Reindeer Co-CEO Yoav Naveh: 'The FDE Model Works With 5 Agents. It Breaks Down at a Hundred'

Reindeer co-founder and co-CEO Yoav Naveh says the forward-deployed engineering model that dominates enterprise AI "works when you have 5 agents" but "breaks down when you have a hundred agents." His company builds adaptive AI agents for core operations in finance, supply chain, logistics and shared services, and it stakes its pitch on a claim most rivals avoid: agents must learn from human experts in production, under governance, without permanent engineering support.

Naveh co-founded Reindeer with Yair after a career that included building ConvertMedia, working within Taboola and investing in founders through INT3. In an interview, he laid out both the company's mechanism — expert feedback, simulated regression testing, human approval — and its candid limits. Notably, he offered no customer-specific accuracy figures or exception rates, even when asked directly.

What problem does Reindeer claim to solve?

"The AI market was too exciting for us to watch from the sidelines," Naveh said. He and his co-founder focused on "the gap between how useful AI has become when we work with it individually and how difficult it remains for enterprises to adopt it at scale."

The platform, as he describes it, "takes enterprises from their first agent to an entire herd of agents that learn from human feedback and improve over time." The goal: companies turn core operations into an agentic AI workforce "without having to rebuild or constantly maintain every agent."

That is a promise about labor economics, and Naveh knows it invites scrutiny. His stated success metrics are concrete: the share of cases agents resolve end to end without a human, expert touches per case, and total handling time per case across agents and people combined. "If the third number isn't going down, we're just moving the work around," he said. No baseline or current numbers for those metrics were disclosed.

Why does Naveh attack the forward-deployed engineer model?

Forward-deployed engineers, he conceded, are "valuable for complex integrations and getting initial agents into production." The problem starts when work processes change and FDEs become permanently necessary to keep agents running.

"If every new workflow or unexpected exception requires more engineering, you're not scaling software, you're scaling expensive human labor," Naveh said.

His advice to buyers carries a skeptic's edge: most vendors claim their engineers eventually move on, so "enterprises should demand proof that is the case before signing a contract that requires FDEs." He draws a parallel to RPA: "We've already seen how brittle RPA became as business processes evolved. We shouldn't expect hand-built agents to be different."

He also reframes the cost question: "The problem with the FDE model is that it often moves work from relatively low-cost operational employees to expensive engineers. That's not solving the labor problem, it's making it more expensive."

How do agents learn without copying human mistakes?

Most enterprises hold institutional knowledge in workers' heads, not documents. Reindeer's agents start with partial knowledge and ask subject-matter experts when they hit unfamiliar situations. The design principle: learn the way a new hire does, not by memorizing "a 3,000 page SOP."

Naveh walked through a concrete exception. An invoice agent receives an invoice 4% over the purchase order and finds no covering policy. It escalates. An expert explains the supplier has an approved exception for fuel surcharges. The agent resolves the case and stores the explanation as context for that supplier only.

If the same exception recurs across suppliers and other experts resolve it the same way, the system proposes a policy change — for example, "fuel surcharges up to 5% are acceptable when the contract allows it." That proposal passes testing and human approval before use. "One expert's answer resolves one case. A pattern across cases becomes a candidate for a new rule."

The distinction matters because individual decisions do not generalize. "Two different employees may have a different resolution for the same customer's exception," Naveh said. The system evaluates circumstances, identifies recurring patterns, reconciles contradictions and flags conflicts with existing policies. Proposed updates run through an evaluation platform Naveh calls "a 'Truman show' for agents," then go to a human for approval.

"The goal is to learn from humans, not blindly replicate human decisions," he said. "Otherwise, you're simply automating mistakes."

What safeguards gate a change before production?

"Continuous learning without governance is dangerous," Naveh said. Every proposed change passes three safeguards:

• Regression testing. The system replays representative cases from expert feedback against both the current agent and the proposed version. Operations work has "no fixed answer key," so Reindeer selects cases, defines comparison criteria and retires test cases that no longer match policy.
• Human approval. A manager reviews the change, the evidence and the test results, and can approve, revise or reject. "Nothing changes in how the agent operates until that approval."
• A confidence threshold. The agent acts autonomously only above the threshold; anything below goes to an expert. Operations teams tune it themselves, starting conservative and loosening as the track record builds.

Actions with side effects, such as sending an email, run in a simulator built from past runs. "Agents have to believe they live in the real world, and their actions have real implications — a record inserted to SAP, or an email sent to a customer," Naveh said. Every policy change is versioned, attributable and logged, with rollback available. If quality drops, the system automatically tightens review thresholds so more cases reach experts.

What happened at Hellmann?

Naveh cited one deployment in detail: Hellmann Worldwide Logistics, where a pricing team spent days manually processing shipping requests across emails, spreadsheets and documents to produce a quote. Reindeer started with roughly 20 sample requests to learn the workflow. Hellmann's pricing experts then worked directly with the system, giving feedback and navigating exceptions.

Today, routine requests are handled automatically and complex cases escalate to the team. Naveh said the agent recognizes missing information "rather than making assumptions." The claimed results: reduced quote turnaround time, fewer manual data-entry tasks, more customer-facing time and an improved quote win rate. He gave no percentages for any of those outcomes. Papita Global — Papaya Global — was named as a customer but not described.

"The objective isn't to eliminate humans from the process, but to continually reduce the amount of routine work requiring their intervention," Naveh said.

Where does the approach stop working?

"Not every process needs an AI agent," Naveh said. The strongest fits are repetitive, high-volume workflows requiring contextual judgment, where information spans multiple systems and employees regularly handle exceptions.

He also named a hard ceiling: "Core operations will not be fully automated, and we don't aim for that." Agents expand coverage gradually, and high-risk cases stay with people by design.

His broader prediction is organizational. As companies move from a few agents to a herd, they need a function — like revenue operations with CRM, or DevOps with cloud — where someone manages agents "the way a department head manages a workforce: accountable for their KPIs and ROI." Conflicting expert practices are not a reason to wait, he argued; the system surfaces contradictions and turns them into policy proposals that this owner resolves.

Can the advantage survive cheaper models?

Naveh expects foundation models to become "increasingly interchangeable as they get cheaper, faster and more capable," and basic agent building to commoditize. "Building an agent isn't the same as making it reliable inside a complex business, over time," he said.

The durable edge, in his telling, is proprietary knowledge of how an enterprise actually operates — "a moving target" where every production case teaches the agent future exceptions. "The winners will eliminate the expensive human engineering infrastructure required to keep those agents working and relevant over time," he said. That thesis is testable in Reindeer's own metric: if total handling time per case keeps falling at customers like Hellmann, the model holds; if not, the company has merely moved the labor, exactly as its co-CEO warned rivals would.