As part of our exploration into the AI tech stack, we’re speaking with a founder who isn’t just using AI, but building the very tools that are defining the future of AI-native businesses.
Marvel Gomulya is the Co-founder and CEO of Covena.ai, a company that provides advanced B2C AI sales agents designed to handle the entire sales cycle. From initial contact and lead qualification to final checkout, all without human input. Their platform is built to eliminate hallucinations, integrate with the tools founders already love, and scale to thousands of leads a day.
We interviewed him on building, selling, and scaling with AI.
The core idea for Covena.ai – an AI agent that can close a sale from start to finish, is a huge leap beyond simple chatbots. What was the specific founder pain point you experienced that made you realize this wasn’t just a “nice-to-have,” but a necessity?
Initially, my hypothesis was grounded in market dynamics here in Southeast Asia: labor is relatively cheap, so most companies are far less motivated by opportunities to save on operating costs. What really moves the needle is anything that directly increases revenue. Our belief was that AI agents, by improving conversion rates and delivering unmatched scalability, could drive more sales than any human team.
One night, our AI agent infrastructure briefly went down, just 30 minutes, and at midnight. Within minutes, my phone was lighting up with calls from clients. They wanted the system up and running immediately, because every minute offline meant lost sales. For our ideal customers, Covena was directly powering their topline growth, almost like a payment gateway.
That was the turning point. When our product became so critical that customers are “pulling” it out of us, and their business depends on its continuous operation. That’s when I felt Covena stopped being a simple chatbot and became a necessity.
Before you even started building Covena for customers, how did your own founding team use AI as an internal team member to build, market, and operate in the early days?
Before we built Covena for customers, our founding team have already been using AI tools for years. We used AI coding tools to speed up every part of building our software – writing, reviewing, and improving code. In fact, around half of our code was written by AI, and AI even checked our code quality during pull requests.
We made it easy for AI to help by designing everything in clear, modular blocks, kind of like building with Lego pieces. Our team got really good at writing instructions (prompts) for AI and reviewing AI-generated code. We also covered the costs for our team to use any helpful AI tools, like Cursor and n8n, so everyone had access. That’s how we built Covena fast and matched bigger competitors in months.
Your site promises an AI agent that “just works” and is deployed by experts. For founders who aren’t AI engineers, what’s your advice on the “build vs. buy” decision when it comes to integrating AI into their core operations?
Absolutely, if a company has the resources, time, and AI expertise, building in-house can be an option. But most founders seriously underestimate the challenge of building a reliable AI agent that works in real business conditions.
Often, people watch a few tutorials, put together a simple AI demo, and think they’re ready for production. But making an AI agent robust enough for real-world use is much harder than it looks. It’s at least 10 times more complex than demo projects, and anyone who’s tried to launch an AI tool in production with high reliability will tell you the same.
Even if you could make it work, doing so would take significant time and could distract the team from the company’s core operations.
That’s why for most companies, I recommend buying rather than building. Even experienced software teams often end up partnering with us because developing an enterprise-grade AI agent is time-consuming and much more difficult than anticipated. Buying lets founders stay focused on growing their business while we make sure the AI solution “just works.”
A major selling point for Covena is that your AI agents don’t hallucinate and stick to the SOP. From a founder’s perspective, how did you solve this massive trust and reliability problem?
The first step was admitting this was a major problem. Even if an AI is 95% accurate per response, that means 1 in 20 responses is wrong. In a typical conversation with 5 exchanges, only about 77% of conversations are fully reliable, which means the AI makes a mistake in roughly 1 out of every 4 conversations. That’s not acceptable for real production.
Knowing this, we took much longer to build our product than most others. We overhauled our architecture 3–4 times before finding an approach that worked. Then we spent additional time developing custom tools to support this reliability. Honestly, the early months were tough, and progress felt really slow.
But now, I’m confident our AI agents are the most reliable on the market for sales. This reliability lets us do what competitors can’t, for example, we charge based on successful transactions or bookings, not per response. That’s how sure we are that our AI works as promised.
