Zillow, the actual property know-how firm, doesn't get one dialog with its clients. They transfer from a telephone display to a mortgage officer to an actual property agent, typically over months or years, and anticipate the context to observe them. A single chatbot might by no means carry that thread.
At VB Remodel 2026, Zillow SVP of Engineering Toby Roberts and Glean co-founder and CEO Arvind Jain described how they constructed AI structure meant to hold context throughout that total journey — and why context, not uncooked knowledge, turned out to be the tougher drawback to resolve. Zillow's merchandise contact roughly 80% of U.S. actual property transactions annually, and the corporate has been utilizing AI lengthy earlier than ChatGPT existed.
"We fairly rapidly recognized that we have been going to want a persistent context layer that was going to satisfy our clients and the professionals wherever they have been," Roberts mentioned.
Information was by no means the arduous half
Roberts mentioned Zillow's AI effort began the place most enterprise AI efforts begin, with the information itself.
"We began with a big push round ensuring our knowledge did have the correct basis," Roberts mentioned. That meant an information mesh method, clear knowledge lineage and a governance construction with permissions and identification hooked up to the information itself.
None of that turned out to be the arduous drawback. The arduous drawback was constructing one thing that remembered the place a buyer was of their journey and carried that ahead, irrespective of which floor they confirmed up on subsequent.
"This context layer has to dwell to have the ability to help you the place you might be at any given level in your journey," Roberts mentioned. Zillow selected to personal that layer itself relatively than depend upon a single exterior chat interface, a choice Roberts mentioned the group reached rapidly as soon as it regarded on the form of an actual transaction relatively than a single dialog.
Why Zillow constructed its personal structure, and the place Glean suits into it
Zillow constructed its personal harness relatively than route clients by a single mannequin API. The group drew on 20 years of machine studying historical past behind merchandise like Zestimate, leaning into smaller, task-specific fine-tuned fashions as a substitute of 1 general-purpose mannequin.
Internally, that harness runs alongside Glean. Roberts mentioned Zillow now has hundreds of Glean brokers in manufacturing, dealing with repetitive duties with tens of hundreds of executions throughout the corporate. Glean's pitch, per Jain, is centralizing that integration work as soon as, by the Glean MCP gateway, relatively than letting finance, authorized and advertising and marketing every rebuild their very own connections to the identical programs.
That centralization can also be a price lever. Jain pointed to 2 mechanisms: mannequin routing, which sends most duties to smaller, cheaper fashions as a substitute of defaulting to frontier fashions, and precomputed context, which avoids an agent burning tokens assembling its personal context from scratch.
"Claude can also be very gradual as a result of the primary a part of assembling that context really takes without end," Jain mentioned. Routing that request by Glean as a substitute, he mentioned, can reduce token consumption by as a lot as half.
What Zillow and Glean's method means for enterprises
Throughout knowledge, value and permissions, the session provided a number of sensible takeaways for enterprises constructing agentic AI on their very own programs.
Construct the measurement baseline earlier than the AI push, not after. Roberts mentioned Zillow's potential to credibly attribute a 40% enhance in shipped code to AI adoption rests on a DORA metrics baseline the group put in place years earlier, not on the AI rollout itself.
Centralize context as soon as as a substitute of letting each group rebuild it. Jain's core argument for Glean's platform is that duplicated integration work throughout finance, authorized and advertising and marketing groups is a hidden value most enterprises haven't accounted for.
Don't assume permission inheritance is sufficient for regulated knowledge. Even with a permissions-aware context platform in place, Zillow layered arduous guidelines and a standing compliance test on high for its most delicate classes, relatively than trusting the structure to deal with it robotically.
Deal with context as a price lever, not only a functionality. Mannequin routing and precomputed context have been the 2 mechanisms Jain pointed to for chopping AI spend, each geared toward lowering wasted token consumption relatively than including new functionality.
"Fashions by themselves should not sufficient to deliver automation with AI inside your enterprise," Jain mentioned. "You do have to attach it along with your enterprise context."
