Predicting things is incredibly hard, but I’ll try. Here is what I said going into 2018. I have a decent spot from which to analyze the AI market. First let me tell you what I’m learning and seeing from each of those spots and then I’ll summarize some key predictions for next year.
As CEO Of An AI Company
I run Talla, which builds AI powered products to automate work. We have an intelligent knowledge base, and a customer support automation platform that can aid reps, or end customers directly. This is the first year we’ve seen real quantifiable gains from customers who use the product the way it was intended. The results are remarkable, often shaving 40–80% of the work off of a support team, and resulting in 7 figure annual savings for organizations as small as a few hundred people.
But, getting people to buy into it is still difficult. More people than ever want AI, but, they want it to work differently than the way it does. Non-AI tools that fake AI (e.g. dumb scripted chatbots) still have a slight advantage in the market, but, I think this will be the year that changes. Buyers are getting smarter about AI, but their mean buying criteria is still “ease of implementation” rather than “how much work it automates once it’s implemented.” The ROI on AI is massive, but, if you calculate ROI as (Performance at Scale) / (Implementation headache), buyers are still overwhelmingly focused on the denominator.
The other issue I’ve seen is that companies that started all over the map (like Talla, originally an HR chatbot) have all consolidated on a few key use cases that sit at the intersection of “economically viable” and “technically feasible.” Expect to see more direct competition between companies that didn’t compete before, which could lead to some market consolidation.
As An AI Investor
A few things stuck out to me this year. First of all, there is a rise in neuromorphic hardware startups. I did a new deal in the space this year, and had been looking to do one after my first AI chip investment, Mythic, in 2016. It’s a difficult space to figure out though, because chips have long development cycles and existing players are massive. But, I have a firm believe that these chips will break an entire industry out of the x86 mindset, which will open up an explosion of innovation, and so I want to be in at the ground level. Plus, most investors are super skeptical of neuromorphic hardware, think Intel won’t let these companies get big, and that they take too much capital. So, in other words, they can be good deals because of the lack of interest. I’m bullish.
The second thing I noticed is that this year, Automation > Prediction. Early AI startups of the past few years were about predicting things. This latest wave is about automating things. Botkeeper’s massive Series A is an example of this trend, also the work UIPath and Automation Anywhere are doing moving into AI from traditional RPA. I’m seeing it in startups focused on automating away new tasks, and have a few more investments there (most recently Invisible). This is going to be a massive massive massive trend, and it was predicted by Bradford Cross in 2017 in this blog post.
The third thing I noticed is that almost all these AI companies are taking longer than SaaS companies to hit equivalent milestones, and most investors haven’t adapted yet. Part of it because now, in addition to an MVP and finding product market fit, Jocelyn Goldfein from Zetta has shown (brilliantly, I think) you have to worry about Minimal Algorithmic Performance as well. It’s a great insight and is very true. I see smart investors reserving more capital for follow ons than they traditionally would, and doing more extensions on early rounds than they would tolerate for SaaS. I think it will pay off.
And finally, buyers are just starting to come around. Almost all of these AI startups are sub $2M in ARR, but, the pace is picking up, in part because buyer’s are finally moving beyond pilots in some cases. I’m not sure 2019 is the year AI adoption explodes (probably 2020) but it will pick up the pace. I see it across my portfolio.
As A Writer and Podcaster in AI
As the Sunday writer for the InsideAI newsletter, and the host of AI at Work podcast, I see lots of news and talk to lots of people who are thinking about AI. The two biggest themes to emerge this year, in articles and in the podcast, are concerns about bias in AI algorithms, and the idea that fast following doesn’t work. Tom Davenport’s HBR article on the latter was particularly interesting.
So what does this mean for 2019? Here are my predictions:
- An AI hardware company will breakout, driving a plethora of new AI hardware startups and investment.
- Automation, predictive analytics, machine vision, and chatbot areas of AI will start to consolidate, as you have lots of companies with a bit of traction but not enough to raise more funding. They will get rolled up into other startups and some big companies.
- Two new jobs will grow. First — “trainers” or “data annotator” have been a small thing for a few years. They will become a big thing. Secondly, you will start to hear about “knowledge mechanics.” These are people who don’t do a process but understand how to fix it when a machine screws it up. Think of a washing machine. We don’t wash clothes by hand anymore, and most of us don’t know how a washing machine works. But we have people who design and fix washing machines. These knowledge mechanics will design and fix applied AI processes in a similar way.
- GANs will start to show up in applications. We haven’t seen this yet, but the tech is 4 years old. It’s about time.
- Google will come out with something new that combines deep learning and symbolic logic processing. I’ve heard Peter Norvig speak on something like this, so I assume Google is working on it, and 2018 was a little bit of a “deep learning is getting limited” year in some areas. Google will drive some new innovation here and once they do, everyone else will jump into it too.
That’s my take on 2018/2019. Hope you enjoyed it, and please leave comments if you have other insights or disagree with any key points.