“What is Explainable Decisions as a Service?”
There seems to be truly no limit to what artificial intelligence is capable of doing. One of the best examples of this is Canadian startup Daisy Intelligence, which works primarily with retailers and insurers to make automated decisions that don’t require having a human in the loop. It’s what founder and CEO Gary Saarenvirta likes to call “Explainable Decisions as a Service.”
Gary joined startup coach Roland Siebelink on the newest episode of the Midstage Startup Momentum Podcast to share what “Explainable Decisions as a Service” means and why he doesn’t like the term artificial intelligence. He also talked about his journey to becoming a founder and Daisy’s progress thus far.
- Why Daisy feels the need to explain its AI decisions to customers.
- How to target multiple industries in the go-to-market.
- What it’s like having a global go-to-market in a competitive space.
- Why Daisy should have tried to become profitable sooner.
- Why startups need the right product if they want to burn cash.
- Navigating a product’s users being different from those choosing to buy it.
Roland Siebelink: Hello and welcome to the Midstage Startup Momentum Podcast. I am Roland Siebelink and I’m a coach and ally to many of the fastest-growing startups around the world. Today, we have with us Gary Saarenvirta, who is the founder and CEO of Daisy Intelligence. Hello, Gary. What an honor.
Gary Saarenvirta: Hi Roland. Nice to meet you and glad to be on the podcast.
Roland Siebelink: Oh, yes. The honor is entirely ours. Gary, for those listeners who have not heard of Daisy Intelligence yet, what do you do? Whom do you serve? And what difference do you make in the world?
Gary Saarenvirta: Our goal is to help companies make smarter operating decisions. We use artificial intelligence - I hate that term; I think it’s a marketing term for the most part. Our software is autonomous - no human in the loop. And we help retailers - is one industry we serve - we help them make smarter decisions around what products should I promote, what products should I not promote, what prices should I charge, how much inventory to allocate to every store, where to put the product in store. Core merchandise planning decisions automated with no human in the loop.
For insurance, as another industry we serve, we do fraud detection and claims automation. Deciding should I pay this claim or should I not pay this claim? Can I pay this claim with no human intervention? Is this fraud or not? What Daisy does is we recommend a decision with no humans in the loop. And we provide an explanation as to why. If a human does wanna look at that, why did we recommend that decision. I would call what Daisy does: Explainable Decisions as a Service.
When the clients execute our decisions in retail, we’ve shown we can grow total company sales by five or more percent - the total company retail sales, which is massive. It’s a testament to autonomous AI, what it can achieve. I think our decision-making is for problems that are beyond human ability, that are very complex. In retail there are tens of thousands of products, millions of customers, hundreds of stores. There’s just too many moving parts for human beings to optimally decide what to do from the perspective of item price and inventory. And similarly in insurance, you have millions of customers, you have millions of claims every year. You have historically hundreds of millions of claims. How can a human being decide if this one claim is fraud or not fraud?
Again, these are beyond human ability problems where technology can make a difference. And our goal is to elevate human beings in the sense that because these jobs are beyond human ability, they’re very difficult. It’s stressful, highly repetitive because there’s millions of decisions that need to be made every day. We want to let the machine take care of these types of decisions, which are ideally suited to computing machines, and let people do what people are good at like interact with other people, figure out how to service customers, and set the strategy.
The technology doesn’t set the strategy. Our goal is to elevate humanity, and I think in that way, if we can elevate the people and the company to be more strategic, then that creates shareholder value. By letting machines do what machines are good at letting people do what people are good at, we can service our customers better, we can create a better world, make the job more enjoyable, increase customer satisfaction, increase shareholder value. And ultimately, our goal is if we can do this in every industry, we can lower the cost of living.
Smart companies, when we create profitability, they reinvest in price. They don’t just dividend the money out to shareholders, they invest back in price because in most industries it’s price competitive. The goal is to lower the cost of living for humanity by being more efficient and smarter on these core operating processes.
Roland Siebelink: Excellent. Okay. That’s a lot to unpack.
But before we go into your take on artificial intelligence and autonomous decision-making, what was the history behind Daisy Intelligence? Can you enlighten us a little bit on how you got into founding this company? What skills did you have that brought it in or the unique insight that was at the root of this?
