Democratizing Machine Learning to Predict the Future with Richard Harris of Black Crow AI

Listen on
June 2, 2022

On today’s pod I chat with Richard Harris, the Founder & CEO of Black Crow AI.

Black Crow is a no-code, real-time machine-learning based predictive software that helps companies understand likely customer behavior.

Richard’s a veteran entrepreneur and has been involved in tech since the 90s. He cut his teeth in the world of consulting, was involved with Travelocity during the dot com boom, then has continued to be a serial founder.

As you’ll hear him explain, the world is turning into a browser. Between mobile devices, computers, wearable tech, self driving cars - real time data will be streaming from every part of our lives. He’s using this insight to build a company to help startups and brands collect and understand their first party data so that they can increase revenue and margins.

In addition to discussing how Black Crow operates and the machine learning industry, we also talk about some strategies for how founders can navigate down market cycles - like the one we’re entering now. He’s been through three of these cycles and has some very helpful wisdom. If you started your company after 2010, then I definitely recommend that you give this one a listen.

Show Links:

Transcript (this is an automated transcript):

MPD: All right, let's jump in. Can you give us an overview of Black Crow?

Richard Harris: Sure. Black Crow is a real time machining learn machine learning platform. And the key thing we do is we ingest massive amounts of data, process it in real time and then produce predictions. And the key area where we work today is any e-commerce where in real time, as users are shopping in e-commerce environment, we predict their future value 15 milliseconds after anyone does anything inside of a brand's environment.

MPD: How does that work? That's that sounds too good to be true.

Richard Harris: It does. It does. And it almost is, except that it's not. So what we're doing, we've made, the core of what we do is a very sophisticated, real time auto ML platform. And it was designed to be able to as I mentioned, ingest massive amounts of streaming data.

So that's literally the event data that's kicked off by a user, interacting with a website, with a browser. And historically that has not been a widely solved problem, streaming data processing it and generating predictions on top of it has not been a widely solved problem. Massive companies like Google do stuff like this.

Like Amazon does stuff like this, but our particular mission is about making this accessible to the middle of the market. People who have historically not have. Not had access to fortune 500 grade machine learning. But the key thing is in order to get after that middle of the market, we needed to make it super simple.

So rather than being a multi-year multimillion dollar project, which is how it works at the fortune 500 level, it's literally one click to install. And what our machine does is it just listens to all of that user interaction, data, all of those real-time events and based on whatever the objective function is, meaning what is the key thing that this brand or store e-commerce company is trying to have happen?

Whether that's a subscription or a purchase, a repeat purchase, the machine starts finding the patterns in that without any work, frankly, on our side or on the customer side. And in about two weeks, we will have. Accurate predictive model built so that every time someone does something, so a hundred percent of visitors to a commerce environment of the machine can say, how likely is this person to do the thing you want them to do?

MPD: So where's the signal and all those. Cause I think about the stream of data flowing through, and I'm sure it varies by use case and by company and by customer. But I would just imagine a lot of it's noise. Is there an area where, Hey, people who click on a certain area or a certain amount of time, is there something that's usually like a higher probability prep, more probabilistic signal in the data stream?

Richard Harris: Yeah, it's interesting. So we obviously look at this all the time and the answer is nothing that is consistent. Now. In the smallest possible sense, meaning there's no, there's nothing I could narrate to you in a human understandable way, or I could narrate anything, not understandable by humans, but when you try to find the story and the predictions, it's often not there.

I can tell you though that, the way we work, the machine is listening to about 450 signals in real time. And it's trying to understand what, if any of those signals are predictive or I can create predictive knowledge and I'll say that. For every brand that we work with and we have our models running on about two hundred and forty two hundred fifty different commerce brands right now.

There's none of them are the same. So each model is built in a bespoke way for each brand. And what the machine is trying to do is figure out which of those 450 signals, they're after each user action is relevant for this brand and then the weighting of each variable and trying to predict what their future behavior and value will be.

