6 Ingredients of AI Product Strategy

Hello Product Folks!

Creating and selling AI products is tough. The tech keeps changing, and companies have to stay alert. The rules for making successful products are still being figured out, so outcomes can be uncertain.

I was lucky to get involved with many AI products and discuss them with many experts in the space. From these experiences, I’ve identified some key points to consider for success.

Focusing on these 6 AI ingredients has helped me align our products for long-term success.

Let’s dive in…

My Favorite Finds

👉️OpenAI announced a new scale to track AI progress. But no AGI?
👉️ Save hours per week in PRD creation with our Free AI PDR (beta)
👉️ Sequoia, Andreessen disagree on Chip Shortage
👉️ Silo AI helps enterprises build custom AI models and products
👉️ Writer lets you upload up to 10M words to a chatbot & then search that info.
👉️ Amazon App Studio helps build enterprise apps from text prompts.

Past Issues:
👉️ How to lead effectively when it's not your call (Link)
👉️ Why AI Agent is the next frontier for the AI field (Link)
👉️ Craft Your OKR Story to Drive Team Execution (Link)

1- Not all AI/ML products are born equal

Let’s start by saying that ML is simply a tool; it all comes down to understanding if it makes sense to use it for the user problem you’re trying to solve - more reference my previous article.

Now let’s look at the different systemic risks, you should be aware of.

Let’s go from A to C.

A. ML-Powered Features Solutions

In the low-low value, ML enables a critical feature, but the product works even without it, like Netflix recommendations. So the risk is mitigated, company-wide.

For these products:

• They provide a friendly environment for deploying ML models.

• The product itself often generates the training data needed to build models.

This means:

• Gradual roll-outs and iterations are possible without compromise

• ML can grow the product’s value and defensibility, but the product's fundamentals should be strong even without data science.

B. “Generic Black Box Products” with ML Models at the Core

For the next up, these are products where the ML model is the core, and customers decide how to use it, like transcription services from Microsoft Azure or Google Speech-to-Text.

For products in this category:

• Be mindful of different inputs customers might send. Your model is trained on specific data, so edge cases can cause issues.

• Google faced this with their e-commerce categorization API. Their systems were trained for specific product names, but customers sent vague inputs like "mobile phone" or some other unrelated terms So some weird stuff can happen.

To handle this:

• Guide customers on correctly using your product through documentation and training.

• Reject incompatible inputs by building a detection layer.

So this means, you need to build a product that handles all edge cases, though this can be complex and have diminishing returns at some point.

C. End-to-End Solutions with ML Models at the Core

These products have ML at their heart and can't function without it, like self-driving cars.

For these products:

• You control how the product is used, avoiding the pitfalls of black box products.

• You can capture more value and expand your market to less savvy customers.

However:

• You need to deeply understand customer needs and engineer the right experience end-to-end deeply.

To conclude, your development team may need to build hardware, software, data layers, and customizable components. These challenges are high but your moat will be huge if you pull this through.

2. How does your product impact your customers’ workflow?

a. Does your product replace human effort?

Replacing humans entirely is tough because people handle edge cases and nuances that models struggle with. Your model will be compared to human performance, and it might fall short.

Plus, the cost-saving aspect isn't as appealing as you might think. Customers won't cut jobs easily unless they are sure your solution is long-term and reliable. Even then, your revenue will often be seen as just cost savings, not value added.

This strategy works best if it allows budget-limited organizations to scale tasks otherwise done on a small scale.

b. Does your product speed up human tasks?

Enabling humans to work faster is usually better - Like Microsoft co-pilot.

This way, your model can handle common decisions, and humans can step in for edge cases.

However, your value will still be seen in terms of cost savings, and your product measured by human labor standards.

c. Does your product solve previously unsolvable problems?

This includes products that:

- Solve problems humans couldn't do (e.g., AI drug discovery)

- Free up human effort where desirable (e.g., self-driving cars)

- Make decisions faster than humans (e.g., real-time code estimation)

These are more unique problems, making them easier to capture value from.

3. What is your defensible unique selling proposition (USP)?

a. Access to data science talent

Selling to traditional industries with slow tech adoption can bring early rewards but isn’t very defensible. 

If your product becomes too critical or competitors enter the market, your position weakens.

This strategy isn’t for fast growth, so don’t be misled by early wins.

b. Unique datasets

Unique datasets can make your models stronger than competitors'. But:

- The best protection is legal access to unique data for a long time.

- If creating training data is what keeps competitors away, they will catch up if the market is lucrative.

- If your advantage is being first or having a large customer base, remember that data network effects aren’t always a strong moat.

c. Innovative model architectures

Having a talented data science team can yield better results on common datasets, but this is rarely a long-term advantage.

Innovations spread quickly due to the open-source nature of the industry.

Combining all these factors — talent, datasets, and innovation — into a robust process and product makes a difference.

4. What is the role of humans in your loop?

AI products need humans to handle edge cases, build datasets, create heuristics, and measure accuracy.

How you integrate humans into your product workflow affects both costs and user experience?