Gary Saarenvirta: I have a master’s degree in aerospace engineering, so computational fluid dynamics, very technical. I came outta school in the eighties and nineties. I went to the University of Toronto, the engineering science program and did my graduate studies there. And there was really no aerospace industry in Canada to speak of at the time, very minimal. I wasn’t prepared to leave Canada at the time. I ran into this accidental career and I got working with big corporations and saw how little math and science they used in decision-making.
At the same time, at the University of Toronto, Goeffrey Hinton was a professor there. The famous deep learning neural net professor, and I went to all of his seminars on machine learning and neural nets in the late eighties and early nineties. I was exposed to this industry way in its early days. When I started working with banks and insurance companies and retailers, I was shocked at how little math and science was used. I created this accidental career for myself.
Gary Saarenvirta: Then I worked for IBM. I worked for a company first called Loyalty One - runs a coalition loyalty program. And I started doing machine learning. They had a lot of retail transaction data from Canadian businesses. I became one of the first worldwide users of IBM’s data mining - was the buzzword for machine learning or AI back in the nineties. I was one of the first worldwide users of that technology and became an expert in that and training clients around the world. And then IBM hired me to run their data mining and data warehousing practice. I worked there for a few years and then I realized that supervised learning - and we can get into this more, my view of the AI - this predictive analytics doesn’t really work for complex problems. I feel the world is having the moment today that I had by myself 25 years ago. I believe it will come to the same realization that I have over the last 25 years and the reason that our technology has gone in a slightly different direction, more into the engineering space.
I started Daisy thinking that IBM’s a great company but they’re huge. They do everything. Maybe they’re good at everything, maybe best at none. I felt with my technical background, I can do this autonomous decision-making much better than they could. I founded the business.
Roland Siebelink: And when was that? When did you found it?
Gary Saarenvirta: In 2003. Many people ask why has it taken you 15 years to get to this point? Well, we’ve written a hundred million lines of code. Doing autonomous decision-making with no human in the loop - it’s not a write 1,000-line mobile app and you’re done. If you look at autonomous cars, I think an autonomous car has 300 million lines of code in it. Our production systems that make autonomous decision-making in retail in production are probably 20, 30 million lines of code, and we’ve written multiple versions over the years.
Roland Siebelink: I can also tell the way you discuss it upfront and explain that there is a lot of deep thinking behind what makes artificial intelligence - for lack of a better term - so powerful, especially the interaction with humans, the elevating the humans. Was this clear to you from the outset or is this more a consequence of the longer term investments in this industry, working with clients, working with people on how to actually find a niche for this?
Gary Saarenvirta: I think it’s been an experience in learning as we went along. First, I got very excited by predictive modeling. I think when people say artificial intelligence today, it’s really predictive modeling or statistical analysis. It’s just statistical analysis rebranded. There’s a human being at a laptop using sophisticated algorithms. I’d say deep learning is just a very sophisticated form of linear regression. It’s the same problem set up for supervised learning or for clustering, you would do an unsupervised learning method. It’s the same problem set up. Linear regression was invented in 1805. Problem set up is the same. I think if predictive modeling was the panacea, it should have run its course in 200 years.
I got excited about it because wow, I could predict, I could find these correlations, and I could build these amazing supervised learning models that were good in certain use cases. When I was at Loyalty Group, it was around direct marketing, so direct marketing targeting. That’s a great application of supervised learning and similar applications like that where the false positive rate doesn’t matter. You can carpet bomb the world with emails and yes, you’ll annoy people, but it’s not like getting a medical diagnosis wrong or something. I found that supervised learning was good for those types of problems.
But when I tried to use it to optimize a business, it doesn’t work for that. In retail, there’s a hundred thousand products. In retail, the pattern products are related. A customer buys ground beef, buys tomato sauce, pasta, cheese, bread, wine, salad for dinner, so there’s this halo effect. If you think of the interactions and products, both positive and negative, cannibalization. I bought Coca-Cola, I didn’t buy Pepsi. There’s negative cannibalization. Sale, pantry loaded, so there’s a pull forward. If you have a hundred thousand products, you have more than 10 billion first order interactions. Well, you can’t build a predictive model with 10 billion variables.
