock and cost items
320 G/L account lines
non-stock postings
From the reconciled Business-Central-ready import; live workbook, 2026-07-06.
Evidence inputs
- Flagship programme, led end to end: a feasibility study modelled across 335 vessels, then purchase-side cost capture — what stock is bought, how it is determined and what it costs — feeding the consumption side that records and evaluates the value created. A three-tier supplier network with high document volume and variable formats.
- Python · LLM tooling · Deterministic validation controls
Transformation
- Combined deterministic parsing and validation with model-assisted extraction for low-confidence residual cases.
- Routed uncertain records to review instead of accepting them silently.
- Cached results by SHA-256 content hash to avoid repeat processing and invalidate changed documents automatically.
Controls & assurance
- Kept every financial figure under deterministic control; the model never posts.
- Reserved model calls for the complex residual to control cost and risk.
Output
- Established full cost visibility across the three-tier supplier network. Once posting goes live, the pipeline is estimated to release roughly 450 hours a year.
Business value
- Anchors the value-creation side of the programme: finance sees what stock enters the operation and at what cost, giving the consumption side a reliable basis for recording and evaluating the value created — while the released capacity returns to analysis.
- Problem
Manual invoice reconciliation consumed several hundred hours a year and obscured part of the supply-chain cost base.
- Approach
Combined deterministic parsing and validation with model-assisted extraction for low-confidence residual cases.
- Outcome
Established full cost visibility across the three-tier supplier network. Once posting goes live, the pipeline is estimated to release roughly 450 hours a year.
Business value
Anchors the value-creation side of the programme: finance sees what stock enters the operation and at what cost, giving the consumption side a reliable basis for recording and evaluating the value created — while the released capacity returns to analysis.
Transformation route
- 01
Combined deterministic parsing and validation with model-assisted extraction for low-confidence residual cases.
- 02
Routed uncertain records to review instead of accepting them silently.
- 03
Cached results by SHA-256 content hash to avoid repeat processing and invalidate changed documents automatically.
- 04
Gated Business Central posting behind automated checks, sandbox rehearsal, two independent calculations and explicit authorisation.
Decision log
- Kept every financial figure under deterministic control; the model never posts.
- Reserved model calls for the complex residual to control cost and risk.
- Treated posting as irreversible and required independent agreement plus explicit authorisation.
This case study proves
- Prompt engineering AI & Automation Strong
- Agentic workflow design AI & Automation Strong
- Cost-aware model routing AI & Automation Applied
- Python Data & Reporting Foundational
What I learned
- Model assistance works best on exceptions; confidence thresholds and deterministic controls govern quality.
Future improvements
- Widen document coverage and calibrate thresholds against a labelled regression set.