The CEO/founder guide: should you do something AI/ML related?

Most of my career has been in building data & AI/ML products. More recently, I’ve used Machine Learning to build the technology for marketing attribution, A solution acknowledged by industry analysts as top technology platform.

In addition, I’ve had the opportunity to meet and advise a lot of Startup founders and CEOs, around data/analytics/product, and it’s clear that AI is a very hot topic.

The typical Startup pitch includes something around data or using machine learning and artificial intelligence to “disrupt the industry”. But in reality, when you talk to CEOs and founders, most of them don’t know if/when/how to start.

Since you are CEO/founder and busy, I’ll first TLDR. Here are the phases of getting started with AI:

Step 1: Should we even do it

  • What’s the use case, for real. Not investor BS but real value
  • Do we have enough data (rule of thumb- less than tens of thousands of datapoints won’t be good enough)
  • What is the designed business outcome. Use as “so what” framework when thinking about using AI (start by describing the use of AI and drill down 3 layers deep by asking “so what” each time

Step 2: Are we ready to handle the trust issue?

  • People hate black boxes. People don’t blindly trust algorithms. If you go the AI route you need to be prepared to:
  • Easily show that “it just works”
  • Have a good, simple way to explain the “under the hood”

Step 3: How should we do it

  • At a minimum, you need 4–5 skills (usually covered by 2–3 people in the early days): data science, analyst, product/business mindset, engineering, potentially data architect
  • Manual first- try to have humans manually perform all or some of the tasks that the machine will do later. That allows you to test the value and prove it

Step 4: How to get started

  • No 1-year-projects. Create a project plan and have actual deliverables that internal/external customers can use
  • AI projects should have these components:
  • Goal- what is the question that we’re trying to answer
  • Data- what is the required data, how do we gather it, how do we prepare it for the machine
  • Algorithm- what is the way we’re going to answer the question
  • Output- what is the way to surface the results (UI? Report? Backend service)
  • Timing- should we start right away, should we wait until we have the right people, enough data, enough customers etc

Now, to the details:

Step 1- should we do it?

Like any hot buzzword, “Using AI to disrupt an industry” attracts investors, prospects, employees and attention. Sometimes this is a good-enough reason to do it.

More times than not, that is not a good reason to invest in AI/ML. You should try to articulate the use case first.

Let’s take an imaginary startup that helps consumers find interesting events when travelling abroad. The story might be “we use AI to find the most relevant events based on other consumers like you in that location

What is the use case? BEWARE: if you are using the word “cool” when describing the use case it’s probably not a strong enough use case (for example: “wouldn’t it be cool if we find similar users and look at what they like” or “wouldn’t it be cool if we build a recommendation engine for events based on past events”).

So without using the word cool, the use case might be: user opens our app and gets relevant events. Is there another way to accomplish the same thing? Sure, we can ask users which events they like and we can curate list of events per location per user or (if there are many users) per user type.

Enough data: in my experience it requires hundreds of thousands if not millions of data points to achieve satisfactory business-valid goals. As a rule of thumb I’m placing it in the tens of thousands range

“So what framework”: this framework helps you understand the real value of what you’re trying to achieve. If you see that as you go down the waterfall it becomes harder for you to rationalize then it’s a good indicator to NOT to use AI. If the bottom line at the end is value that can be easily achieved in another way, probably should NOT use AI:

  • Dos: think about the use case, be honest, audit your data assets, use “so what”
  • Don’ts: don’t use the word cool, fall into your own trap, ignore the data

Step 2- Are we ready to handle the trust issue?

“Well I don’t care what the data says, I’ve been doing this for years and your data is wrong”: this is almost an exact quote from one of our clients a few years ago. People tend to have a view of the world and it’s hard to change their mind.

People will always question AI and machines (and sometimes rightfully so) unless they see that it works (for example: if every netflix recommended movie was spot on and you loved it) or they believe the underlying mechanics that you tell them (for example: “we use automated a/b testing to look at two versions of this creative and how many people respond to each”).

AI and ML will never work if you don’t trust the results. Understanding how you can get there is fundamental to starting any project.

  • Dos: think about trust as a key factor, plan the “why should people believe it” part
  • Don’ts: assume that people get it, assume that machines are better

Step 3- How should we do it

Fail fast is almost always true. If you can examine the assumptions that are at the core of wanting to start with AI (for example the value to customer) early before heavily investing in infrastructure and scale you can save millions of Dollars.

For example: you want content recommendation system for customers? Assuming people will be more engaged? Can you curate recommendations for 50 customers and check their engagement levels? In a sense, this is what Facebook did with messenger M.

You also need the right team. Data scientists are not one-stop-shop. The need to get the right data, understand the business outputs, scale the solution, deliver on time etc. implies that you need much more broad skillset. Make sure you have a small, well-rounded team.

  • Dos: make sure you have the right team and skill-set
  • Don’ts: build before you test manually, hire data scientists without complementing them with different skill-sets

Step 4- How to get started

OK! So now we know that we definitely want and can start with AI/ML. But how to actually get started? When to get started? How to plan?

No 1-year-projects: Create a project plan and have actual deliverables that internal/external customers can use. AI is very non-deterministic in nature, research can take time and outcomes can be surprising. If you don’t break it down to tasks that allow early feedback by others you will end up not delivering at all.

Success of AI projects depends on a lot of factors, but you have to understand the structure of “what is a AI/ML project” in order to make sure you get clarity from the team on each of these aspects:

  • Goal- what is the question that we’re trying to answer. This is the most important part. Without defining the goal there will be no product. For example: trying to find content that people like could be completely different from trying to find the next piece of content that they would want to consume
  • Data- what is the required data, how do we gather it, how do we prepare it for the machine. You can never “just have all the data in one place”. That will usually cause a fortune, not to mention that usually you don’t have all the data and you need to get it from somewhere. Designing the data process is the foundation of the algorithm
  • Algorithm- what is the way we’re going to answer the question. This will most likely change over time but the way to talk about this topic is to understand from the data scientists why this is the best approach to answer the business question (achieve the goal)
  • Output- what is the way to surface the results (UI? Report? Backend service). If you don’t design the way to use the results you will most likely fail and be left with the algorithm with no way to actually drive value to customers

Timing: should we start right away, should we wait until we have the right people, enough data, enough customers etc. If you start too early you are wasting time and money (it’s like building a racecar when people just want to drive 2 min to the grocery store), if you start too late you’re behind and might have already been disrupted by another company.

  • Dos: Be skeptic, focus on the core value to the customer
  • Don’ts: be tempted to build a ferrari day 1, run after “shiny objects” (like only try to use latest development in AI)

If you have additional steps, or you want me to drill down more just comment. If you liked this guide and find it valuable- share with others.