Roland Siebelink: One thing you mentioned in the introduction was interesting to me, that it’s not just a black box that makes decisions for people but also explains why it’s making that decision. And this chimes with me because I worked with an early AI company in advertising - to your point from before where it doesn’t really matter what you recommend - the patterns that it came up with based on predictive modeling were things like, “Let’s place ads where people are mostly engaged.” For example, at adult entertainment sites, which obviously was something the advertiser wouldn’t really appreciate. How do you find the rules that need to constrain that artificial intelligence and how do you balance that human judgment, that human intervention with the target orientation of your reinforcement learning?
Gary Saarenvirta: We have a dynamic, so the differential equations and there’s an explanation, so it’s math. There’s this halo effect that I described. Some products were bought together, all of a sudden there’s a negative positive pull forward. Then there’s seasonality, price elasticity, promotional elasticity, recent purchases. Did you just promote a product last week and therefore demand has been reduced in the market? All the normal factors that a retailer would think of. And then the context is - the differential equation is if you make better decisions today than you did yesterday, that will create incremental value.
The explanation is why did Daisy recommend ground beef on the front page of your website at 99 cents a pound. Well, we would say because that product has a large positive halo. People buy ground beef, they buy many other products. That’s one explanation. It’s a good week because you haven’t promoted in three weeks. It’s the right season. It’s the summertime. People are barbecuing and eating more meat in the summertime. It’s the right price point because it’s a very elastic product. When you discount it, people buy more and they’ll buy a bigger halo as well. And this is a better price point than you’ve done in the past. It’s a better price point than your customers. It’s highly promotionally elastic as well when you put it on the front page of the flyer. What we do is we compare the decision today to the decision in history and show that the decision is better than what you’ve done historically. That’s the delta explanation for retail.
Roland Siebelink: I wanna move a little bit away from technology and more to your go-to-market. You mentioned already that a lot of the applications of Daisy Intelligence are in retail and insurance. Two very different industries. How do you combine that targeting of two very different industries? And maybe there’s even more that you target.
Gary Saarenvirta: Yeah, the underlying technology’s the same as the technology description we had. They’re not as different on the surface. But underlying, the making decisions beyond human ability is different. Our product is very similar architecturally, so when we go to market, we go to industry-focused events and do marketing and industry-focused vehicles. We attend a lot of retail events, to a lot of advertising and retail publications, insurance events, insurance publications. I think it’s an industry-focused go-to-market.
The industries we’re in, probably in the world, there’s maybe 2,500 retailers and 2,500 insurance companies who could use the technology. We’re going to move it down to scale that a single mom-and-pop store could use our technology. We’re not there yet in terms of price point, but I think we have the capability to do that. But today, we’re working for companies that are typically a hundred million in revenue and up. That limits the size of the market.
We know the market. If I was to get to $100 million in revenue, which is a target of most companies, that means my deal size is quite large. We’re selling roughly $1 to 5 million a year as a current price charge. Let’s say I need to get to a hundred customers. A hundred million-dollar customers gets me to a hundred million. Am I going to have a hundred customers in the United States? Likely not; it’s a very competitive market. Out of the 600 retailers, you know - grocers or hypermarkets - in the US, if I capture 10 or 20 of them, that would be a coup.
For me, it’s to go around the globe. I’m getting one or two customers in each country. Maybe in the bigger countries, in markets like the U.S., Brazil, the hope is to have five, 10, 15 customers in those markets. The go-to market is really global, so we’ve met a lot of these global customers at trade shows, and we’ve found channel partners in other markets that help us sell, and that’s how we’ve gotten to where we are. We got to 100 customers, we’re so fortunate to get there. You look at our customer base and you’d see it’d probably be spread across 10 different countries, multiple geographies.
Roland Siebelink: And who -because these are relatively large companies if you talk about a hundred million revenue - who is the typical decision-maker that you target in your go-to market? Do you have a very clear strategy of this is the exact person we need to talk to?
Gary Saarenvirta: Yeah. We’re going after the C-suite. We’re selling a technology that is - the change management is very difficult since we’re replacing some of the human job roles - not replacing people because I don’t want to replace people, I don’t think that should be a goal of our AI. But we’re taking some of the job roles away. If you’re changing a job role that way, it needs to align with the C-suite vision. I’m not gonna go in and say, “Hey, we’re gonna alter your most important process,” which is merchandise planning in retail, claims processing in insurance and say, “Hey, we’re gonna automate that,” and never engage the C-suite. It has to align with the C-suite vision.
