Microsoft's Jeff Hollan, who leads the Agent Platform in Microsoft Foundry, is drawing sharp distinctions between actual AI agents and the chat interfaces everyone's calling agents these days. According to Hollan, Partner Director of Product at Microsoft, true agents have "reasoning capability and ability to work toward a concrete goal" — they break work into steps, track progress, and continue until they meet that goal or know to ask for help. Chat interfaces, by contrast, are simply reactive tools that respond to prompts and maybe call a function or two.
Hollan's perspective matters because his team builds the infrastructure that lets developers deploy enterprise AI agents at scale, and he's seen what actually works in production versus what dies in pilot purgatory. The successful enterprise agents he's tracking aren't replacing entire roles — they're automating "tasks requiring a lot of effort that are repeatable and well-bounded." Think support case triage, first-level analysis, research prep, and sales engagement preparation. The common thread: clear scope, trusted data sources, and designed handoff paths to humans.
The biggest production blocker isn't the AI itself — it's data access. Enterprise context lives scattered across office docs, knowledge bases, and data lakes, and pulling that together while meeting security and compliance requirements is where most agent projects stall. Hollan's team is betting that the agents that succeed will be the ones that solve this integration problem first, not the ones with the fanciest reasoning capabilities.
For developers building agents, this suggests focusing less on making your agent "smarter" and more on making it reliably connected to the right enterprise data sources. The winners won't be the most impressive demos — they'll be the ones that can actually plug into existing business workflows without breaking compliance or requiring massive infrastructure overhauls.