What Is Azure OpenAI Service and Why SpeedMVPs Uses It
Azure OpenAI Service is Microsoft's managed deployment of OpenAI's models - GPT-4o, GPT-4 Turbo, GPT-3.5 Turbo, DALL-E 3, Whisper, and the text-embedding-3 series - hosted within Azure's global infrastructure. Unlike the standard OpenAI API, where your data is processed on OpenAI's infrastructure under their terms of service, Azure OpenAI processes data within your chosen Azure region under Microsoft's enterprise data processing agreements. SpeedMVPs reaches for Azure OpenAI in a specific set of situations. First, when a client's IT security policy prohibits sending data to non-Microsoft cloud providers - common in financial services, NHS-adjacent organisations, and large enterprise environments already committed to Microsoft EA agreements. Second, when the project requires UK data residency with a binding contractual commitment: Azure's UK South (London) and UK West (Cardiff) regions provide GPT-4o access within the UK, satisfying ICO guidance on international data transfers post-Brexit. Third, when the client needs Azure's compliance portfolio: Azure OpenAI is covered by ISO 27001, SOC 2 Type II, Cyber Essentials Plus, and NHS Data Security and Protection Toolkit alignment, making procurement sign-off significantly faster. The tradeoff is deployment friction: you apply for model access, wait for approval, deploy models to your own Azure subscription, and manage capacity through Provisioned Throughput Units (PTUs) rather than a simple pay-per-token API. For the right client, that friction is a worthwhile price for the governance it enables.
Setting Up Azure OpenAI Service in a Production AI Project
Azure OpenAI setup involves more steps than the standard OpenAI API, but each step maps to a real governance requirement. Here is the production setup sequence SpeedMVPs follows. First, apply for Azure OpenAI access through the Azure portal. Approval typically takes two to five business days. You will need to describe your use case and confirm acceptable use policy compliance. For regulated industries, be specific about your use case in the application - healthcare and financial services applications have been approved without issues in SpeedMVPs' experience. Second, create an Azure OpenAI resource in your target region (UK South for London data residency). Choose the pricing tier appropriate for your expected throughput. Set up a dedicated resource group for AI services so access control and cost tracking are isolated from other Azure resources. Third, deploy the specific model versions you need within your Azure OpenAI resource. Model availability varies by region - not all GPT-4o versions are available in UK South, so check the regional model availability matrix before committing to a region. Fourth, configure network access. For enterprise deployments, disable public network access and use a Private Endpoint within your VNet. This means all traffic between your application and Azure OpenAI stays on Microsoft's backbone, never traversing the public internet. Set up Private DNS zones correctly or API calls will fail silently. Fifth, create managed identities for your application services rather than using API keys. Managed identities eliminate the risk of key rotation failures and mean secrets never appear in environment variables or configuration files. Assign the "Cognitive Services OpenAI User" role to the managed identity. Sixth, configure Content Filtering policies appropriate for your use case. Azure OpenAI's content filters are on by default and can block legitimate requests in some domains (medical information, legal advice). Apply for filter modifications with documentation of your use case. Finally, set up Azure Monitor and Diagnostic Settings to capture all API call logs. These are required for compliance audit trails under NHS Digital DSP Toolkit, FCA SYSC requirements, and GDPR Article 30 records of processing.
Key Features and Capabilities
Azure OpenAI's most important differentiator for enterprise clients is the data processing boundary. When you use Azure OpenAI, Microsoft contractually commits that your prompts and completions are not used to train OpenAI models, are processed only within your chosen region, and are retained for abuse monitoring for a maximum of 30 days (configurable to zero with an approved zero data retention policy). For NHS Digital-regulated data, FCA-regulated financial data, or any data subject to UK GDPR, these contractual commitments are what make production deployment viable. Provisioned Throughput Units (PTUs) are Azure OpenAI's answer to rate limits. Instead of sharing capacity with other API users and hitting token-per-minute limits, you reserve dedicated compute capacity. This matters for production AI products where inconsistent latency degrades user experience. PTUs are priced by reservation period rather than per token, so the economics shift for high-volume applications. Azure OpenAI supports the same API schema as the standard OpenAI client library, which means switching between the two is a configuration change rather than a code change. The openai Python and JavaScript SDKs both support Azure endpoints with an azure deployment parameter. This makes hybrid deployments practical: use Azure OpenAI for production with data residency requirements, and fall back to standard OpenAI for development environments. The integration with Azure Active Directory (now Entra ID) means you can enforce conditional access policies on AI API usage - restricting which identities, from which network locations, can call the LLM. For financial services firms under FCA SYSC 8 third-party outsourcing requirements, this level of access governance is expected. Azure OpenAI also supports fine-tuning on GPT-3.5 Turbo and some GPT-4 variants within your Azure subscription, with fine-tuned models stored and served entirely within your chosen region.
