Should AI be in your MVP v1, or added after the workflow works?

There has been a noticeable trend among development teams lately, where they are incorporating large language models (LLMs) into the initial version of their products. As the cost of using these models continues to decline, it may seem economical to include them from the outset. However, the affordability of accessing these models does not equate to their ease of integration or the complexity of ensuring their effectiveness.

The author advocates for a simple rule to guide this decision-making process: if the product's core functionality becomes significantly diminished or ineffective without the model's capabilities—such as tasks like parsing documents, matching data, or summarizing intricate inputs—the model should indeed be integrated into the first version. The primary focus, in such cases, should then shift towards rigorous evaluation and managing potential failures during the implementation of the model's output.

Conversely, if the product's primary value proposition remains intact despite a human's manual intervention at certain stages, it may be prudent to launch with a basic version and subsequently introduce the AI model once the team has a clear understanding of what constitutes correct functionality.

The decision to embed AI capabilities directly into the MVP (Minimum Viable Product) version one or to introduce them post-launch carries its own set of trade-offs. On one hand, including AI in the initial release can generate early feedback that may be tainted, as users might attribute their dissatisfaction to the product itself rather than the AI-generated output. Conversely, delaying the integration of AI could necessitate significant modifications to the user experience to accommodate the model's presence.

For applications operating in highly regulated sectors such as finance, healthcare, or any domain governed by stringent compliance requirements, the author strongly recommends adopting a cautious approach by implementing a human-in-the-loop strategy from the start. This approach ensures that there is a clear, documented record of the correct answers, which is crucial for maintaining regulatory compliance and transparency.

Ultimately, the question arises: is the AI the central hypothesis that the product is built around, or is it an added feature intended to enhance the user experience? The choice between these two approaches depends on the specific nature and objectives of the product being developed.