It's Tuesday morning in Johannesburg, the kind where the highveld sun is already sharp by 9 AM, and you are deep into an interview with a senior engineer. Nomusa has a decade of experience, the last five at a major fintech firm, and she's exactly the kind of talent a startup like yours, Asenda, needs. She’s not just looking for a job; she’s assessing whether your AI product deserves three years of her best work. She doesn't ask about salary or benefits yet, not directly. Instead, she asks: "What's the riskiest assumption baked into your AI's core logic, and how often has it broken in production during your pilots?" This question reveals that a candidate isn't just looking for a role; they are evaluating the product's foundations and the credibility of its builders, seeking evidence that the work ahead is genuinely impactful and well-conceived.
The Engineer's Real Interview
Nomusa wasn’t just interviewing for a role. She was interviewing the product, its underlying assumptions, and the founder's capacity to navigate real-world challenges. Her question cut past the pitch deck’s promises, probing for the kind of practical insight only a founder who has truly shipped product can offer. It's a signal that she understands the difference between a working demo and a resilient system, between an idea and a product that faces users.
This isn't an isolated incident. Across the African continent, in Lagos, Nairobi, or here in Joburg, top-tier engineering talent is increasingly scarce and discerning. They’ve seen enough "innovative solutions" that never ship, enough grand visions without a concrete plan for production stability. They want to know what breaks, how often, and what the team actually learned from it. This practical interrogation of the product is crucial for early-stage founders, particularly those building AI or SaaS products on limited runway. It forces a founder to articulate not just what works, but what risks are still active, and how those risks are managed. This kind of honesty builds trust, which is a powerful asset in attracting talent.
Learning from Production Breakages
My own journey building AI products with Asenda Ltd across Africa, Germany, and the US has been a masterclass in uncovering risky assumptions. One incident stands out: an AI agent we developed for a logistics client in Ghana consistently misclassified certain delivery exceptions during early pilots. Our initial models performed well in sandboxed environments, but real-world data, with its unpredictable noise and edge cases, exposed a critical flaw. The agent, designed to categorize delays, often conflated "customer not available" with "incorrect address," leading to misrouted support tickets and delayed resolutions.
The team could have dismissed this as a minor bug, something to fix with more data. But it was a deeper problem with the underlying inference logic. We had assumed a clearer separation between these exception types than existed in practice. It was an assumption that held up in a controlled environment but crumbled when faced with the raw, messy reality of daily operations.
Our response was not to throw more data at it, but to fundamentally rethink the feature. We redesigned the agent to prompt for additional, human-validated context in ambiguous cases, effectively creating a feedback loop that refined its understanding over time. This meant a slight increase in human intervention initially, but it vastly improved the accuracy and reliability of the overall system. It also meant a frank conversation with the client about the limitations we discovered and the iterative path to improvement.
The Ask and the Trust
Nomusa's question on that Tuesday morning in Johannesburg wasn't just a technical probe; it was a character test. She wanted to know if the founder would dance around the truth, offer platitudes, or, like the Coca-Cola executives in 1985, acknowledge a misstep and articulate a plan for recovery.
In 1985, The Coca-Cola Company launched "New Coke," a reformulated version of its flagship beverage, after extensive blind taste tests indicated it was preferred over both original Coke and Pepsi. However, the launch was met with widespread public backlash, demonstrating a profound disconnect between taste preference in a controlled environment and the deep emotional and cultural attachment consumers had to the original product. Just 79 days after its introduction, Coca-Cola announced the return of "Coca-Cola Classic," acknowledging its mistake and responding to overwhelming customer sentiment. This reversal, extensively documented by publications like the New York Times at the time, highlighted the critical importance of understanding not just feature performance but also the underlying, often emotional, assumptions about product value and user connection.
Like the Coca-Cola executives realizing that taste alone didn't define their product, I knew Nomusa wasn't just assessing the technical 'taste' of our AI. She was looking for evidence that we understood the broader context: the messy realities of deployment, the inevitability of unexpected challenges, and the founder's willingness to confront them head-on. My answer, detailing that specific misclassification issue in Ghana and our iterative solution, allowed the conversation to shift from hypothetical performance to demonstrated resilience. It showed that we didn't just build, but we learned, adapted, and were honest about the journey. That transparency is what ultimately convinced her.
Comments
No comments yet.