The AI Integration Manifesto

Introduction

AI as we know it today (specifically, Generative AI) has turned into a polarizing topic. On one hand, it is a great piece of technology that is applicable in specific use cases. On the other, the marketing-budget-driven hype surrounding it is blowing the applicability of the tech out of proportion.

Based on my research and own experience with AI tools, I’ve come up with the following DOs and DON’Ts for integrating AI into a product or service. I call it the AIIntegrationManifesto. This is a slightly modified version of a post that I shared internally at work.

AI Integration Manifesto

DON’Ts

1. DON’T choose AI tools based on pure hype (“Everyone is doing it, so should we”)
2. DON’T choose AI tools based on fear (“If we don’t do it we’re doomed”)
3. DON’T turn AI into a hammer-and-nail situation. There are specific situations where AI shines. Find those use cases and concentrate there instead of trying to sprinkle AI everywhere.
4. DON’T turn AI into a solution in search of a problem. Start with the problem and if AI has a role to play in the solution, great. If not, that’s also fine.

DOs

1. DO prefer purposeful, focused AI tools over general purpose ones
2. DO take into account how you measure the success of AI integrations. Example: In a vacation planning app, if the user was able to better plan a vacation using AI, how do you know what AI prompts/inputs led to this improvement?.
3. DO take into account how the feedback loop for AI integrations works. Example: If the user reports that their vacation planning experience worsened as a result of the AI integration; how do you improve the prompts?
4. DO have a clear end-goal for the AI integration.
   • Does it help reduce the percentage of uninstalls or un-subscriptions
   • Does it help users save money while planning a vacation?
   • Can you quantify the impact that AI had on these goals?

Conclusion

In conclusion, remember that Generative AI has very narrow applications. Specifically, it is useless if your problem domain relies on facts, accuracy or determinism. You wouldn’t use agricultural tools for surgery, so why use “generative” tools where correctness matters?