A practical guide to Microsoft AI Builder, how UK businesses use document intelligence, form processing, and object detection to eliminate manual data entry, cut processing times, and meet compliance requirements.
Manual document processing is one of the highest-cost, highest-error activities in UK business operations. AI Builder (Microsoft's no-code AI capability embedded directly in Power Platform) makes it possible to extract structured data from unstructured documents with production-grade accuracy, without writing a single line of machine learning code.
What Is AI Builder and How Does It Fit into Power Platform?
AI Builder is Microsoft's embedded AI capability for Power Platform, available inside Power Apps and Power Automate without requiring Azure ML expertise or data science resource. It provides pre-built AI models for common business scenarios and a no-code model training interface for custom document types.
Critically, AI Builder is not a separate product with a separate learning curve. It is a native component of the Power Platform environment that passes extracted data directly into Power Automate flows and Power Apps forms. A document arrives, AI Builder extracts the fields, and the data flows directly into Dataverse or a downstream system without manual intervention.
AI Builder vs Azure Document Intelligence AI Builder uses Azure Document Intelligence (formerly Form Recogniser) under the hood but abstracts the technical configuration entirely. For most UK business document processing scenarios, AI Builder is the right tool. Azure Document Intelligence directly is appropriate only when you need API-level control, custom post-processing pipelines, or integration outside the Power Platform ecosystem.
The AI Builder Model Types: Which One Fits Your Use Case
Document Processing (Custom) When you have a consistent document layout, supplier invoices, insurance certificates, application forms, purchase orders. You train the model on 5+ sample documents and it extracts named fields from future documents. Typically 90–97% field extraction accuracy on well-formatted documents with 20+ training samples.
Invoice Processing (Prebuilt) For general supplier invoices where you do not control the format. The prebuilt model handles diverse invoice layouts without training, extracting vendor name, invoice number, date, line items, amounts, and VAT. 85–92% accuracy on standard invoice formats. Works out of the box with no training required.
Receipt Processing (Prebuilt) Expense management, petty cash processing, and supplier receipt handling. Extracts merchant name, date, items, totals, and tax from printed and digital receipts. Well-suited for expense automation across retail, hospitality, travel, and professional services contexts.
Identity Document Processing (Prebuilt) AML and KYC onboarding, right-to-work verification, and employee onboarding where ID document data must be captured into a system. Processes UK passports, driving licences, and international identity documents. Must be used within a compliant data handling framework, GDPR and AML considerations apply.
Text Classification (Custom) When you need to categorise inbound documents, emails, or text by type, routing complaints vs enquiries, classifying legal correspondence by matter type, or triaging IT helpdesk tickets. Effective with 10+ examples per category. Works inside Power Automate to route items automatically.
Object Detection (Custom) Quality inspection on production lines, asset condition monitoring, and inventory counting from photographs. Trained on images of the specific objects you need to detect. Production-ready for manufacturing inspection with 50+ training images per object type.
UK Business Use Cases in Production
- Accounts payable automation: AI Builder extracts invoice data, Power Automate matches against purchase orders in Dataverse, routes exceptions for review, and posts approved invoices to the finance system. UK mid-market businesses are achieving 75–85% straight-through processing rates.
- Legal matter intake: UK law firms use document processing to extract client details, matter type, and key dates from instruction emails and new matter forms, eliminating the 15–20 minutes of manual data entry per new instruction.
- AML client onboarding: Financial services firms extract data from identity documents and proof of address to pre-populate KYC records in their CRM, reducing onboarding time from 45 minutes to 8 minutes per client.
- NHS referral processing: NHS trusts use document processing to extract patient details, referral urgency, and pathway codes from inbound referral letters, eliminating manual transcription and the errors it introduces.
- Property and insurance forms: Property managers and insurance brokers use custom document models to extract data from surveyor reports, valuation certificates, and broker forms into their management systems.
AI Builder Licensing: What It Costs in the UK
AI Builder is priced on a credit consumption model. Each AI Builder service plan provides 1 million credits per month (approximately £400/month for a standalone add-on).
| Operation | Credits per Run | Runs per Month (1M credits) |
|---|---|---|
| Document processing (custom model) | 1 | 1,000,000 |
| Invoice processing (prebuilt) | 1 | 1,000,000 |
| Receipt processing (prebuilt) | 1 | 1,000,000 |
| Identity document processing | 1 | 1,000,000 |
| Text classification | 1 | 1,000,000 |
| Object detection | 1 | 1,000,000 |
At 1,000 invoices per month, a single AI Builder service plan covers the processing at approximately £0.40 per 1,000 documents, significantly cheaper than the fully-loaded cost of manual processing at any UK wage rate.
Implementation: What a Production-Ready AI Builder Deployment Looks Like
- Collect 20+ representative sample documents, the more layout variation, the more robust the model. For invoices, collect samples from your top 10 suppliers.
- Train the custom model in the AI Builder portal, labelling each field you need to extract. The training process takes 15–30 minutes and requires no coding.
- Test extraction accuracy against a held-out set of 10–20 documents not used in training. Target 90%+ field accuracy before proceeding to production.
- Build the Power Automate flow: document arrives via email attachment, SharePoint upload, or API → AI Builder extracts fields → conditional logic handles confidence thresholds → high-confidence extractions auto-process, low-confidence items route to a human review queue.
- Implement a human-in-the-loop review interface in Power Apps for the exception queue. This is essential for production deployments and should not be omitted.
- Monitor extraction accuracy monthly and retrain the model when new document layouts appear in the training set.
