A practical guide to integrating Azure OpenAI Service with Microsoft Power Platform, covering Power Automate custom connectors, AI Builder prompts, Copilot Studio knowledge, and the data governance requirements for UK businesses.
Azure OpenAI Service (Microsoft's enterprise deployment of GPT-4o and other foundation models within Azure's security and compliance boundary) integrates directly with Power Platform. For UK businesses that need the generative AI capabilities of large language models without routing sensitive business data through consumer AI endpoints, this integration is the pattern that reconciles capability with compliance.
Why Azure OpenAI Specifically (Not Just OpenAI)
The distinction matters for UK businesses, particularly those in regulated sectors. Azure OpenAI Service deploys the same models as OpenAI's API but within Microsoft's Azure infrastructure, meaning data stays within your chosen Azure region (UK South and UK West are available), the service is covered by Microsoft's enterprise security SLAs, and there is no model training on your data.
For FCA-regulated financial services firms, NHS organisations, and law firms with client confidentiality obligations, this boundary is not a preference. It is a compliance requirement.
The Azure OpenAI Data Residency Commitment: Azure OpenAI Service in the UK South and UK West regions keeps prompt and completion data within the UK boundary. Microsoft commits in the enterprise agreement that customer data is not used to train or improve the foundation models.
The Four Integration Patterns
Pattern 1: AI Builder Prompt Actions in Power Automate
The simplest integration path, available to makers without code. Create a prompt action in AI Builder, connect it to your Azure OpenAI deployment, and invoke it from Power Automate flows as a standard action. Best for document summarisation, email drafting, classification, and structured data extraction from unstructured text. Complexity: Low.
A UK insurance firm uses this pattern to summarise claim descriptions into standardised fields for Dataverse, reducing manual handling time by 12 minutes per claim.
Pattern 2: Custom Connector to Azure OpenAI API
When you need more control over prompt engineering, model selection, temperature settings, or response format than AI Builder exposes. A custom connector wraps the Azure OpenAI REST API and makes it available as a first-class action in both Power Automate and Power Apps. Complexity: Medium.
A UK law firm uses a custom connector to send contract clauses to GPT-4o with a structured system prompt and function calling to extract defined clause types into Dataverse fields.
Pattern 3: Copilot Studio Generative Answers with Azure OpenAI
When you want a conversational AI agent that answers questions grounded in your own business knowledge, SharePoint documents, Dataverse records, or custom data sources. Copilot Studio's generative answers node can be configured to use your own Azure OpenAI deployment rather than the shared Microsoft service. Complexity: Medium.
A UK professional services firm uses this pattern for an internal knowledge agent that answers staff questions about policies, procedures, and client engagement guidance, grounded in SharePoint documentation.
Pattern 4: Azure Function Middleware with Power Platform
When the AI logic is complex enough to require back-end orchestration, multi-step reasoning, vector database retrieval, model chaining, or custom fine-tuning. An Azure Function acts as the AI orchestration layer, called by Power Automate via HTTP or custom connector. Complexity: High.
A UK bank uses Azure Functions to orchestrate a multi-model pipeline (GPT-4o for reasoning, Ada for embedding search, and a custom classification model) called from a Power Automate flow triggered by new document uploads.
UK Business Use Cases by Sector
Financial Services: AML narrative generation, summarising transaction patterns into structured SAR narratives; credit memo drafting from structured application data; Consumer Duty fair outcomes assessment from product terms documentation.
Legal: Contract clause extraction and classification; due diligence report summarisation; precedent suggestion from matter management system knowledge base.
NHS and Healthcare: Clinical note structuring from dictation transcripts; discharge summary drafting; referral letter triage and urgency classification.
Professional Services: Proposal generation from CRM opportunity data; meeting note summarisation and action extraction; client report drafting from structured project data.
Manufacturing: Quality non-conformance report generation from inspection data; maintenance work order drafting from engineer voice notes; supplier correspondence from structured deviation records.
UK Data Governance Requirements for Azure OpenAI Integration
- GDPR data minimisation: Only send the minimum personal data required for the AI task. Strip personal identifiers before sending to the model unless they are essential to the output.
- Data subject rights: If AI-generated outputs are stored against individual records, document how data subject access requests and erasure requests will handle those AI-generated fields.
- AI transparency obligations: Under UK GDPR, where AI is used to make decisions that significantly affect individuals, you must be able to explain how those decisions were made.
- Sector-specific requirements: FCA operational resilience requirements mandate that AI systems in scope of critical business processes have documented fallback procedures. NHS DSPT requirements cover how AI-processed clinical data is handled and retained.
- DLP policy alignment: Configure Power Platform DLP policies to explicitly permit the Azure OpenAI connector in environments where you are using it, and block it from environments where it should not be accessible.
Prompt Engineering for Business Reliability
- Define the output format explicitly, structured JSON, a specific list format, or a defined number of bullet points. Unstructured natural language output creates parsing problems downstream.
- Include boundary-setting instructions, what the model should say when it cannot answer confidently, rather than generating a plausible-sounding but incorrect response.
- Test against adversarial inputs, what happens when the input document is incomplete, in an unexpected format, or contains contradictory information?
- Version control your prompts as code, store them in your ALM pipeline alongside the Power Automate flows that use them, so prompt changes go through the same review and testing process as code changes.