But I'd say that the interesting thing is th there's, even though there's no human narrative story, what's so interesting is that this data that brands have that commerce companies have is data they own. And historically, because it's so hard to process time streaming. Historically, they haven't really thought of it as an asset that they can exploit.

And that's what we're trying to do is really using machine learning, which is the only way to understand when there's so much data using machine learning. How do we turn this thing that they own into a huge asset, right? When you can predict what users are going to do and how valuable they're going to be.

It changes the game on a whole bunch of elements of a

MPD: commerce business. So this could be as simple as someone hovers over a picture and then clicks it hovers over the buy button. And those two things in conjunction over a huge amount of data sets. You, and I won't see the pattern, but the machine will figure out that combine with two or three other things and a duration, boom.

This person's a high probability buyer.

Richard Harris: Yes. It's really as simple as that but I can tell you, but you're hitting on the right things, which is, what kind of event data is actually there. And if you think about it, Interacting with the brand that there's a few different buckets, right?

There's like, how did that user get to where they are with the brand? Where did they come from? Was it some sort of marketing campaign or did they just show up at the website or they come through social? And then all the normal stuff you get when you're just integrated into a browser, like, where are they, what time is it?

What kind of device? All that kind of stuff you get for free. So that's one bucket. We call it like the referring data, but then there's the rich in session data. How does how's this user behaving? Are they, as you said, lingering on images, are they looking at a lot of images? Are those images for the same product or different products?

Are they scrolling deeply on a page? Are they moving quickly through the funnel? So there's all that rich in session data, which can be extremely predictive. And then there's how that evolves over time. So are they coming back? Are they looking at the same products? Are they doing different things that deepen their search and all of their past interactions with the brand.

And then we also look at product economics of different categories have different values to the brand. And all of that gets crunched together in 15 milliseconds to say, boom, this user will rely is part of a population that will reliably convert at 65%. Whereas another user may be part of the population that will convert at 0.02%.

And when you just ask a brand like, Hey, if you knew this in advance, you'd, if you had 10 deciles, or low, medium high, if you knew that someone was part of a population that will convert at 60% versus something close to 0%, what would you want to do differently? And then we get back in extremely long list because.

Kind of want to do everything differently, right? Certainly how they market, maybe how they price or offers promotions, UX with their merchandising how they prioritize their customer service queue. So there's so many elements of a business, how they text or email them. There's so many elements of business that could be really optimized.

Once you have this future knowledge, which hasn't been available till now, that's

MPD: phenomenal. Can you, what can you extrapolate from that? Continue. Can you get customer profiles to the point where before someone even lands or right when they land, what bucket they're going to fit into or they've got to come self-identify through behavior and then once they're on.

Richard Harris: Yeah, no, it's a great question. And we because we use only first party data, so we're not merging profiles, stuff that like the ad tech business used to do, we're not involved in that in any way. We're literally just a data processor of that brand's own first party data. And so we have to be able to provide a predict prediction on the first landing, right from pixel zero, from the moment, a user lands on a page.

And we're able to do that. Obviously, the more data you have about a user's behavior, the more high fidelity that prediction gets, but we can users from anonymous users who had just cleared their cookies, et cetera, et cetera, based on that real-time data, we're able to D average them and provide a prediction better than just thinking about all your users on that.

Which is what most brands have been forced to do. Yeah.

MPD: The idea of segmenting the customers down this way is not the typical you talk about in marketing, right? Customer segmentation usually is based around some sort of high level conceptual demographic, like a geography or something else thinking about,

Richard Harris: or it's like some psychographic, these are empty-nesters or nomadic millennials or whatever. And it's okay, that's interesting, but what really do you want to do? You can make stuff up about what you'd want to do differently for a baby boomer or a nomadic millennial, but ultimately your brand, trying to curate a relationship with the user centered around this product, whatever value you're delivering to the user and how valuable they are to you as a brand, it's like the number one factor, right?

