Case study · Finance platform · Generative AI
DataTwin enabled informed decisions through GenAI-based business transaction analysis.
DataTwin · Modern finance platform
Their invoice AI needed five man-months of retraining per 50,000 invoices. We replaced the retraining with a Bedrock workflow.
80%
Reduction in invoice processing time
50%
Faster financial book close
The opportunity
An AI that had to be retrained every time an invoice format changed.
DataTwin automates accounts payable, accounts receivable, fixed assets and reconciliations for finance teams. Its invoice processing used an AI-based system to extract data from supplier invoices-which worked, until the format changed.
Adapting to evolving formats meant manual retraining, and that retraining consumed five man-months for roughly 50,000 invoices. The rigidity was the problem, not the accuracy: DataTwin needed extraction that stayed accurate without the retraining cycle.
Manual retraining required
The model had to be retrained to handle evolving invoice formats.
Five man-months per 50,000 invoices
That was the cost of keeping extraction current.
Invoice formats kept changing
Supplier variation meant the target moved continuously.
The model was rigid
Adapting it required retraining rather than reconfiguration.
The impact
Formats change; the workflow adapts without retraining.
80% faster processing
Vendor details, invoice numbers, dates, line items and amounts extract automatically against predefined checks.
Structured to JSON
Bedrock normalises extracted data into a standard format, so downstream systems read one shape.
Real-time invoice visibility
Status tracking and payment-due monitoring come from the structured store rather than a report run.
Audit trail maintained
A comprehensive record of processing activity supports regulatory requirements and audits.
The stack
What it runs on.
- Model layer
- Amazon Bedrock, prompt tuning
- Orchestration
- LangChain extraction from emails and attachments
- Structure and storage
- Standardised JSON in a queryable database
- Response generation
- LLM-generated email replies for internal agents
The road ahead
Extraction adapts on its own. Next is what it feeds.
With invoices structured on arrival and the audit trail automatic, the procure-to-pay cycle can be optimised on live data rather than reconciled after the fact.
- 01
Procure-to-pay optimisation
- 02
Wider document classes
- 03
Supplier risk signals
Talk to us
Bring one process. Leave with an outcome architecture.
30 minutes on one function, mapped to the architecture behind it. No slides.