Real-World Workflow: Azure OpenAI Service in an AI MVP
SpeedMVPs recently delivered an AI contract analysis tool for a UK financial services firm regulated by the FCA. The product needed to extract key terms, flag non-standard clauses, and summarise obligations from supplier contracts. The compliance team's requirements were non-negotiable: no data could leave UK infrastructure, all API calls had to be logged for audit, and the vendor had to hold relevant compliance certifications. Azure OpenAI in UK South met every requirement. The application ran on Azure App Service with a managed identity, communicating with the Azure OpenAI resource through a private endpoint inside the client's existing Azure VNet. No API keys were stored anywhere - managed identity tokens were obtained at runtime and rotated automatically by Azure. The prompt design used GPT-4o's structured output capability to return JSON with extracted clause data, confidence scores, and citation offsets into the original document. Diagnostic logs went to a Log Analytics workspace retained for 90 days, satisfying the FCA's requirement for records of automated decision-making processes under SYSC 10A. Content filtering was configured with elevated thresholds for the legal domain after the default filters flagged some contract language as potentially harmful. Microsoft's filter modification process took three business days with a written use case justification. The entire system - including infrastructure provisioned via Terraform, the Next.js frontend, and the Azure Functions backend - was delivered in 16 working days. The client's IT security team had approved the architecture within the first week because the Microsoft compliance documentation was already familiar to them from their existing Azure estate.
Cost and Pricing Considerations
Azure OpenAI pricing follows the same per-token structure as the standard OpenAI API for pay-as-you-go deployments, but with a slight premium over direct OpenAI pricing in some regions. GPT-4o in UK South currently runs at comparable rates to the standard API, though prices shift and should be verified at the Azure pricing calculator before budgeting. Provisioned Throughput Units change the economics significantly for high-volume applications. PTUs are priced per unit per hour on a reservation basis (hourly, monthly, or annual). A single PTU provides approximately 2,500 tokens per minute of consistent throughput. For applications processing thousands of documents per day, PTUs often cost less than pay-as-you-go at scale, and the predictable capacity eliminates throttling risk. The compliance infrastructure adds cost beyond the model API: Private Endpoints carry a small hourly charge, Log Analytics retention costs per GB ingested, and Azure Monitor alerts add marginal cost. For a typical AI MVP, these infrastructure costs add GBP 50 to 150 per month on top of token costs. For UK organisations with Microsoft EA agreements, Azure credits may offset Azure OpenAI costs during an evaluation period. SpeedMVPs helps clients navigate Azure pricing and EA consumption to find the most cost-effective path to production, and includes a cost model covering 6 and 12-month projections in every project handover package.
Alternatives to Azure OpenAI Service
The standard OpenAI API is the obvious alternative for teams without strict data residency requirements. It offers the same models, simpler setup, no approval process, and access to the latest model releases often before they reach Azure. The tradeoff is governance: your data is processed under OpenAI's standard terms, which may not satisfy enterprise procurement or regulatory requirements. Anthropic Claude API provides long-context reasoning and strong safety properties, with Claude Sonnet available via AWS Bedrock in EU regions for teams needing non-Microsoft hosting with strong compliance frameworks. Claude is particularly well-suited to document analysis tasks where context windows above 200,000 tokens matter. Mistral AI is a French LLM provider with EU-hosted infrastructure and explicit EU AI Act alignment. For organisations where the requirement is EU data residency rather than Microsoft certification, Mistral Large or Mistral Small may satisfy compliance needs at lower cost than Azure OpenAI. For NHS Digital-regulated deployments specifically, NHS-approved cloud providers (AWS UK, Azure UK, GCP UK) all provide the data processing agreements required for NHS DSP Toolkit compliance. The choice between them often comes down to existing organisational relationships rather than technical factors. SpeedMVPs has delivered compliant AI products on all three for NHS-adjacent clients.