From which decisions should be made and now they can be made in real time. Now

MPD: the behavioral. Profiles, you're able to create those tend to correlate with any of the psychographic or demographic profiles that we're so used to thinking about, or do they correlate more with an intent or an interest or a need or some sort of state of mind?

Richard Harris: Yeah, again, we're super data nerds. So we haven't really connected those things, which is if we can tell you how likely a population of your user base is to buy or subscribe or whatever it is. That's the number one thing, eventually we'll get to stitching that together to wait, who are these people?

These people who convert at 60% versus 0%, is there anything common about them that differentiates them right now? We don't know. But it's also not the number one way to add value for our customers, which is what we're focused on is delivering the hardcore. How valuable will this person be in.

MPD: Cool. So what do companies do when they get this data? They find out customer a is a 65% probability of buying something. And customer B is a 2%, there's a whole litany of things you said they could throw at it. What do you typically see people

doing?

Richard Harris: Yeah, the number one use case. And because we work in e-commerce and we work with flooded direct to consumer brands, the number one use case, meaning the place people want to point these predictions, those future knowledge is into their marketing workflow.

So if you think about it, besides product costs and sometimes not even CAC customer acquisition cost is the number one line item on the P and L of most commerce brands. And so when that's the case, If you can add some efficiency to that process, if you can bring down your customer acquisition costs, it can have a really big impact on your P and L.

And so that's where it gets pointed first. And, just to give you a sense of how that works. So we're predicting the future value of every user. In real time, it's only, we only use first party data. So that prediction that we pushed back, like an API fire hose of predictions, that's also a unit of first party data that the brand owns.

And because basically all software, all tools, every part of the e-commerce stack is set up to ingest a brand's own first party data. We just push that right into those platforms. And so now Facebook, for example, which is the biggest place where most brands are, especially direct to consumer brands are spending their marketing dollars.

Our predictions to show up as audiences, right in the Facebook kind of ad manager. And now, Hey, there's this population. Expected 60% conversion rate reliably. And another with 2%, as you said, do you want to bid the same way or do you want to make sure you have the same share of voice for these two groups and spend your money peanut butter across them equally?

Definitely not. And so by just making more rational decisions about your marketing spend in line with the value of these different cohorts, these different segments, you can improve your return on ad spend by 25 to 50%. So it's pretty amazing. And you can also scale up. Yeah.

MPD: So you mentioned it's a lot of direct to consumer brands currently.

Where else do you see the supply?

Richard Harris: So really we think about it, for now we're, hyper-focused on, on commerce and TTC and ex you know, moving beyond paid marketing, which we're already doing to expanding the number of. Of use cases. So we already have customers plugging this into Klaviyo and postscript and Zendesk and gorgeous and dynamic yield.

So all these places where, as we were saying, if you knew this in advance, if you knew who was, who in advance, what would you want to do differently? So that's our roadmap for the next few quarters. But to get back to your question, really anyone that has a CAC LTV equation at the core of their business, meaning there's a product or a service or some value delivery at the center, but then the business lives or dies by how much does it cost to bring a consumer to that product and get them to interact or buy it.

And then what is the lifetime value? So how does that customer acquisition costs pay off over time? And so if you think about. There's so many verticals like consumer financial services, consumer software education that even healthcare, there are all these places where there's something unique at the center, but CAC LTV is how you live or die.

So those are the places where we think real-time machine learning and this sort of predictive power can have the biggest impact.

MPD: I love the way you frame that on the CAC LTV bit internally and interplay. We actually, I believe all companies live or die on the CAC LTV. The there's organizations that famously like direct to consumer pencil and track all of the data, they know exactly what they spend to buy a customer, the CAC, and they know exactly what the lifetime value is. The value of that customer over time. And there's organizations that don't do it historically, enterprise. But at the end of the day, it is quantifiable. And we actually look at this internally when we're evaluating investments we'll see a enterprise company come in, we'll say, great.

