Intelligent Automation • Cloud
Document Automation Service
An Azure service that classifies and extracts data from inbound documents, routes low-confidence cases to a human, and writes verified results into the business system.
- Client
- Regulated services provider
- Industry
- Regulated services
Note. Portfolio concepts shown here are anonymized and representative of the kinds of intelligent business systems Aqlyst delivers. Client names, screenshots, and measured results are published only with written approval.
Challenge
The situation
High-volume manual data entry
Inbound documents arrived in several formats and were read and keyed by hand, which limited throughput and made quality dependent on individual attention.
No confidence signal
Earlier automation attempts treated every extracted value as equally reliable, so errors reached downstream systems unchallenged.
Failures that went unnoticed
Scheduled processing had no alerting, so a failed run was typically discovered by the business rather than by the team running it.
Approach
How we worked
- Confidence-driven routingExtraction results carry a confidence score. Values above the agreed threshold flow through; anything below is queued for human verification rather than guessed.
- Human review as a designed stepThe review queue is part of the service rather than an exception path, with a simple interface showing the document alongside the extracted values.
- Observability from the first environmentStructured logging, health checks, and alerting were implemented alongside the processing logic, not added after an incident.
- Reproducible infrastructureBicep templates and automated deployment so environments are recreated from source rather than configured by hand.
Solution architecture
How the system fits together
An event-driven pipeline with an explicit human verification step between AI extraction and write-back to the system of record.
Intake
- Azure Storage
- Event trigger
- Format detection
Processing
- Azure Functions
- Azure AI Document Intelligence
- Confidence scoring
Review
- Verification queue
- Side-by-side review
- Approval
Write-back
- Business system API
- Dataverse
- Document linking
Operations
- Application Insights
- Alerting
- Bicep deployment
Delivered
What was built
- Classification and extraction pipelineAzure Functions orchestrating Azure AI Document Intelligence across the supported document types with per-field confidence.
- Human-in-the-loop review queueLow-confidence extractions route to a verification step with the source document and suggested values side by side.
- Verified write-backOnly verified results are written to the business system, with the source document linked for audit.
- Operational monitoringApplication Insights dashboards and alerts covering throughput, failure rate, confidence distribution, and queue depth.
Outcomes
What changed
Figures are published only once the client has approved them. Where a value reads “Illustrative”, the improvement is real but the number has not been cleared for publication.
- Documents requiring manual keying
- Not yet approved for publication
- Extraction confidence threshold
- Agreed with the business before launch
- Processing failure visibility
- Alerted, not discovered downstream
Technology stack
- Azure Functions
- Azure AI Document Intelligence
- Azure Storage
- Power Automate
- Application Insights
Next step
Working on something similar?
Bring the problem as it actually is, including the parts that are messy. That is usually the fastest route to a useful first conversation.
We reply to every message, usually within one business day.
info@aqlyst.ai · (901) 232-2944