CHICAGO — Enterprise AI agents will not be differentiated by model quality going forward, according to Yoav Naveh, founder and CEO of Reindeer. That conviction is reshaping where AI startups are placing their strategic bets.

In an interview with FreightWaves at the Supply Chain AI Symposium in Chicago, Naveh described an industry that has cycled through two distinct waves of adoption without delivering the returns executives anticipated — and is now pursuing a third, more ambitious wave focused on core business operations. It is a shift taking place as enterprise generative AI spending continues to climb, yet surveys of CIOs repeatedly show that scaling pilots into production remains the top barrier to realizing returns.

Two Waves of AI Adoption Have Fallen Short

Naveh traced the trajectory through clear stages. "The first wave: assistants. Everybody got a copilot, Claude, ChatGPT," he said. "We've seen a lot of success with assistants around software engineering. Coding agents work really well. Maybe legal, not as well. But generally we haven't seen assistants make companies 10x or 100x more efficient."

The second wave zeroed in on customer support and call centers. "I think that has been a very big low-hanging fruit, really," Naveh said. "One big workflow, a lot of similar case types in one. I definitely see some success there."

Now, enterprises are turning toward the area that accounts for the majority of their labor expenditure: core operations. "Let's take all these core workflows that represent 70, 80% of the work that everybody does," Naveh said.

This market is considerably more complex than customer support. Naveh estimates that a single enterprise may contain dozens, if not hundreds, of such workflows — spanning finance, procurement, treasury, accounting, and supply chain — frequently with no single department holding clear ownership.

The Fragmentation Problem Investors Overlook

Naveh pointed to a telling example from one client he identified as among the largest consumer packaged goods companies globally.

"They say, 'Hey, we're doing accounts payable in my company,' but your accounts payable is different from their accounts payable," Naveh said. "We work with one of the largest CPG companies in the world. Five different departments, five different accounts payable processes within the same company."

This internal fragmentation is precisely why Naveh is skeptical of the forward-deployed engineer model that has gained traction among AI startups, where vendors embed technical staff directly inside a client organization to build agents one process at a time. The model, closely associated with Palantir's commercial expansion and since adopted by a new generation of AI startups, is prized for delivering bespoke deployments but has faced persistent questions about gross margins and scalability.

"I'm going to throw three engineers in and sit with your team and build the agent for that particular process. Are they ever going to come back and be available to be sent to another department, or are they stuck maintaining whatever they built as processes change and evolve?" Naveh asked. "I think the math doesn't work because it doesn't scale."

Why the Model Won't Be the Moat

Reindeer's strategy challenges a prevailing assumption in enterprise software: that owning or fine-tuning a proprietary large language model constitutes a durable competitive advantage. Naveh contends the opposite is unfolding — and the trajectory of open-source models lends weight to the argument. Releases such as Meta's Llama series and Mistral's models have rapidly narrowed the capability gap with frontier proprietary models, reinforcing the view that model quality alone is a diminishing differentiator.

"Strong belief that LLMs are going to be a commodity," Naveh said. "We see that open-source models are maybe six months behind the frontier models. But what we're thinking is, I don't need the frontier model for everything anymore."

Reindeer's approach is model-agnostic in practice. "We work now with a large European insurance company. They decided to build their own LLM. We're happy to plug in their own LLM," Naveh said. "We're happy to restrict the list of allowed LLMs based on the company's data privacy policy, and figure out how to run their case with what's whitelisted."

If the underlying model becomes interchangeable, any competitor whose pitch relies primarily on model superiority faces a narrowing advantage.

Building Is Easy. Maintaining Isn't.

In Reindeer's framework, the real differentiator lies in the maintenance layer: the capacity to detect when a deployed workflow drifts from its original configuration, reconcile the change, and recalibrate the agent without requiring a fresh engineering engagement. The challenge Naveh describes mirrors a well-documented problem in machine learning operations, where model and data drift gradually degrade production systems that performed well at launch.

"Even after you deploy an agent, the workflow keeps changing," Naveh said. "If you don't have a method to have agents detect these changes, reconcile, self-heal, self-learn, we're not really going to be able to realize what agents can do for you." To address this challenge, Reindeer developed what Naveh calls a two-loop system designed to manage workflow drift.

Naveh noted that the economics of building AI agents have already shifted decisively in the industry's favor — which is precisely why maintenance has emerged as the critical bottleneck.

"I think everybody wants to build. Building is easy," he said. He recounted the experience of building an internal dashboard for a colleague in a single evening, only to face a continuous stream of bug fixes and feature requests thereafter. "I don't want to do that," he said. "That's not fun. Maintaining is not fun or easy."

"Five years ago you couldn't build anything, now you can build amazing things," Naveh said, pointing to live coding assistants as tools that have dramatically narrowed the gap between novice and expert engineers. "The expertise that is needed now is exactly what you said, around all these different things after it's built, and make sure that what was amazing on the first evening you released it is still amazing six months later."

For enterprises evaluating AI vendors, Naveh's framework points toward a fundamentally different diligence question. Rather than asking which company demonstrates the strongest model at the proof-of-concept stage, the more durable question may be which vendor can sustain agent accuracy well beyond launch — as the workflows underneath continue to shift.