What is your cost of acquiring a customer? And they'll say we don't track that. And we'll say, great. What do you spend per salesperson? And they'll say X and they'll know it. Fully loaded, travel and entertainment. The whole thing. How many customers does that person acquire every year? And they know that number just a little division and suddenly you figure out, oh, it cost you 30 grand to get this customer.

And they usually know that they can back into some sense of LTV, longer cycles for those. But I think it's so fundamental to all companies. I think there's probably arguably a bigger trend here in sounds like you're well-placed for it. Where increasingly data-driven management teams are going to be looking at that ratio LTV.

Up and down the scale of like transactional sales to field sales, to relationship sales and beyond, I don't know if you think this has a place in the world of enterprise or not, but

Richard Harris: yeah. I certainly hope so. Like the question of, do we focus on B to C businesses or B to B businesses that was something we obviously thought about and grappled with at the beginning, you have to find a use case in a vertical where the market poll was just so strong.

But we think about it ourselves. We're a, we're a technology company, but I know when I go to raise my next round of fundraising, that, the first thing a VC is going to ask is what's your customer acquisition costs? What was your growth? What's your, those things are so fundamental to SAS, businesses, software, like so many B2B businesses and where there's a good data set.

So were a lot of the interactions are happening digitally. I think our predictions for.

MPD: These concepts were known when I started in VC. Oh 6 0 8 timeframe, but they have become mainstream expected KPI, vernacular. Yeah. I told you when raising money needs to be thinking about this on some level and you shouldn't just be doing it for your investors are looking at it because it's a key driver of the success and the health of the company, all management should be looking at this in their own accord.

If you want to spend a dollar, you want to make three. It's that simple. It's that simple. It's that simple. If you're not doing that calculus, who knows what you're spending. Yeah. Who knows if the test, it sounds long-term.

Richard Harris: Yeah, exactly. If the, if you can't either see now or a path to those sort of micro unit economics making sense, which is if I spend a certain unit on sales and marketing and I don't get a unit out the bottom of that whole.

That will over time exceed that by some meaningful margin, then it's hard to imagine what is this, right? Is this a business or is it something else? Is it just a, a product or a feature that belongs inside of somewhere else where you can leverage a customer acquisition? Pretty much yeah. I totally agree. You have to see that path,

MPD: Richard, how did you start this? What's your story that got you here?

Richard Harris: Yeah, so my, big picture story I've been working in software and digital. Startups from the first one I've founded or co-founded in 1999. But this one in particular it was, so it was a really interesting confluence of things.

So in my last company completely unrelated to what black Crow does, but we had a skunkworks that was helping to solve a very different problem, but that was working on predictions. And that team started cracking some very interesting problems. Again, those sort of not widely solved problems and the biggest ones being, how do you do hyper fast, 15 millisecond machine learning?

How do you do this in such a way that the prediction is available? Like the moment it's required the moment after some new event or piece of data has. And this is something that, you may hear about predictions in the in the CRM world or in other places, but those are usually predictions based on static databases, right?

Like your CRM file, right? If someone purchases eight times, I can reasonably assume that they're high value. But being able to do that, before someone is in your CRM file, when they're just some anonymous internet user that hadn't been solved. And so that was a key unlock that said to us, oh wow.

We actually figured out how to do the data pipeline, what tools to use. And we also started figuring out, where it had been done before this was done, like by one company internally. And the way they would usually work is, and this is the way sort of enterprise machine learning works is if you're like some think of a fortune 500 company, if you're Pfizer pharmaceuticals and you decide, I need predictions inside.

You'll go out and buy Databricks, which is like the, privately held, but worth tens of billions of dollars ML infrastructure company. So you'll go buy a contract on that. That'll be, $10 million over a few years to go hire an army of data scientists and data engineers. And you'll start working in building internally on top of these sort of developer tools and data science tools, those projects, by the way, depending on who you read Gartner wall street journal, they fail somewhere between 50 and 80% of the time, but eventually these guys get there and yeah, it's insane.

