What Actually Works with AI: Why Most Founders Get It Wrong (From a Seasoned AI Engineer)

Building an app used to take months of code, thousands in hosting, and at least one engineer who knew what they were doing. At Supermoon's second OpenClaw for Startups workshop, two builders showed a room full of founders how that timeline is collapsing, and where it still isn't.

Eric Manganaro led a hands-on session at WeWork in New York. Eric covered OpenClaw installation and basics, then opened the floor. The audience drove the rest. Founders asked about their own use cases, their own broken workflows, their own startup problems, and the two walked them through it live. Not a slide deck. A working session.

How To Start Leveraging Your Agents.

The first thing Eric, made clear: there is a difference between asking AI a question and using it to build something that runs on its own. Most founders treat AI like a search engine. OpenClaw works differently. You give it a natural language instruction and it creates a hosted application, running on a server, handling real tasks without you at your computer.

"You can also ask ChatGPT what is the weather," Eric said. "But that is not an app. You cannot expose it to others. In this way, what you did is you created an app."

Throughout the workshop, Eric demonstrated this live as audience members pitched scenarios. Automations, scheduled tasks, multi-channel messaging. Each time, the tool translated plain English into something functional in seconds. The pattern was consistent: what would normally require writing code, configuring a server, and connecting APIs took one sentence.

Ship Your MVP Before You Burn Out

The biggest reaction came when Eric talked about product-market fit timelines.

"That used to take six months. By the time you get there, you just lost your time, lost bunch of money, and you basically burn out."

His argument: OpenClaw compresses the early validation phase. Build, ship, get feedback, iterate. The loop that used to cost six figures now happens in days. But he added a qualifier that most AI evangelists skip.

"People are sometimes saying that software engineering is not required. That's not true. Software engineering is at most required. And domain expertise is also going to be required."

The point was not that AI replaces engineers. It compresses the window before you know if anyone wants what you are building.

Why One-Prompt Tools Fall Apart

Most founders burn money and AI tokens without any ROI, Eric addressed this frustration that came up repeatedly from the audience: why does AI-generated output feel so generic? His answer pointed to how most people use these tools. They ask for everything at once.

"Make me a whole website at one shot, and it'll just make this long column of trash. It is a website. It is there. But why am I scrolling for so long?"

His approach: work iteratively using atomic design. Build small components, then combine them into larger structures. The difference is that OpenClaw learns your preferences over time. It remembers your brand constraints, your design patterns, your taste. A single prompt to a generic tool does not know you. OpenClaw, used correctly, starts to.

The Honest Limits

The audience pushed hard on security, web scraping, and where these tools hit a wall. Eric pointed to Trusted Execution Environments as one emerging solution, but was honest: privacy layers on open frameworks are still early. Eric did not oversell either.

"If you don't own a process, you can't own it. If it doesn't have an API, it does not want you to use it."

The workshop closed on a point that tied the evening together. OpenClaw is not a silver bullet. Google is getting better at detecting AI-generated content. You still need humans with domain expertise and taste. The tool makes the iteration faster. That is the value.

About the Workshop

OpenClaw for Startups was held April 15, 2026, at WeWork in New York. Led by Eric Manganaro, with Supermoon's Joanna Orlova, Christopher Micheal and Elena Obukhova, as well as Khurram Kalimi of Vinncorp, co-hosting. Second edition of the series, following strong demand after the first session.

‍

‍