- Insights This article was written by a TIA community member. Insights pieces undergo the same rigorous editorial process that newsroom-produced articles have.
Marvel Gomulya · · 6 min readWhat I learned from moonlighting as our AI agent
After weeks of testing, my company, Covena, finally shipped its AI agent. We were all in high spirits that Tuesday morning: The agent’s prompts were solid, the responses were accurate, and the demo had gone cleanly.
That did not last long, however. By afternoon, we knew we had a problem.
Image credit: Timmy Loen
Our AI sales agents handle end-to-end lead conversations on WhatsApp for consumer services companies. Back then, we were on our second client deployment, working with an edtech firm whose AI agent handled inbound WhatsApp inquiries from professionals looking to enroll in online training classes.
Our AI agent was proficient at handling simple, FAQ-style questions like “What are your prices?” and “What hours are classes?” But within a few hours of deployment, we started seeing more complex queries that AI couldn’t handle.
For example: “My friend told me you guys have a package for around 2 juta (US$119.20), is that the one with the coaching package or not?” That’s one sentence with a price reference, a product question, and a clarification request stacked together.
It then dawned on us: We had built an FAQ engine and called it a sales agent, and those aren’t the same things.
To put it simply, our AI agent broke. There was no single crash, no clean error log – it simply could not navigate the ambiguity. Responses came out either confused, incomplete, or flatly unhelpful.
See also: Agents, not ads, are the next frontier for online sales
We had two choices: pull the plug and lose the client, or step in ourselves while we rebuilt.
We chose the latter. We told our client upfront, set a hard deadline for a return to normal, and treated it for what it was: a temporary stopgap.
And so my co-founder got into the system and started rewriting the architecture in real time, while I opened the backend and started answering every message myself, posing as our AI agent.
From the ground floor
That went on for days. The first few ones were brutal as I juggled five to 10 conversations at once, each at a different stage of the funnel. There were long stretches of silence followed by 15 messages in five minutes, with no way to predict when they’d arrive.
By day three, I understood viscerally why customer service agents burn out. The psychological weight of being always-on is relentless, and I was only living it for a week.
4 lessons from my side gig as an AI agent
Do tasks that don’t scale
Stay ahead in Asia’s tech landscape
This is premium content. Subscribe to read the full story.
When our AI agent broke mid-production, I was forced to step into its role. That allowed me to see the gap between automation and real work.
We know this is not ideal. ⌛ Sign up in 20 seconds. Cancel anytime.
Our subscriber community includes professionals from these companies:
Stay updated on the go with our mobile app.
Get latest insights with smoother, more personalized experience through TIA mobile app.