The idea of “onboarding” AI like a new teammate is key for me. What does the “training” process look like when you deploy a new Covena agent for a client? How do you teach it the nuances of a specific business?
A big part of “training” one of our AI agents is actually about training the person who builds the agent. We have a unique role on our team called Forward Deployed Engineers (FDEs)—a concept that started at Palantir in Silicon Valley but is still very rare in Southeast Asia.
Here’s how our process works:
- Our FDEs work directly with clients, spending time at their offices to deeply understand the specific needs, nuances, and workflows of the business.
- Once they have a clear picture, the FDEs translate those requirements into instructions and examples that the AI agent can understand.
- We build an initial version (v1) of the AI agent tailored to the client’s needs.
- The agent is then tested in a limited deployment, with FDEs continuing to collect feedback from both the client team and the agent’s real customers.
- Based on this real-world feedback, we iterate quickly, fine-tuning the agent to maximize both its reliability and conversion rate.
In short, it’s a hands-on, collaborative process where our experts adapt the AI to each business through rapid iteration.
Your AI is built for complex, multi-step sales flows. Could you walk us through an example of a challenging sales conversation that a traditional chatbot would fail, but Covena is designed to handle?
Absolutely! Let me share a real example from just a few days ago, where our AI handled a complex sales process end-to-end over more than a week, something that would be impossible for a typical chatbot.
The case involved a client in the education sector. We needed to guide a parent through enrolling their children in lessons, which turned into a multi-step, multi-day sales flow, requiring the AI to follow up seven times before converting the lead.
Here’s why this scenario was so challenging (and how Covena excels):
- Multiple children and schedules: The AI managed checkouts for more than one child, each with different desired class times.
- Changing decisions: Even after a schedule was confirmed, the customer later changed their mind. Our AI seamlessly updated everything and communicated clearly.
- Persistent, contextual follow-ups: The AI tailored each follow-up based on where the lead was in the process, whether still asking questions or already at the payment stage, never sounding robotic or repetitive.
- Dynamic class schedules: When available slots changed mid-conversation, the AI nudged the customer to select a new schedule, showing adaptability.
- Handling objections: It responded politely and knowledgeably to concerns about pricing and curriculum, maintaining trust.
- Conversational finesse: The AI managed double and triple texts naturally, and it reacted appropriately when the customer replied with simple “Oke thank you”-type messages, keeping the flow human-like.
- Proactive support: When a payment link expired, the AI noticed and automatically generated a new one for the customer, rescuing the deal.
In total, the conversation spanned 20+ turns and was managed entirely by our AI, without any need for human intervention. In fact, our incentive model means that if even one escalation was needed, we wouldn’t have been paid.
A traditional chatbot would likely have failed at one of these many points, losing track of the context, mishandling objections, or dropping the conversation flow. Covena, on the other hand, is designed precisely to handle these messy, nuanced sales conversations that real businesses face every day.
For founders just starting to build their AI stack to use internally, what is the single biggest mistake you see them make when trying to automate parts of their sales and customer interaction process?
Many founders assume they can just plug in an open-source or no-code tool, connect everything via API, and quickly automate their sales or customer interactions. In reality, this almost always fails—especially once you go beyond very simple customer service cases.
Most no-code or off-the-shelf solutions are just too generic. They can’t handle the nuanced workflows, exceptions, and business-specific logic that pop up in real sales processes. As a result, teams end up spending a lot of time working around the limitations, or the automation simply doesn’t deliver the desired results.
What do you think is the biggest misconception that people have about AI agents?
The biggest misconception is that people tend to swing between two extremes:
Either they believe “AI agents can do everything and AGI is just around the corner,” or they decide “all AI agents fail and are useless.”
What usually happens is that people start out overly optimistic, try an AI agent on a tough problem, see it struggle, and then swing to the other extreme, dismissing all AI. But the truth is somewhere in the middle.
You can absolutely build reliable, high-performing AI agents but typically only for specific verticals or well-defined use cases.
AI agents shine when they’re designed for a clear purpose, with the right data and integrations, not as one-size-fits-all solutions.
