I’ve been spending quite a bit of time pondering the role of AI in modern entrepreneurship. It’s a conversation I frequently find myself having with fellow angel investors and startup founders. There’s no denying AI’s ability to transform how we operate, but I’m always left with the question: are we relying on it too much, or not enough?
I recently sat down with a founder who was beaming with excitement about an AI tool his team was using to generate product mockups. They could churn out polished designs at the snap of a finger. This kind of efficiency is exactly what is celebrated in the startup world. But his excitement was tempered by a lingering question: how do we ensure that this efficiency doesn’t come at the cost of genuine understanding?
AI tools can indeed make prototyping more accessible. They can draft copy, suggest code, and even organize notes. But as I’ve often said, the ability to produce a polished demonstration is not synonymous with understanding the customer’s problem. The danger lies in being lulled into a false sense of security by the glitzy facade AI can build.
When I first started investing in startups, I noticed that many founders were spending more time perfecting their pitch decks than understanding their market. AI, with its ability to automate certain processes, can sometimes exacerbate this issue. The ease of production means there’s even more value in choosing the right work. And the right work is understanding the problem before you present the solution.
One approach I’ve found useful is to begin with a task where the output can be verified. For instance, you might ask AI for three clearer versions of an interview invitation. Or you could have it propose categories for anonymized feedback and then compare those categories with your original notes. The founder must remain responsible for deciding what the material means. AI can assist, but it should never replace the human element of understanding and interpretation.
This leads to another crucial aspect: setting boundaries. AI requests need specificity, much like a brief to a human developer. A useful request might be: “Create a mobile screen for reviewing an order with ten items. Show product, quantity, and an unresolved-item flag. Use invented data. Do not add payments, analytics, or inventory features.” Tight scope is as valuable in an AI request as it is in a brief to a human developer.
But let’s not forget the elephant in the room: AI-generated content can sometimes be confidently wrong. OpenAI’s research describes hallucination as an ongoing problem, one that fluent output doesn’t necessarily resolve. For a founder, the practical response is to verify consequential claims and design a way to stop or escalate uncertain outputs. It’s about creating a system of checks and balances, much like you would in any other area of business.
Take Asha, for example, a founder experimenting with converting voice orders into structured lists. She should measure errors against human-checked examples, including local product names, mixed languages, and background noise. A high average accuracy can conceal mistakes in quantities that matter greatly. Begin with human approval before dispatch, and record both error frequency and correction time.
The importance of human oversight cannot be overstated. A demonstration running smoothly once on your laptop says little about its reliability in someone else’s working day. Generated code still needs checking. Test normal inputs, missing fields, duplicate submissions, and interrupted connections. Before handling real customer data, arrange a competent review of the relevant access controls and storage.
The same caution applies to customer data. Simulated personas can suggest questions, but they cannot supply evidence that a real buyer needs your product. Protect customer information: use invented data where possible, obtain necessary permission, and understand a provider’s current data terms before sharing sensitive material.
And then there’s the cost — not just financial, but in time and resources. Count the complete cost of using AI, including usage charges, checking, corrections, and exceptions. If automation saves ten minutes but creates twenty minutes of review, you have learned something useful, but you have not yet improved the business. Adopt the part that works and keep investigating the rest.
In my experience, this is not unlike the process of diagnosing a complex fertility case. We can run all the tests and gather all the data, but at the end of the day, it’s the interpretation and the human empathy that truly make the difference. Just as in investing, where due diligence feels like peeling back the layers of a puzzle, the human element remains irreplaceable.
AI is a tool with immense potential, but it must be wielded with care. The founders who will succeed are those who understand that AI’s efficiency must be balanced with a deep understanding of their customer’s needs and a robust system of checks and balances.
In the end, leveraging AI is not about replacing human effort; it’s about enhancing it. And isn’t that what we all strive for — to become more efficient, not by doing less, but by doing more of what truly matters?
As we embrace these technologies, let’s do so with our eyes wide open, aware of both the possibilities and the pitfalls. The promise of AI is real, but its success depends on how we choose to integrate it into our human-centric processes. Here’s to doing that thoughtfully and with integrity.
Part of the Building Frugal Startups: Lessons from the Trenches series — read the full guide.