Since 2006, I have worked within companies that have integrated AI (or machine learning or data science) into their solutions. These companies have ranged in size from micro start-ups all the way to large multi-nationals. As to their level of success, I will let their board members and customers be the judge of that. This article is a group of things that I wish I had felt able to say to the various business leaders in those companies. I am writing this as as open letter, as it felt like the right way to do it.
You say you want AI in your product. Actually, the word want is too casual. You must have it. It is the best thing. Investors are asking why we don’t have it. We could raise more money. We have the data, we can use it better. Our competitors have it. Our customers would pay for. I know that the list of reasons could go on and on.
You are smart leaders. You admit that you don’t know everything about the technology and how to implement it. But you have hired smart people, and surely we can figure this out.
Here are a few things I’d like you to think about as we attempt this:
- We are not Google or Apple. Our team are nubes, even if you have just hired that amazing CTO or VP engineering or (insert cool techie title here). We still need to ramp up. Unless you have the stomach to lay out the cash to hire the entire vertical from one of the big guns in this business, then that’s just the way it is. Even when you hire in, we are integrating them into an existing team that don’t have the knowledge base you just acquired. I could tell you stories of long hours trying to convince people to do things differently in order to put ML into product.
- We are going to make mistakes. Sometimes the mistakes will happen before we integrate it into the product, sometimes after. And because we have an iterative product development process, v1 is just not going to be that spectacular when we finally get it working. Truth.
- Machine learning is research, our problem is not already solved. Many things will be tried that do not yield the desired results. Yes, there are tools and pre-trained models that we can leverage. But a lever provides an advantage to an exerted effort, it does not in and of itself provide the solution. Let me tell you about the frustration that I have heard at the amount of time taken to get to the first model.
- Machine learning is not the answer to every problem. Sometimes the customers’ problems can be solved with something simpler. Unnecessary complexity comes at a cost. I would recommend you think about that when you are looking at your balance sheet.
- Other product features still matter. Our customers won’t give a shit about the fancy backend if the product is difficult to use or the results are unreliable. One of the first things I learned as a PM is that it is better to have a kick-assed half product than half-assed complete product.
- While data driven machine learning can be more powerful than traditional coded solutions, it is more complex to maintain. Model development is not one and done, every model degrades. Let me tell you how that happens.
- Have you thought about pushback? Not every customer or industry is keen on having their data mined, even though you provide a DUA. What are the laws around data use that we need to think about? What about transparency/explainability? The existing data governance in the organization will need to change to this new usage. Please don’t leave these important items to your technical team.
I hope that you made it to the end of this letter. I think that you are pretty great for wanting to do things differently. Please show your staff the same respect that we are expected to show you.
Your ML PM