Why Most AI Projects Fail
Everyone has artificial intelligence (AI) on their 2017/2018 roadmap these days. Bottom-tier innovation verticals like HR, multi-level marketing, entertainment, fashion, medical, supply chain (anyone else we should throw under the bus?) are even starting to talk about it. Everyone wants to hire a data scientist. Fortune 1500 companies are throwing out multi-million dollar data leadership positions to lead their data teams to success. Despite the hype and excitement, the majority of companies that commit to tackling AI projects will fail. Even that $1M+ hire won’t save you from failure.
Based on my own personal failures and companies I have interacted with, here are some of the main reasons why your AI project failed or will fail:
1) Science project sharks
Many of the companies we consult with are surrounded by science project sharks, who say things like, “Wouldn’t it be cool if we could do (names some niche geek-fetish)?!”; “We want all user uploaded images to align with our brand using AI!”; “This data set has value, let’s extract it!” Instead of focusing on such vague projects, the question we ask to cut through the crap is:
“Which projects have the largest impact on your revenue and KPIs?”
Wow-factor won’t grow your business or feed your family; increased revenues will. You can’t afford for your first AI project to be a failure, or that could set you back significantly behind your competition. It would be better for you to postpone jumping into AI than to fail on your first attempt. Failure will also burn political capital and excitement to pursue the next project. I have even seen senior people fired for AI failures. Don’t be that guy/girl.
Once you align your projects with BHAGs (Big Hairy Audacious Goals), KPIs (Key Performance Indicators), or Revenue you are speaking a familiar language, the language of business. Your first AI project must be business-focused. If you are successful, your executives/board will love you and they will back you with more resources for your next win. For many, this may fall into hiring, customer churn, and marketing email messaging. If the majority of your revenue comes from an unoptimized email marketing campaign you would be an idiot to start anywhere else.
2) Breakdown in communication
The majority of data scientists can’t speak dollars and cents. For being amazing data geniuses, they don’t know much when it comes to running a business. They want to bring their jargon and personal geek passions to a conversation where they don’t belong.
Four years ago, when I sent an executive at HireVue a gorgeous bootstrap plot, he said to me, “What the hell am I looking at?! We know you are smart, that is why we hired you; you don’t have to remind us.” These words were truly inspiring and impactful on my data science career since.
Your job is not to educate your executives on data science methods or complicated jargon. They are literally drowning with priorities focused on keeping the company growing, hitting revenue quarter goals, appeasing investors, and making sure you have a damn job. Their plate is overflowing, so simplify the data science discussion. Make whatever you say to-the-point and actionable. Give them a number with a dollar sign attached, and if you need assumptions to get there, do it. Speak their language, avoid all possible jargon, and align your AI projects with the highest priorities for your company.
3) Fail before you start
Following lean startup principles, what is the least amount of time or effort you can spend to find out if you are going to fail on a particular project? Is there any way to fail before you start!? Wouldn’t that be great! I was involved with a fantastic AI project using deep siamese nets only to find out two months later the customer wasn’t willing to pay for it. The AI project was dead before it started. Imagine instead pitching your best customers with a slide deck showing a successful AI project you haven’t done yet. Fake the results. Sample a user focus group and get feedback, and make sure you get full engagement and buy-in before you pull the trigger.
4) Not having a data warrior
Data science newbies are kind of like Russian Roulette. I have seen companies hire new college grads for compensation discounts or out of necessity (“Nobody good wants to work for us, sorry join the club!”). I don’t care where they are coming from (Stanford, MIT, Oxford, etc…), without real-world experience, you could be hiring a loose cannon or worse, an academic! An academic will waste your time and resources and have an unlimited number of excuses why their AI project never crystallizes into anything useful. You want someone who has worked for someone else, made mistakes and shipped AI product. You want a Scarface.
5) Homegrown talent/software
We have internal data science talent, homegrown baby! Most of the time that means you have soft/distracted talent. So you are telling me you have internal data science talent who have been sheltered from the upside-down world that exists outside of your organization’s bubble?
If they haven’t mixed with the surrounding data science community/job market [data battlefield] how are you ensuring you are using the latest and greatest? Are they 3 years behind? Probably. Can a scarface data scientist eat them for dinner? Absolutely. Even worse than homegrown talent are companies that think homegrown software is a good idea. If a third party vendor/partner can do something better/faster (i.e. DataRobot) use them! I have found that most companies that attempt to create homegrown software end up doing it for more time/money and get lower quality. Leveraging third-party software, I can outpace your entire homegrown data science team singlehandedly, no comparison. This gets back to the business objectives, and a true business-minded leader won’t mind paying for value. I prefer to have my data talent focused on the hard/custom problems that aren’t turnkey yet.
6) Start simple!
Right now you are getting 0% value for your AI project if you haven’t implemented it yet. The moment you roll out a dead simple Bayesian method you are realizing 80% of the value. Sure, a gradient boosted regression or deep network can take you to 97% but why not bag 80% value today!? Some AI projects are overcomplicated with time horizons that are too long. I always recommend focusing on a 30-60 day proof of concept. Executives also like this because it usually means you are derisking the problem by reducing the cost. I would much rather have you fail on a 30-day project than a one year Hail Mary pass that ends up being incomplete.
Ben will be speaking at the Data Science Go conference in San Diego from November 10-12 — use speaker code BEN50OFF for 50% off the conference price. He will also be in Mountain View, CA November 1-4 and in LA November 19-21, and would be happy to connect over data science or machine learning topics — feel free to reach out to schedule a conversation!