Frontier intelligence currently is a buyer’s market. GPT 5.5, Opus 4.7, Sonnet 4.6, GPT 5.4 — for nearly every enterprise workflow, they’re functionally interchangeable, and any new flagship from one lab is matched within weeks by another. Rahul Dey, a product manager leading LLM infrastructure at the enterprise AI company Glean, thinks that shift is the most under-discussed change in the industry. It’s also why he believes most teams are asking the wrong question when their AI deployments fail.
“When a deployment is broken in production, almost no one actually needs to switch models,” Dey says. “The system either can’t see the data it needs, or it can see the data but the answers aren’t trustworthy, or the unit economics fall apart the moment the customer scales. Model choice is the last variable I’d touch, not the first.”
Two years ago, that wasn’t true. The gap between the best and the second-best model was wide enough that picking correctly mattered. The ceiling on what an enterprise AI deployment could do was set, in large part, by which model the team chose. That ceiling has since been lifted. The labs ship a new flagship every few weeks, and most of them clear the bar that any serious enterprise workflow actually requires. What hasn’t changed is everything around the model: the data integrations, the evaluation loops, the cost structure. The constraint moved, but most teams kept looking in the old place.
Dey came to Glean in 2025 from Microsoft, where he spent three years on deepfake detection and on integrating Copilot into Microsoft Teams for frontline workers, with a year in between at the healthcare-operations startup Athelas. Two high school friends who were engineers at Glean offered him the chance to work on infrastructure that could matter across every industry at once. Glean indexes data across more than 130 enterprise applications — Slack, Salesforce, Jira, NetSuite, Google Drive, dozens more — and layers AI on top so employees can ask questions and finish work without hunting through siloed tools. The company has been valued at over $7 billion. Dey leads the team responsible for building the platform that powers most of Glean’s AI Products. He focuses on: model selection, evaluation, and cost.
Two initiatives sit at the center of his work, and they’re each a direct response to the new market dynamic.
The first is evaluation. Glean is defining golden-standard outputs for every major use case the platform supports — what a great answer to a sales rep’s question actually looks like, what a great Jira ticket summary actually reads like — and benchmarking every new frontier model against those standards using a combination of human reviewers and LLM-based judges. The point is to make model selection an empirical question grounded in customer workflows, not a marketing one driven by benchmark scores.
“The labs ship a new flagship every few weeks now,” Dey says. “Without a rigorous way to measure whether each one is actually better for the workflows your customers care about, you just chase the leaderboard. But the leaderboard isn’t your customer.”
The second is intelligent model routing. Frontier inference is expensive, and most enterprise queries don’t need it. A simple lookup against an indexed document doesn’t demand the same compute as a multi-step reasoning task spanning five systems. Dey’s team is building the routing layer that matches each query to the most efficient model that can answer it correctly. It’s a margin project as much as an infrastructure one and it spares customers from being charged for a drill when a screwdriver would do. The same effort produces customer-facing controls over token spend, so enterprises can cap AI usage the way they cap a corporate card.
Dey has unusually direct visibility into the convergence he’s describing. Glean partners with the major research labs and gets early previews of unreleased models, sometimes weeks before public launch. From that vantage point, he says, the homogenization at the frontier is obvious as are the consequences. When everyone can buy roughly the same intelligence by the token, intelligence stops being the differentiator. What replaces it is who can convert that intelligence into measurable customer outcomes most efficiently.
That, Dey argues, is the work most enterprise AI teams are still avoiding. Connecting the systems. Defining what a good answer looks like for each customer’s job. Routing each query to the cheapest model that can handle it. Capping the bill before it surprises the CFO. None of it is glamorous but it all compounds. And in a market where the model is no longer the variable, it’s where the next decade of enterprise AI margin and adoption will actually be decided.
“In two years, no enterprise customer is going to switch vendors because someone integrated a slightly newer model,” Dey says. “They’ll switch because someone delivered a measurably better outcome at a lower cost. That’s the only competition that matters now.”
Disclaimer: GeekWire newsroom and editorial staff were not involved in the creation of this content..
