Chen Markman is the CEO of Avon AI, a company he started in early 2025 with his co-founder to help organizations deploy and manage GenAI agents. Speaking with Laurent Cochet in Tel Aviv, he explains why the real work begins after an agent goes live, how regulated companies such as banks and insurers can gain trust in conversational AI, and why he believes the manager of an AI agent should be a human.
Lessons from machine learning in the field
Markman served in an elite unit of the Israeli army called Talpiot, then worked on machine learning research at Pagaya, an Israeli fintech company traded on the NASDAQ. There he saw that you cannot simply train a model, put it in the field and forget about it. Most of the machine learning experts, he says, were not building new models but monitoring existing ones to keep them behaving as intended.
He went on to manage the AI group at Dream Security, working on agentic AI and LLMs for cybersecurity. The lesson repeated itself: AI is not plug and play, and reality is, in his words, the worst enemy of the AI researcher. He believes too few people invest in what happens after a capability is deployed, and that this is the key to success over time.
The gap between demos and deployment
His co-founder, Amit Segev, managed the AI and data department at One Zero Bank, a digital-first bank that Markman says was one of the first banks in the world to deploy a real GenAI chatbot that now handles most customer engagement end to end. Despite the noise around GenAI, he argues, most companies people interact with do not yet deploy high-quality agentic AI, and the gap is biggest where errors are costly: insurance, financial organizations and banks.
Companies such as Lemonade and Klarna have deployed high-quality agents, but he notes they are tech-focused, with many engineers and domain experts keeping quality high after launch. Avon AI aims to give more traditional companies the tools for that work: monitoring performance, finding problems and their root causes, fixing them and testing that the changes are robust.
Treating AI agents like new employees
Markman compares the platform to a human customer service organization, where supervisors listen to conversations and give feedback, other teams gather knowledge, and new employees are trained and tested before speaking with customers. Avon AI brings those processes to AI agents, treating them as a new type of employee managed by someone from the business domain.
Monitoring means catching conversations that loop or customers who repeat the same question without getting an answer. Governance means business-specific rules, such as making sure an agent gives the appropriate disclaimer whenever it discusses investment. Because agents hold thousands of conversations every day, he says, no one can review them manually, so the system has to surface the most problematic ones.
In his view, the manager of an AI agent should be a human. The goal is to empower business leaders to bring their domain expertise to managing agents, not to replace people, which he considers neither feasible nor right in complex settings like insurance and finance.
Trust and the most common mistakes
The biggest worry for regulated companies, Markman says, is that GenAI is a very open technology: it might promise something outside policy or talk about competitors or politics. More control and visibility, he argues, is what builds the trust needed to deploy. Banks can start with general information, move to user-specific information and then to tasks, though even general answers about insurance coverage carry accountability.
A common mistake is assuming a demo is nearly a product. By his account, it is easy to build an agent that works 70 percent of the time, but getting it to 98% is really tough. He also tells customers not to overthink use cases: deploy something, then measure what customers actually want. Without monitoring, he notes, publicized cases have included an AI agent selling a truck for $1.
Where adoption goes next
Today, he says, most agents work internally as copilots or assistants rather than open conversational agents facing customers. He expects many more companies to deploy in 2026, starting where errors are less costly, such as online retail, with insurance, financial and health organizations next. He believes tier one tasks will be automated, since there is no reason for a human to answer the same question for 100 customers every day.
Successful companies, in his experience, treat GenAI as an organizational effort, with business teams monitoring conversations and opening tickets for engineers. He sees a new role emerging: people who used to manage people will manage GenAI agents. People who come from managing service or customer service teams, and are quite technical, can do this very successfully, he says, without being developers. His message to CEOs: an AI system is not magical, it is still software, and it needs effort to manage over time.
Watch the full conversation
Watch the full conversation above. Chen Markman and Avon AI can be found on LinkedIn.
AI InterConnect hosts conversations with founders, investors and operators building with AI. Explore more episodes on the AI InterConnect podcast page.