There’s a corporate value proposition. We’re gonna grow your net income in retail by more than 100%. These are the bright people, so we’re selling to the chief merchant, the chief financial officer. Those are the people who ultimately see the value proposition. The folks who would use our software and get the benefit are the category managers, merchandise buyers in retail, and in insurance, it would be selling to the same C-suite. The people using our software would be the investigators or adjusters. Our user audience is different than the decision-making audience, which is the C-suite.
Roland Siebelink: What is the unique value proposition that hopefully comes through when you talk to C-Suite people?
Gary Saarenvirta: We have no humans in the loop. We’re automation. It’s intelligent automation. You don’t need to hire any data scientists. We have an infinite number of data scientists running on millions of GPUs. It’s completely autonomous, there’s no human.
Roland Siebelink: Not hiring data scientists is a big benefit I think these days.
Gary Saarenvirta: Our users are business people. It could be completely automated so that the business people don’t even have to review. It’s complete autonomous decision-making, so that’s unique. And we do risk sharing. We stand behind the financial value proposition. We’ll guarantee the financial results. And we do risk sharing with many of our customers where we take a percentage of the incremental net margin created, so it’s net margin sharing.
Roland Siebelink: What can you share with us, Gary, around the traction of Daisy intelligence? Number of customers, number of industries,how long people stay with you typically. You already mentioned the $1 billion sales impact on your largest customer. That’s a great number too.
Gary Saarenvirta: We’re currently at about 15 customers between retail and insurance. Our biggest challenge is this human change management. Where we failed is where if you don’t use the technology, nothing happens. Because we’re replacing 50 to 60% of certain people’s job roles, there’s a lot of human fear in using this technology. I think that’s where we fail. The technology has shown that it works in every client we’ve ever had. As a tech company, in the early days, we were less well equipped to do the change management. And so we’ve been building more and more change management capabilities over time that makes the adoption much better.
Before the pandemic, we were growing at about 100% percent a year. The pandemic - we did a Series A raise, so we’ve raised about $15 million in equity capital going into the business since about 2019. The pandemic hit five months after our capital raise, our first capital raise. And we were flat during the pandemic. No one was going, “Hey, let’s automate my most important process right in the middle of the pandemic,” so the pandemic was like a giant pause button for us. Given that we had raised VC money and were burning cash, we got an early start to run the business to profit. I’ve run the business to profitability over the last two years, so that takes the pressure off fundraising.
I believe, in hindsight, I think it would’ve been smarter to run the business that way always. I think this burning cash is not a good strategy unless you have a product that sells like hotcakes. If you invented the iPhone and your problem is you can’t manufacture enough of them to sell, then maybe raising capital is a good idea or burning cash is a good idea. But raising capital and burning cash is probably not a good idea if you’re not that type of product. The discipline of running a business profitably is something that I’ve learned over the last two years. That’s what I did before I raised money. I self-funded for 12 years and then raised money and got caught up in let’s burn money and raise money and burn money. And I found going back to building a profitable business - and I think the investment environment is more now looking for businesses that are on the path to profitability. We got there and are there now and are working on growing EBITDA and building a profitable business.
Roland Siebelink: Especially since your sales cycle is challenging, I understand. And it takes a long time. And then also the implementation takes a long time. That may indeed not fit as well with a raise fast and sell fast model in that sense.
Gary Saarenvirta: Yeah. We’re a 12-to-18-month sales cycle. Since the pandemic, I’d say we’ve now grown again. We started growing again. I think we grew about 50% this year, and we expect to grow 50% to 100% next year. The growth is still healthy. We’ve also increased our pricing dramatically, deal size larger. And we’re trying to get longer term deals because the change management doesn’t happen in six months, so we like to sign multi-year deals with customers. It’s a commitment. And the business case is there. We’ve shown that on a net margin basis, you can get a 10-to-1 return on Daisy. And we’ll revenue share that.
Roland Siebelink: And that’s even with the higher prices, I’m assuming, right?
