My AI Project is DOA – What do I do now?

Dear fellow C-person,

Let me save your time and be brief.  The main reasons for AI strategy or program failure in a large company are:

  1. Not doing your homework
    1. If you don’t know what’s the difference between ML and AI you should do some reading RIGHT NOW.  I mean it.
    2. If you don’t have a specific problem for AI to solve, then you ARE the problem.
  2. Trusting the hype
    1. No, AI cannot and will not replace all of your employees.
    2. AI can help your company grow, but it cannot magically replace your (fill in the blank).  In fact, it cannot magically do anything.  Magic does not exist.
  3. Dirty data
    1. If your data management stinks, so will your AI.
    2. Not listening to your scientists/engineers

Go talk – and more so listen – to your engineers. They are the way and the truth and the life of your company.

Now in more detail

If you are reading this, chances are your AI project is screwed up already (I mean the title is, “My AI Project is DOA…”).  But fear not!  There are no unfixable mistakes in AI or ML adoption[1].  What’s ML, you’re asking?  ML is Machine Learning.  You see, AI comes in two main flavors: the kind that learns and the kind that doesn’t.  You want the kind that learns – that’s ML.

Which brings us to the first point – do some reading.  I recommend The Artificial Intelligence Imperative by Anastassia Lauterbach and Andrea Bonime-Blanc. It’s a great entry-level book for decision-makers.  Another worthwhile read is Prediction Machines: The Simple Economics of Artificial Intelligence by Ajay Agrawal, Joshua Gans, and Avi Goldfarb.  You can also read our whitepaper, What is AI? Tips on How to Implement it in Your Organization if you don’t mind giving us your name and email address.

The second point is – don’t trust the hype.  If you think that an AI agent can learn like a person but faster, think again. That kind of AI only exists in Hollywood movies or Super Bowl commercials.  AI agents don’t learn as humans do.  They don’t have a few billion years of evolution behind them, and they don’t take 20+ years of real-world learning to start an entry-level job.  Rather, for an AI agent or ML algorithm to learn anything useful, it needs data: data that is clean, well kept and follows a certain standard (or “schema”, in database lingo).  Also, while AI is great at learning from data and can surpass a human on any specific, well-defined, data-based task, the price the AI pays for that is that it completely, entirely lacks the real-world context of the task and any human cultural conventions and connotations associated with it.  Simply put, present-day AI has no notion of common sense; and (despite the claims to the contrary) the socially-,  culturally-, and situationally-aware real-world AI agents do not yet exist.  On the up-side, that means AI won’t drive in a thunderstorm to buy a lottery ticket and a pack of cigarettes or get a badly misspelled tattoo.  On the down-side, that means that (a) it needs data to learn, and (b) its training data is its entire world.

It takes a lot of skills and experience to make sure an AI agent or ML algorithm would generalize gracefully or respond in a non-absurd fashion to data it hasn’t seen before.  And it takes even more skills and experience to teach AI to do what you wanted it to do rather than what you told it to do.  Most of all, this emphasizes my third point – the importance of clean data. It is the data janitorial services that are crucial for your success.

But before we get to data needs and to AI training, let’s step back and ask ourselves, what do we need AI for?  Be as specific as possible.  This is not a test.

Here are some typical honest answers:

  1. Because I need to detect anomalies in my process.
  2. Because I watched a commercial that told me my business needs it.
  3. Because my advertising ROI (or some other KPI) leaves a lot to be desired.
  4. Because my Senior VP of something-something thinks it’s the best thing since the guillotine.
  5. Because I need some actionable insights from my sales data.
  6. Because I tried everything else and it’s not working.

These are all valid answers.  Remember, this is not a test.  However, if you answered 2, 4, or 6, you have a longer way to go than if you answered 1, 3, or 5.

Here’s how you can get off to a much better start (or re-start).

Key to success is the understanding that AI is not a goal, it’s a tool.  More specifically, it’s a tool for making sense of your data (or your customers’ data) and for acting on it.  The three elements here are “data”, “making sense”, and “actions”.  “Data” is the first element.  It may come in a variety of forms: sales data, sensor readings, video, customer calls, and so on.  Without it, AI can do nothing.  “Making sense” is the second element. What kind of questions do you want the AI to answer?  Do you have well-defined questions?  Do you have enough data for the AI to work with, so that those questions can be answered with any certainty?  “Actions” is the third element.  What do you want the end result to be?  Do you envision AI acting on its own?  Do you envision AI augmenting your personnel?  Do you envision AI generating recommendations or alerts for you or your personnel?  Do you envision AI optimizing your processes?  Do you envision embodied or cloud-based AI components or functionality in your products or services?  Do you envision AI as a component of your IT security suite?  Do you envision AI as a vehicle to make your company more attractive to investors and/or consumers?

These are all valid and doable, but these all require different approaches to AI, different ML techniques, and vastly different amounts of data, time, and effort.  As long as you have a clear answer to what it is that you want to achieve — data, questions, and actions — you are already a long way ahead towards successfully launching your company’s AI initiative.  Hopefully, you are starting to see why off-the-shelf AI programs are usually NOT the answer you are looking for.

A second key is to start small. Instead of trying to boil the world’s oceans (why would you ever want to do that — do you hate fish that badly?), find low-hanging fruit; a relatively straightforward project where you know that

  1. the data is available
  2. the questions are answerable by the team with a certain degree of success
  3. the goals are well defined
  4. the results can be improved (i.e., there is an actual pain-point that can be addressed by the AI).

Now back to point number four – ask your most experienced employees (the scientists and/or engineers you trust, not the yes-men) whether this AI initiative is

  1. doable in principle (reality check)
  2. worth doing (a back-of-the-envelope risk/benefit analysis)
  3. can be done in-house within the time and budget limitations, and/or
  4. can be done by AIaaS experts within the time and budget limitations
  5. has a clear path of growth within the company (as opposed to a one-time effort or an effort that is outside your company’s main focus).

Trust me, this would make a world of difference for you and your company.

Don’t have the team of experienced data scientists and engineers you need in place to define and execute your program? It’s okay.  Most companies don’t.  Remember, there are only about 10K qualified data scientists on the planet.  We can help.

[1]                There is, however, Chapter 7; but then, again, you can always try again, and succeed.

About the Author

Dimitry Fisher is a Chief AI Officer at Analytics Ventures / Dynam AI.  Dimitry has 20+ years of R&D experience in academia, government, and private sectors, spanning multiple areas of physics, neuroscience, ML and AI.  Dimitry received his PhD in Physics from Weizmann Institute in 2002. He did independent research in physics before shifting his focus to neuroscience and to mechanisms of learning in neural networks.  Dimitry joined Brain Corporation in 2012, doing bleeding-edge R&D for Qualcomm, DARPA, and several internal projects, including the first self-driving janitorial AI.  Dimitry saw Brain Corporation grow from 10 people to 100 before resigning his position to join Analytics Ventures as the Chief AI Officer.  He has 12 granted US patents and over 40 publications in scientific journals.