At this point, you probably have a million different ideas on how you can incorporate AI into your workflow. Before you start trying to figure out exactly how to use AI, here are a few vital rules of thumb to ensure you pick initial use-cases that set you up for success.

1. Easy to verify, labour-intensive to generate:

This is the most important rule. AI is extremely powerful but comes with underlying unpredictability - and so you will need to verify the outputs. You always want to pick use- cases where the work to come up with the answer is much greater than the work to check the answer. Otherwise, you'll be creating more work for yourself than you're automating.

• Positive Example: Captions for an Instagram post: you might spend HOURS thinking of a good one, but you can instantly know if its good or bad when you read it.
• Negative Example: Intricate documentation: The work needed to verify the accuracy of every single line of text usually far exceeds time saved by using AI. Models tend to hallucinate with confidence, such that even experts overlook mistakes.

2. No Right Answer - Many Acceptable Ones

Generative AI is inherently probabilistic. This makes it naturally better suited to tasks where there is no one right answer but rather, you’re searching for a good-enough answer that is often tedious to arrive it. When learning to swim - don’t fight the tide, swim with it.

• Positive Example: Blog Title Generator: There isn’t a ‘right’ blog title, but many great ones. AI’s inherent randomness makes arriving at one easier and faster than usual.
• Negative Example: Numerical calculations: There is always a correct answer with math problems. Given AI’s probabilistic nature, you are better suited avoiding these.

3. Self-contained

The ideal initial AI use-case is one that doesn't require external information as this introduces friction in usage and requires engineering work to orchestrate. That is, pick use- cases where you don't need to integrate real time/dynamic information into the AI system.

• Positive Example: explain like I'm five - for most foundational models, you can easily use their in-built knowledge to understand most concepts quickly.
• Negative Example: Daily news summarization. This would require you to connect the AI to an external news source, which, while doable, requires engineering work.

4. Inherently Human in the loop

AI results vary drastically across use-cases. That is, while AI might be fantastic at some things, it can be quite bad at others. And the differences are not pre-determinable. As such, to start - focus on using AI as an augmentative tool vs an automation solution. AI is best used in scenarios where multiple levels of review already exist, like the first draft of a deal memo or pitch document.

5. Low stakes

This one is relatively obvious. Try not to jump into the deep end directly. Start by tackling non-critical but strenuous tasks, then graduate onto higher risk tasks over time. As you move towards higher stakes tasks, we highly recommend moving away from boilerplate generalized models and investing in specialized solutions hyper aligned to those tasks.