Gary Saarenvirta: Yes. Even with the higher prices. With our previous prices, there were crazy ROIs. That’s part of the reason we weren’t profitable. We weren’t charging the right for the service. But I think as we built more credibility - I had clients like Walmart Canada has been a client for a decade. In the Middle East, we have Carrefour, so some really large brand names lend credibility to the technology. And these are forward thinking companies who see this future of intelligent automation. I think it will roll out to the mid-market.
As I said, I think we’ve been ahead of product-market fit, which is what investors look for. I speak to hundreds of investors; they all reach out to us because we’ve won awards and people see our blogs and podcasts. They’re all very formulaic and looking for product-market-fit metrics. What’s your growth rate? What’s your TAM? What’s your unit cost? What’s your growth rate, retention rate, net revenue retention, all these metrics. For me, when we’re an enterprise SaaS and have a million-dollar ACB - those metrics don’t really apply when you have a dozen customers. I think we’re a different animal, that’s what we’ve learned over time.
Roland Siebelink: Fair enough. How big is your team at the moment?
Gary Saarenvirta: We’re currently about 35 people. I think at our peak during the pandemic, we were about 50. And then I think we’ve right-sized the business. I think we were hiring too much ahead of the curve. And I think that discipline of - I think you’ll never have enough resources. Even when I was at IBM, I never had enough resources. The idea that you’ll be fat, dumb, and happy with more resources than you need, I think that’s invaluable.
Roland Siebelink: Many founders have to learn that lesson. And also, that the more resources you add, the harder it becomes to manage the business and lead the business in the right direction. That’s very good. What is next for Daisy Intelligence?
Gary Saarenvirta: Well, I think this year the goal is - we’ve really figured out this change management - making sure that the existing customers we have are committed for the long term. Run the business to profitability, which we just have turned the corner on this past year and start to build up a nice net income over the next year. I think if we grow 50-100% in 2023, stabilize our base, build a little bit of a cash pot that we can use to then invest into growth into 2024 and beyond. I think it’s really stabilized the business in this new environment.
Again, there’s a recessionary headwind, so since we’ve raised money, we’ve been experiencing headwinds as a business. We feel very lucky that our customers have stuck with us, that we’ve been able to get the business to where it is today. And I think we’re just about to turn the corner to have a really stable, repeatable business. And then look at how we grow this intelligently as opposed to, we’ve been very opportunistic to date. Being more strategic about where to grow, perhaps raising more capital. Looking at perhaps either a majority sale or a minority investment raise with the parties who can continue to help us grow together with our existing shareholders. I’d love to see our existing shareholders get a return for their investment that they put in.
Roland Siebelink: Yeah, of course. Very good. Whoever has made it to the end of this podcast, how can they help Daisy intelligence? How can they help Gary the most?
Gary Saarenvirta: I think if you’re a company looking for technology automation, AI - I don’t like that term - if you’re looking for AI automation, I think we’re the real deal; would love to talk to you. If you’re looking for employment and you’re someone who likes what we talked about here, we’re always looking to hire smart young people who are motivated, eager to learn, and create the next generation of technology. You’re a channel partner who wants to resell, you have customer relationships, so anyone looking in those ways to help us, we’d be more than happy to have a conversation. Reference this podcast, say, “Hey, I saw you talk to Roland Siebelink, I loved what you had to say.” I do get a lot of outreach. I tend not to accept the requests unless they come with some kind of context. I’m on LinkedIn, you can look up my last name on LinkedIn. Look at daisyintelligence.com. You can find me there as well.
Roland Siebelink: Perfect. Okay. And of course, if anyone knows me and doesn’t know Gary yet, I’m happy to provide an introduction as well. But this has been a great interview, Gary. A lot of extra new learning around what I think we should by now call autonomous decision-making without a human in the loop instead of those horrible two letters that you mentioned before.
Gary Saarenvirta: Explainable Decisions as a Service, that’s what I call it.
Roland Siebelink: Even better. Thank you for putting that. We may actually put that as a headline of this podcast. Thank you so much for your time, Gary Saarvenvirta, the CEO and founder of Daisy Intelligence. Thank you so much for joining us.
Gary Saarenvirta: Great. Thanks, Roland. Happy to be on the show.
Roland Siebelink talks all things tech startup and bring you interviews with tech cofounders across the world.