Reindeer Co-CEO Yoav Naveh: Agents Must Learn or They're Just RPA

By James Calloway

Reindeer co-founder and co-CEO Yoav Naveh makes a blunt claim about enterprise AI: the industry's dominant deployment model — teams of forward-deployed engineers (FDEs, vendor staff embedded at the customer to build and maintain agents) — collapses as agent counts grow. "The FDE model works when you have 5 agents," Naveh said in an interview. "It breaks down when you have a hundred agents."

His company builds what he calls adaptive agents for core operations — finance, supply chain, logistics, and shared services — that learn from expert feedback in production rather than requiring engineers to re-code them when processes change. It is a direct attack on how much of the agentic AI market currently operates. The stakes are straightforward: if Naveh is right, a large share of today's enterprise AI revenue is services dressed up as software. If he is wrong, the FDE vendors can point to their production deployments and Reindeer's promise of self-improving agents remains a claim to verify.

What does Reindeer actually sell?

Naveh's background is operational: he built ConvertMedia, an ad-tech company, continued at Taboola after its acquisition of that business, and invested in founders through INT3. He co-founded Reindeer with Yair Adato. The pitch he offers now is not better models but a different labor economics.

"The AI market was too exciting for us to watch from the sidelines," Naveh said. He and his co-founder identified "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."

Reindeer's platform combines workflow execution with expert feedback, evaluation, and human approval. The stated goal is to take enterprises "from their first agent to an entire herd of agents that learn from human feedback and improve over time" — without rebuilding or constantly maintaining every agent. Take that framing with a caveat: it is a founder's framing, and the interview names customers but offers no third-party benchmarks.

The market context matters here. Enterprises have watched RPA (robotic process automation, rules-based software that mimics clicks and data entry) age badly as business processes evolved. Naveh invokes that history deliberately. "We've already seen how brittle RPA became as business processes evolved," he said. "We shouldn't expect hand-built agents to be different."

Where embedded engineering stops adding value

Naveh does not dismiss FDEs outright. He concedes they are "valuable for complex integrations and getting initial agents into production." The problem starts when the processes themselves change and the embedded engineers 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," he said.

His sharpest point is contractual, not technical. Most vendors deploying FDEs claim their engineers eventually move on to other tasks. Naveh's advice to buyers: "Enterprises should demand proof that is the case before signing a contract that requires FDEs." That is a rare public invitation to interrogate vendor staffing assumptions, and it cuts against an industry that books much of its AI delivery as billed human hours.

How do agents learn without replicating mistakes?

The core of Reindeer's technical claim is a distinction between resolving one case and changing agent behavior going forward. Naveh walks through a concrete example with an invoice agent:

• The agent receives an invoice 4% over the purchase order and finds no matching policy.
• It escalates to a human expert rather than guessing.
• The expert explains that this supplier has an approved exception for fuel surcharges.
• The agent resolves the case, and stores the explanation as context for that supplier only.

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

What changes in the system is not model weights (the learned parameters inside a foundation model) but the agent's context and skills: workflow logic, policies, and instructions on when to apply each. Naveh is explicit that copying humans is the failure mode. "The goal is to learn from humans, not blindly replicate human decisions," he said. "Otherwise, you're simply automating mistakes." He even notes that two different employees may resolve the same customer's exception differently — a contradiction his system is built to surface and reconcile rather than encode.

His design analogy is onboarding a new hire. Most agentic vendors start by documenting a process exhaustively and coding it from interviews and historical data. Naveh argues people don't learn that way: "They learn a few skills, see a few examples, start working, ask questions when they see something they do not understand, and build their knowledge. This is the process we built." A 3,000-page SOP, in his telling, is not how institutional knowledge actually transfers.

What testing stands between a proposed change and production?

Continuous learning without governance is where most skepticism about agents concentrates, and Naveh names it plainly: "Continuous learning without governance is dangerous." Every proposed change passes three safeguards:

• Regression testing against representative cases replayed for both current and proposed agent versions.
• Human approval — a manager reviews the change, evidence, and test results, and can approve, revise, or reject it.
• A confidence threshold that controls when the agent acts autonomously; anything below it routes to an expert.

