The Legal Front Door: Multi-Agent AI with Human-in-the-Loop Matter Creation

Inside the client's legal organization, employees needing legal help had no single place to go and no consistent path from question to resolution. Taller built an AI-powered “front door” that answered routine questions instantly and, when needed, created and routed a legal matter within the same conversation.

Before this platform, an employee needing legal help had to navigate scattered SharePoint sites, wikis, and email aliases, then guess which legal team owned the issue. Lawyers manually triaged each request, interpreted intent, classified the legal domain, and created and assigned matters by hand across disconnected tools. There was no single front door, no AI assistance, and no consistent path from “I have a question” to “a matter is created and routed.” The process could not scale across employment and immigration, contracts and licensing, marketing, and regulatory compliance. Static knowledge bases, FAQ pages, distribution lists, and ad-hoc ticket queues had all been tried; each misrouted request added hours to days of delay across thousands of monthly interactions.

Taller built a full-stack legal-services portal with a multi-agent Copilot on Azure AI Foundry (a managed cloud platform for building AI agents). A front-door agent mapped each question to a known legal domain; the domain agent then answered using retrieval-augmented generation grounded exclusively in approved SharePoint documentation, so the assistant drew only on vetted sources and could not invent guidance. When human follow-up was needed, deterministic escalation detection handled the handoff, replacing an earlier approach that relied on an AI model to route and proved unreliable. A drafter agent then collected the required fields one at a time through the platform's built-in human-in-the-loop mechanism, created the matter in Dynamics 365, and returned the matter ID in the same chat session. The platform paired a .NET 10 API and a React 18 front end with Entra ID authentication, and ran as region-isolated infrastructure defined entirely as version-controlled code in Bicep and the client's internal deployment tooling, with private networking and least-privilege managed identities. Taller delivered it end to end with a tech lead and seven full-stack engineers.