It's insane, but they get there and it starts driving these sort of predictive outputs. They're not as sexy as the AI articles you read about like robots and self-driving cars and computer vision, but they are the thing that is really driving meaningful sort of economic value inside of the very largest enterprise.

And so when we took a look at that market and said, wait a second, where is so knowing we cracked a bunch of these problems that would make it possible to do real-time analysis, knowing that the market worked on in this fortune 500 model, meaning it was just by definition confined to a very small set of pretty global enterprises.

We said wait, if we built something as good as that centrally two things. One, could we create bespoke instances of it and deliver it as a service so that you didn't need to do this build a million times? And then second, if we could do that, could we deliver it and our sort of mission. For less than the cost of one data scientist annually.

And so some of those problems we were working on in the skunkworks inside of my last companies, before starting black Crow, we knew they could be solved. And so we had an incredible head start getting black core up and running because we knew they could be solved, but it also helped us clarify the mission, which is the fortune 500 is going to do things one way.

But the rest of the market this thing needs to be democratized. Everyone means that the stuff I just described to you about real-time predictions and using it all these ways, Amazon's doing that because they built their own system. But the middle of the market, think direct to consumer brand, a Shopify store with 20 or $80 million of GMV.

Why shouldn't they have access to that? Why shouldn't they be served? And that's what we set out to do at

MPD: BlackRock. That's amazing. What are some of the companies using you guys now? Because you've brought this service downstream, so it's not going to be all the IBM's in the world. Who else is touching.

Richard Harris: Meaning our customers are in the, yeah, so we work with a lot of brands that you may know brands like farmer's dog Cotopaxi daily harvest. So a lot of folks who have a really important direct relationship with their consumers, they may be selling more one-off goods, or it may be more of a subscription service.

But there are people who are very much focused on acquiring customers efficiently, and then making sure that relationship that they start is as valuable as a candidate.

MPD: Your, you mentioned a lot of very forward-thinking marketing organizations, often venture backed. Are you finding that, more traditional marketing teams or, maybe more dormant brands are grabbing onto this concept too?

Or is it just the frontier, the.

Richard Harris: I'll be honest with you. So we're, we've started with the avant-garde or the, yeah. Those folks who are a little more digital native think get this immediately. Like they know that they're living in time by the calculate LTV equation, as you mentioned, not everyone is fully digested that, especially in more legacy businesses.

But that's a function of not whether we can add value or not, but really whether who's going to have a fast sales cycle, who's open to testing and learning. And the fact is, even for those brands, I mentioned, we work with lots of other ones, like Missouri or magic spoon that, companies who are very forward-thinking.

But the interesting thing about this is even though they're, forward-thinking, we're selling machine learning to someone who ultimately doesn't give a crap about machine learning, right? What they care about is the outcomes, right? Can you deliver value? And so that's why, we made it easy, I mentioned it's a one-click install. So this three-year $10 million Databricks project has turned it into one click and then the model builds in two weeks, no work required. And then we just let our potential customers use it for a month. No obligations, no money changes hands. So everything we build delivers value in 30 days or less.

That's one of our missions as well as being able to do it for less than the cost of a data scientist. And when those two things hold it's a pretty compelling, it's a pretty compelling proposition and certainly for the avant-garde, but I think we'll be elsewhere. I shouldn't say we have some very large and medium size multi-channel retailers.

Who've been around for a long time who are seeing the value in this as well. Do you think

MPD: this tool scales to be an enterprise solution or are there things you need an enterprise that are never going to make sense and a productized SAS.

Richard Harris: Yeah. Yeah. It's so it's interesting question. We, as you well know just finished fundraise, I don't know, seven, eight weeks ago.

