Why Enterprise AI Agents Get Stuck Without Unified Context
AI agents at big companies usually don’t fail because the language model is weak. They fail because the company’s information is spread out across lots of separate systems. Customer records are in CRM databases, money balances are in ERP platforms, and project decisions are buried in emails and chats. At Microsoft Ignite 2026 in November, Microsoft launched an expanded Microsoft IQ. It brings together Work IQ, Fabric IQ, Foundry IQ and Web IQ, plus more than 350 third-party connectors, into one intelligence layer for companies.
Teams don’t have to build a separate retrieval pipeline for every bot anymore. Microsoft puts permissions, structured records and outside knowledge under one managed setup.
How the expanded Microsoft IQ works
According to Microsoft’s Azure Blog announcement, the unified layer combines four specialized engines with five main platform features.
Business Applications in Work IQ and single-call retrieval
Business Applications in Work IQ is now generally available. It adds CRM and ERP records to the data agents can use, along with Microsoft 365 collaboration history. It works with Retrieve, which is also generally available. Together they let agents get Microsoft 365 content and connector data in just one call, and the results only include what the user is allowed to see. Developers don’t have to chain lots of shaky queries across different APIs.
Detailed MCP governance rules
To stop agents from seeing too much data, Microsoft made detailed Model Context Protocol (MCP) policies generally available. Instead of one on/off switch for the whole tenant, admins can control agent access by specific data path, user group and trusted app.
Plugins across ecosystems and Foundry IQ Serverless
Plugins for OpenAI and Anthropic are now generally available, and they bring Work IQ data into the ChatGPT and Claude ecosystems. Foundry IQ Serverless, which has usage-based pricing, becomes generally available in December 2026. Since it’s serverless, companies don’t need to run their own infrastructure for agents that only work now and then or in sudden bursts.
Our take: what to think about before using it
In our engineering view, putting the four IQ engines together fixes a common problem, which is keeping access rules in sync across separate vector stores and operational databases. Opening Work IQ to ChatGPT and Claude also shows that companies use many models, not just one assistant.
Still, architects should think about two things before choosing Microsoft IQ as their standard.
Permission cleanup across 350+ connectors
Single-call retrieval follows the identity rules that already exist in Microsoft 365 and outside connectors. But old oversharing in SharePoint or ERP roles that are too wide can still show private records unless admins check the permissions first.
Predictable costs
Foundry IQ Serverless makes it easier to get started when it arrives in December 2026. But with usage-based pricing, companies need strict query limits so agents running in fast loops don’t cause surprise bills.
What to expect for adoption
By combining federated retrieval in one call with path-level MCP controls, Microsoft turns company context into a managed service. For organizations running agents across many models, whether the workflows stay accurate and secure will depend on unified data and strict access limits.