What I Wish I Knew a Year Ago About Starting an AI Company

Prolego is my second attempt. I hope you learn from my journey.

Building a company feels a lot like this. Credit: Patrick Hofer

Last week Russ Rands and I launched our boutique AI services company, Prolego. Although we’re at the beginning, getting here has been quite a journey.

What follows is my story, I hope you learn from my successes and failures.

ScribbleIQ — AI for Content Marketing

In the Summer of 2016 I wrapped-up an engagement with the awesome MadKudu team. Working with the Kudus was a blast but I had the itch to start another product company. Machine Learning is been a lifelong passion and the technology is finally mature enough to build practical applications.

Timing was good because my friend Stephen Tse — one of the few people I want as a co-founder — was also interested in machine learning startup opportunities. After some exploration we decided to launch ScribbleIQ, AI for content marketing.

A big problem. A bad domain for an AI startup

I wanted to solve the content marketing challenges I experienced. Content is expensive and difficult to measure; I wanted to use AI help content marketers identify the best writing topics to generate sales.

At least we had a cool mascot, Mika.

A few weeks of customer discovery interviews lead to some early sales. That’s when the problems started.

• Although topic selection was a problem for many marketers, the bigger challenge was execution; quality writers are hard to find.
• My original idea was to train the machine learning algorithms using sales conversion metrics as success outputs: leads, trials, transactions. Unfortunately the data was just too noisy and fragmented to get useful results.

We realized ScribbleIQ would only solve immediate customer problems if we packaged it with a content writing service. This didn’t interest us.

We spent a few months exploring other ideas in financial services and fraud but nothing excited us. Sometimes the best outcome is a quick end, especially among friends, so we decided to go our separate ways.

(Stephen is now working on a state-of-the-art AI productivity voice assistant. )

AI lessons learned from ScribbleIQ

Now for my biggest lesson: building an AI solution isn’t like building workflow software.

Software solutions for workflow problems are now easy to build. In fact, many startups focus on distribution first and just solve whatever customer problem they discover. This approach doesn’t work with AI products because solutions are highly dependent on quality training data.

My biggest mistake was failing to realize how dependent AI is on data. The sell-and-then-figure-it-out approach didn’t work.

Rebooting

Although ScribbleIQ failed I was more excited than ever about AI. I wanted to explore custom deep learning solutions for large enterprises rather than building my own AI products.

2 months or 2 years too early?

I knew I was ahead of the market. Although AI was getting a lot of business hype very few companies were investigating deep learning technology.

The upside to being early is having time to build expertise and establish oneself as a thought leader.

The downside? Being TOO early. My top priority was deciding if the customers needed help bringing deep learning to the enterprise.

My goals in April, 2017

So I set three 6-month goals for myself:

1. Build my hands-on expertise in AI.
2. Talk to 50 qualified enterprise clients.
3. Establish myself as a thought leader in enterprise AI.

Screenshot of the goals I created in my iPhone. I hit my goals a month early.

Here is the actual screenshot from the goals I created 5 months ago in Things.

This quickly became 3 full-time jobs.

Goal 1 — Building technical expertise in AI

Building hands-on technical skills is extremely time consuming. This type of investment carries a high opportunity cost because every moment I’m coding is one I’m not talking to customers.

Unfortunately I didn’t see a different option. AI is just too hard to learn and we have few case studies.

I dove into the books, course, and papers. As I described previously, fast.ai was the best resource. I plan to continue this investment because AI is changing so fast.

Goal 2 — Talk to early prospects

I met Russ Rands through a mutual friend and discovered we were exploring similar opportunities. Russ is also a startup veteran and knows the necessity of market validation.

We teamed up and started doing early customer discovery together. How many companies did we talk to? A lot more than my goal of 50.

At first we asked questions about existing AI initiatives but soon began giving briefs about the trends we were seeing. Their questions and concerns gave me writing ideas.

The strongest interest was from companies looking to build new revenue streams or defend existing ones with AI.

Goal 3 — Establish myself as an enterprise AI thought leader

99% of AI content isn’t helpful for a business audience: too theoretical, too fluffy, or just plain nonsense. Very little answer practical questions like:

1. What is AI?
2. What can I do with it?
3. How do I get started?

These were the questions product managers, corporate innovation teams, and CEOs were asking us. So I started writing about product patterns, talent development, and getting projects funded.

Interest trickled in but we needed a way to accelerate everything. Russ suggested writing a book, so we started writing AI for Business Leaders.

Connect with Russ on LinkedIn to get your copy

Since 90% of writing an ebook is distribution we started promoting on LinkedIn on day 1. Anyone who wants a copy must comment on LinkedIn and ask for it or connect with Russ on LinkedIn. The comments gave us social proof and hiring ranking in LinkedIn.

We turned our LinkedIn profiles into landing pages and Russ also spends hours per day doing direct outreach. We started distributing early drafts as fast as I could type.

Just in case you missed that … and think writing a book sounds like a good idea … promotion. takes. hours. every. day.

But it worked.

The inbound interest picked up. More companies wanted advice or to discuss specific projects.

But was it time? Was the market really ready for a boutique AI service? Or were we just having more conversations?

Prolego is born — why we decided to take the leap

The book also helped us get invited to an NVIDIA-hosted event in Charlotte for financial services companies.

I can’t say enough good things about NVIDIA. You already know they create the best GPUs on the market, but they also make a big contribution to supporting the AI community through eduction and framework support.

The presentation was excellent and informative. The attendees were well-vetted and we met some great people.

If you get an opportunity to attend an NVIDIA-hosted event — do it.

A rare combination of smart, fun, nice people — thanks team NVIDIA!

The event, conversations, and subsequent meetings was our trigger. We realized an entire ecosystem is being created to bring deep learning to the enterprise — and we wanted to a part of it.

So last week we took the leap and decided to launch Prolego, a boutique AI consulting company to help Fortune 1000 companies bring AI to the enterprise.

So … what’s next?

I’d love to tell you we have all of the answers, but that would be a lie. We have enough interest to take the next step but still struggle like any startup to define our offer and match market needs.

Regardless of how it turns out I’ll continue to share what we learn.