How do you bridge the GenAI Divide and make your production agents profitable? Léo Martin reveals Malt's internal methodology, proven across more than 80 active workflows with concrete ROI.
Imagine the scene: a company invests tens of thousands of euros in a revolutionary AI agent designed to automate customer support. Two weeks after deployment, they completely backpedal: the agent hallucinates, misses key context, and annoys users. This is the daily reality for 95% of companies that stumble with AI and get zero ROI. This is what researchers call the GenAI Divide.
💡 What is the "GenAI Divide"? It is the massive chasm that separates the prototype (the cool "POC" that works in a closed testing environment) from industrial AI (robust, monitored, secure, and profitable at the scale of an entire company). Crossing this chasm means stopping "gadget" innovation to build real infrastructure.
To cross this barrier, our teams chose to treat AI not as a magical tool, but as a genuine system engineering subject.
In 24 months, Malt has deployed and stabilized more than 80 assistants and workflows in production. At the latest AI Malt Academy, Léo Martin (AI & Automation Engineer at Malt) shared our internal methodology. His golden rule for moving from a simple experimental "POC" to industrial efficiency? "Test, Train & Throw."
“Building an AI agent in a test environment takes two days. Making it reliable, monitored, and adopted at scale takes months of iterations. At Malt, 40% of AI initiatives are voluntarily abandoned along the way. That is the price of true infrastructure.”"
What you see on social media: An autonomous agent that navigates company tools perfectly and delivers flawless answers on the first try.
The reality on the ground experienced at Malt:
Léo Martin shares an essential golden rule: "Test, Train & Throw." At Malt, 40% of AI initiatives are voluntarily stopped along the way. Let’s be honest: in advanced engineering, knowing when to abandon a path that does not create value is not a failure—it is a proof of maturity. An AI project that never evolves or encounters no obstacles often hides window-dressing innovation. To build a solid infrastructure, you must accept testing, learning, and sometimes pivoting.
To cross the GenAI Divide, Malt structured its teams into two distinct expert profiles, supported by our cross-functional teams:
Léo Martin points out that to successfully deploy a workflow in a company, technical talent alone is not enough. You need to bring together 3 key profiles from the very first week:
A curious C-Level: To secure budget, fast decision-making, and political cover (at Malt, CTO Claire Lebarz and VP Platform & Agentic Systems Anaïs Ghelfi were major sponsors from day one).
An operational manager: Without whom there is neither field adoption nor validated use cases.
An AI-native junior: To bring unfiltered creativity and avoid simply re-implementing old processes with an AI overlay.
This transition toward industrialization does not come out of nowhere. Data from the latest Malt Tech Trends shows an unprecedented acceleration: demand for the creation of AI agents has surged, while workflow automation projects on tools like n8n have experienced spectacular growth of +1,390%.
Yet, the market faces a double challenge: search volume is exploding, but team trust still needs to be built. A complementary study highlights a major paradox: 90% of employees admit to using ChatGPT on their personal accounts, while only 55% trust the internal tools provided by their company.
It is to resolve this friction and capture this "Shadow AI" that Malt's robustness framework was designed.
To transform the chaos of raw data into business value, Léo uses a pragmatic approach divided into four technical pillars.
Before automation comes a fine understanding of the business need. For example, for Malt's Customer Care department, teams started with 6 months of 100% manual processing.
Expert tip: Do not try to code immediately. Use this manual phase to identify real bottlenecks and map out edge cases. If the initial data is of poor quality (Garbage in, garbage out), the agent will only amplify the mess.
Ne demandez pas un "Agent" quand un simple script séquentiel déterministe suffit. Léo recommande de catégoriser précisément le besoin:
The AI Assistant: conversation-oriented (e.g., redirecting a user to the right internal documentation).
The AI Workflow: a sequence of deterministic tasks structured as a linear process (e.g., using n8n to process product feedback step-by-step).
The AI Agent: an autonomous decision tree capable of choosing its tools based on context (e.g., automated routing and pre-drafting of complex support tickets).
The classic mistake is forcing employees to open a new tab or yet another web interface. Your main competitor in a company is not OpenAI; it is the reflex to open chatgpt.com on the side. The solution? Push AI directly into the existing workflow: a Slack bot (like Malty AI at Malt), a Chrome extension, or a native integration in Salesforce or Notion.
You cannot improve what you cannot see. From day one of production, the architecture must integrate a strict logging system (like Langfuse) to allow Product Managers to modify prompts and datasets without depending on developers. Evaluation must be hybrid: combining classic Machine Learning metrics and an LLM-as-a-Judge validation layer for the semantic aspect.
The experience gained by Malt across its 80+ production workflows can be summarized in 6 major key learnings:
The shift to automation: do not automate until the business team is fully onboarded AND the outputs are 100% reliable.
The trap of anonymous users: in Malt's Help Center, 98% of traffic is anonymous. Without a clear identity, it is impossible to push deep automation.
Classify before solving: tag tickets by priority and tackle a single, very specific use case at a time.
Substance over form: your agent's tone of voice should be worked on after the relevance of the substance. Reversing this order wastes an average of 3 months of development.
Day 1 monitoring: integrate a feedback button (tags, emojis) from the very first day in production, otherwise you are flying blind.
Specialize for quality: it is better to design 5 highly specialized agents than a single generalist agent. Specialization is what unlocks industrial-grade quality.
Industrializing AI represents a strategic investment, but contrary to popular belief, the cost is not where you expect it. At Malt, the budget breakdown highlights a clear reality: the vast majority of the investment is human (dedicated to the time of Builders and the Core Team), while the purely technological line (Gemini API, Claude, platforms like n8n or Dust) represents only a minor fraction of the overall budget.
The return on investment was achieved in 12 to 18 months, with major gains equivalent to 12.5 Full-Time Equivalents (FTE):
A company's competitive advantage no longer lies in the choice of its language model, which has become a commodity accessible to all. Value is found in the intelligence of the orchestration, data governance, and the ability to support humans.
For freelancers and companies alike, boundaries are blurring: the roles of Growth Marketer, Product Manager, and engineer are merging into a pivotal profile: the AI and automation engineer. Today, this is the most strategic hire for scaling organizations.
“The future of operational efficiency is not to replace humans, but to empower them to oversee autonomous digital departments capable of managing end-to-end processes.”"
Want to dive into the code, discover behind-the-scenes details of our internal workflows, and avoid production deployment pitfalls?
👉 Watch the full replay of Léo Martin's Malt Academy on our YouTube channel.
Léo Martin is an AI & Automation Engineer at Malt. He designs and deploys multi-agent architectures, automation infrastructures, and LLM Ops pipelines that run the company daily. His expertise helps bridge the gap to industrialization, turning technological adoption into a sustainable engine for ROI.
Need to automate your business processes? Consultez les profils de nos experts IA sur Malt.