And it was a question that almost every investor asked and here's, what's interesting. Certainly there is a set of fortune 500 who are doing this on their own, that's why Databricks is worth however many billions of dollars. But there are many large companies who are nowhere on this front.

And the question is can we deliver value there? I'm very confident that the answer is yes. There's not that many companies who are doing real-time machine learning. They're just aren't. Are really, the only question we have about it is the go to market at this stage for our business, going to be a sustainable one.

Sales cycles are longer. There's a lot of compliance and et cetera. So we'll get there eventually. But for right now, Like in the, just thinking like middle market e-commerce companies, 20 million to a billion of GMV. There's 40,000 of them in north America, which is often a surprising number. And so we're going to focus there, learn a lot, and then we'll figure out what's our approach to enterprise value delivery.

I'm confident. This

MPD: seems like such a no-brainer. And I know the sales cycle is very short for those, for the, as you said, avant-garde the forward-thinking tech folks. Yeah. What did naysayers say? Why do people not buy this? This seems if you're moving a product online, it's a must have.

Yeah.

Richard Harris: So our conversion rate, once you get someone into a trial, the people who say no are very small, right? It's, single digits, low double digits of people who get through a trial, see the value and no, it happens, but it's really. It's really getting people to take that first click and to run the one month trial.

And there's a whole range of things. The biggest one is this sounds too good to be true. Which is why we develop the whole, just use it for 30 days. See for yourself. There's a lot of resourcing constraints, meaning I know it's only one click to integrate, but I don't have the team right now that has the bandwidth to make sure that this test works.

But ultimately, and we've seen so much of our business comes from, word of mouth and people on, in the marketing channels of DTC companies and the more proof points we get out there, the more sort of logos we can put on our website. I think we're starting to maybe crack through a little bit where a lot of brands are thinking, ah, I think we should try this.

And I do have this sense that I'm sitting on a goldmine of first party data that I'm not leveraging. And especially in the face. There's been so much turbulence in the iOS landscape the way, people can use their own data outside of their own environments. That's prompting people to say, okay, I need some solutions here

MPD: before this.

I know you did intent media. You also did Travelocity. You were there for a while in a senior executive role. And that's obviously a famed tech company in the first lap here. What was like the big internet? Boom. Yes. Any stories or nuance from that experience that kind of informed your view of the world?

Richard Harris: Sure. Yeah. Trouble Aussie known by Expedia, and there's been a lot of consolidation and travel. But my path there was super interesting. I was part of a one of the co-founders of a startup called site 59. Probably haven't heard of it. It was a travel technology company and we were working on the problem of this was like 99, 2000 working the problem of distressed inventory in the travel industry.

So how can you let suppliers use pricing as a demand stimulant without completely crashing their their economics, wherever, and just wait to the last minute, buy us a cheap flight. So that business, I, 59 grew like a rocket ship and was we sold it to Travelocity. This would've been 2002, 2003, and we thought, we had some amazing tech.

We thought they would integrate our technology and fire us. We were a bunch of 20 somethings in New York city and it turned into quite the opposite. It was, it turned into a minnow swallowing the whale actually. So our team ended up taking over and running Travelocity. So we went from being part of the a hundred bruising company in New York city to running this multi thousand person publicly traded entity led by literally, the CEO, my co-founder Michelle Peluso of site 59 became the CEO of Travelocity.

I ran a $2 billion piece of the business, but our CTO, our head of Europe, our general counsel. So we were just running this thing and it was a real culture shock to go from a little startup to a big global company. But it was, it was an amazing experience. It was a bit of the wild west early on in the first.com.

And it was Texas, not New York. They're based in south lake Tesco, Texas. So I spent a lot of time there. But it was an amazing experience.

MPD: How does that work? How do you go, how do you deal with being an agile, small team operator to wading through mud? How does that work?

Richard Harris: Yeah, it is it's the question, isn't it.

It's how do you bring that sort of very lane fast pivoting had. Mindset into large organizations and there's an entire, multi-billion