The regression testing deserves scrutiny, because Naveh admits "operations work has no fixed answer key." Reindeer selects representative cases from expert feedback, defines comparison criteria, and replays them. Actions with side effects — sending an email, inserting a record into SAP — run in a simulator built from past runs. Naveh calls the evaluation platform "a 'Truman show' for agents": agents must believe they live in the real world for the evaluation to mean anything. Stale test cases that no longer match current policy are retired.

The confidence threshold is tuned by the customer's operations team, not by Reindeer. Teams can start conservatively and grant autonomy as the agent's track record builds. If quality drops after a change, the system automatically tightens review thresholds, pushing more cases to experts whose feedback surfaces the problem. Every policy change is versioned, attributable, and logged, with rollback to the previous version.

That is the design on paper. The interview does not include independent audit data on how often rollbacks occur or how many proposed changes survive testing — the numbers a buyer would want before trusting the loop.

Is Reindeer reducing human work or just moving it?

Naveh's most quotable answer addresses the obvious objection: cutting engineering hours may simply shift effort onto operations teams reviewing exceptions. He reframes the FDE problem as a cost transfer. "The problem with the FDE model is that it often moves work from relatively low-cost operational employees to expensive engineers," he said. "That's not solving the labor problem, it's making it more expensive."

Reindeer measures three numbers, and Naveh commits to a falsifiable standard:

• The share of cases the agent resolves end to end without a human.
• The number of expert touches per case.
• 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. It is the strongest line in the interview, because it hands buyers a metric that would expose failure. Reindeer does not disclose current values for those metrics across its customer base.

What happened at Hellmann Worldwide Logistics?

The one deployment Naveh details is Hellmann Worldwide Logistics, the global freight forwarder, on its pricing team. The team manually processed shipping requests across emails, spreadsheets, and other documents, often taking days to produce a quote.

Reindeer started with roughly 20 sample requests to understand the workflow and train the agent. Hellmann's pricing experts then worked directly with the system, providing feedback and navigating exceptions. Reindeer tracked accuracy and speed through testing and continued monitoring after production launch.

Today, routine requests are handled automatically and complex cases escalate to the team. Naveh highlights one behavior as critical: the agent recognizes when information is missing rather than making assumptions. Reported results are a reduction in quote turnaround time, fewer manual data-entry tasks, more time for customer-facing work, and an improved quote win rate. He offers no specific percentages for any of those outcomes — a gap worth noting in an otherwise concrete account. Papaya Global, the payroll and payments company, is also a named customer, but Naveh does not describe that deployment.

Which workflows are ready — and which are not?

Naveh is candid that agents are not universal. The strongest opportunities are repetitive, high-volume workflows requiring contextual judgment, particularly where information spans multiple systems and employees regularly handle exceptions. He also sets an explicit limit: "Core operations will not be fully automated, and we don't aim for that. Agents expand their coverage gradually, and high-risk cases stay with people by design."

His structural prediction is organizational. As companies move from a few agents to a herd, they need a function — analogous to how revenue operations emerged with CRM and DevOps with cloud — where a person or team manages agents the way a department head manages a workforce: accountable for KPIs and ROI, deciding where agents take on work, and approving how they change. "The companies already building this function are moving fastest, and they are the ones positioned to win," he said. He adds that conflicting expert practices are not a reason to wait; resolving them is part of that agent-owner's job.

Where does the durable advantage sit?

Naveh expects commoditization at the bottom of the stack: foundation models will become "increasingly interchangeable as they get cheaper, faster and more capable," and building basic agents connected to enterprise systems is already getting easier.

His bet is that the durable asset is proprietary knowledge of how a specific enterprise actually operates — a moving target, refreshed by every production case. "Models will get cheaper. Building agents will get easier," he said. "The winners will eliminate the expensive human engineering infrastructure required to keep those agents working and relevant over time."

The claim is coherent, and Reindeer's guardrails — evaluation, versioning, human approval, the three-metric success test — are the right architecture on paper. What remains unproven in this interview is scale: no agent counts at customers, no disclosure of autonomous-resolution rates, no third-party verification of the Hellmann results. Naveh has defined the correct test for his own company; the burden now is publishing the numbers.