Share with your CTO
PONS, a legal AI startup running on Microsoft Azure, has published a detailed breakdown of the three architectural decisions behind its regulated-market platform. The core bet is a hard separation between a continuously updated public legal corpus and isolated per-customer data stores, with neither pipeline touching the other. The platform runs on managed Azure services pinned to Sweden Central for EU data residency, uses queue-based decoupling so the AI Engine scales independently, and backs enterprise sales with SOC 2 Type II and ISO 27001 certifications. It currently serves law firms and in-house legal teams across multiple EU jurisdictions.
What this means for your business
The startups building fastest in regulated verticals right now share a specific trait: they treat compliance architecture as a sales asset, not an audit checkbox. PONS is a clean example of this pattern. Its hard data-separation boundary isn’t primarily a security measure; it’s the answer to the procurement question “will our documents train your model?” Any CTO whose team is pitching AI into legal, financial services, or healthcare buyers faces that same question, and the architecture either answers it or it doesn’t. If yours doesn’t yet, that’s the gap to locate.
The managed-services choice deserves more credit than it usually gets. Writing this off as “just use Azure” misses the actual decision, which is about where a small engineering team places its bets. PONS offloaded scaling, identity, secret rotation, and regional configuration to Azure App Service, Azure SQL, and Azure Key Vault, then directed its scarce engineering hours toward legal-data quality and evaluation pipelines, the parts customers actually differentiate on. That’s not laziness; that’s a correct read of where defensibility lives. The recurring failure mode for early-stage AI startups is the opposite: building custom infrastructure that looks sophisticated internally and is invisible to buyers.
The queue-based decoupling between the backend and the AI Engine, where work is published to Azure Queue Storage rather than called directly, is the detail worth stealing. It means the inference layer scales on its own curve, independent of request volume on the application side. For any team where AI processing is the bottleneck and cost driver, that architectural separation is what keeps cloud spend predictable as usage grows. If your current design calls the model synchronously in the request path, the first spike in usage will show you the problem. PONS built the fix in early; most teams discover it under pressure.
Based on reporting from How to build AI for regulated customers on Azure with PONS, originally published 2026-08-28 16:03:00.
