Let’s be honest. If you’re running a company you spend most of your time trying figure out what the hell is going on.
Then you spend the rest of your time — what’s left of it — trying to get your people to do the right thing (based on what’s going on).
Computational power is a moving target. Moore’s law is still in effect. We’ve had to completely upgrade our systems several times. We’re constantly adding and improving new systems and taking down old ones.
The first part you could call, “analytics,” and the rest you could describe as “leadership.”
So there you have it. We’re very focused on leadership and analytics.
Recently we hosted the Vendo Partner Conference in Barcelona. We got together with our other industry leaders and discuss our collective experience with both topics. At the end of the conference we decided to organize some of our thoughts and reactions.
This is a kind of “capstone." A capstone is the last piece you put on an arch. It connects the two sides. Our experience with artificial intelligence (AI) relates to both leadership and analytics.
What started us on this journey towards AI was a story.
Imagine you are in a medieval market. It’s a medieval market in the middle east. You arrive at the market around sunset. The market’s dusty. The men selling their wares are starting to close up for the day. You walk up to a stall. The man is selling oranges and almonds. There are no prices because he can’t write. Prices wouldn’t matter anyway because you can’t read. Literacy was for kings and priests back in those days.
The shopkeeper sizes you up. You’re wearing nice clothes. It’s the end of the day so he knows you are running out of options. He starts to ask you questions. “Where are you from?” “What do you do?” He watches while you squeeze the oranges.
He’s gathering lots of information so he can give you a price. It’s a one-to-one sales process. He wants to give you a price that gets you to buy and spend as much as possible on his products. He’s calculating probabilities of conversion and lifetime value in his head. He’s pretty good at it. He has lots of experience to draw on. You’re not the first customer and those aren’t the first oranges he’s ever sold. He finally settles on a price.
That’s the way it has been for most of human history. It’s still that way in many places.
We’ve had that story in mind as we’ve thought about the barriers to one-to-one selling online. We have a vision of moving towards one-to-one selling online because it makes the most sense. Well, dollars and sense. It is a return to how things were done in the past. It generates the most sales and the most revenue. One-to-one selling is also very helpful for risk. It keeps your accounts active within risk limits. It’s dynamic.
It generates sales you didn’t make because your fixed price was too high for the initial or even the rebill.
It creates revenue you didn’t make because your fixed price undercharged a shopper who was happy to pay more. It lets good transactions in and keeps bad ones out.
When we started out we had the wrong team for AI. It was a team designed to do normal analytics.
One of our partners has a brother who used to work in the CIA. He was a spy. He had two jobs. One was to stop nuclear weapons from getting into the hands of terrorists. His other job was to report to the US Congress so that they could make decisions. He liked the first job and he was good at it. No terrorists blew up any cities with nuclear bombs while he was there, even though lots of them tried.
He didn’t like his other job. There’s a reason that the US Congress has a 7 percent approval rating, or said another way, 93 percent disapproval rating. They earn it. Part of the problem that congress faces isn’t really its own fault. They are asked to make too many extremely complex decisions, one after another. Imagine a congressman’s day. He goes from a hearing on farming policy to fund raising to celebrating his local basketball team to talking with reporters about the latest crisis to campaigning to a hearing on covert operations against terrorists trying to get nukes. And he’s expected to make the right decision at each moment. It’s impossible. It’s asking too much of him.
We realized that information overload was also problem for us. We weren’t going to be able to review and analyze the data and make the right decision each time. There were too many situations. Too much data.
An average client of ours with an average site has around 100,000 visitors a day. That’s more than one per second. We can’t possibly figure out which transactions are good and which ones will chargeback or which price is going to convert best and generate the highest lifetime value each time. Imagine each different shopper coming to the site. It’s happening right now. One ... two ... three ... up to 100,000.
How is this visitor similar to another? How is this visitor different? How is a risk decision or price you gave someone similar six weeks ago performing today? Should you go tighter or looser on risk this time? Should you price higher or lower this time?
To figure out the answer we needed a new key performance indicator (KPI). Andrew from Bang spoke during the conference about this in a session he hosted on “key metrics.” For risk the metric became how many risky transactions we could locate in an ever smaller percentage of transactions. For pricing, sales wouldn’t work because a lower price always generates more sales. Lifetime value didn’t work because a high price would always give you higher LTV. We needed to combine them and focus on revenue per shopper. This lead us to revenue per click, or RPC, as the most useful KPI.
In situations this complex, with this much data, it is obviously impossible for the human brain to make a contribution. It’s like the congressman spending 30 minutes on terrorists at the end of a long day of appointments. He’s probably not going to add very much to the conversation. If he takes the right decision, it’s simply because he got lucky. That’s how we felt, years ago, as we looked at the risk and pricing tasks in front of us. We literally had no idea. And we were not going to get a good idea about whether or not to block that shopper, or the right price for that shopper. It was asking too much of our human capabilities.
We had a team that could give us advice but, like the congressman, we couldn’t really act on it effectively. We decided that we needed to change the way we were working.
Charles from Gamma shared his experience with reorganizing teams during a session he hosted at the conference. It’s one of the biggest challenges a leader can face. It brings together strategy, people and execution. The three areas a leader needs to focus on each day. And it’s full of risk. It’s scary. He showed a lot of courage in the way that he reorganized. We all learned a lot from that session.
For us to pursue AI effectively it was clear that we needed a different team. We needed one that could build a system that would act on its own. Today there isn’t a single person on our analytics team with more than four years inside the company. The tools they are working with are new. Many of those tools didn’t exist or had not proven themselves valuable until recently.
Our analytical team members need both collaboration and solitude. Working remotely — even just from home here in Barcelona — is key for them to develop ideas during uninterrupted blocks of time. Julius from Playa Media spoke about criteria he uses for deciding on which roles can be remote in the “remote working” session at the conference.
They needed new ways of meeting together, too. Sean from Revolution Force spoke about meetings that are really, in essence, about meetings during the session he hosted on the first day of the conference. We were embarrassed to discover how quickly our meetings could devolve into massive wastes of time. Too many people, too unfocused. This lead us to shift more and more to what Buddy from Vendo described in the session on “collaboration tools” as a Basecamp 3 centered organization. He spoke about the importance of having tasks all in one place with due dates and names attached for each task. Agendas and minutes are there, too, for anyone who wants to see them.
Another part of deciding that we were ready to develop AI for our vision of one-to-one selling was to make sure we met other requirements for AI.
Did we have enough data?
Did we have the computational power? Could we do it economically?
If we didn’t have either of these two then the third key element, the team, wouldn’t matter.
Fortunately we have data from large, medium and small companies across the industry. We host the pre-join page and apply our AI for pricing and risk there. Because we host this page we get a ton of data. We’re ranked 8,000 on Alexa. That may not sound like much but those are all shoppers who’ve reached the end of a tour and are making a purchase decision. If by chance you aren’t familiar with Alexa, it’s like golf, a lower score is better. Matt and Fadi from MindGeek have a property that’s ranked 60.
We have data on hundreds of millions of shoppers. And we’ve got data over time. We can see the effect of our decisions on pricing, risk and other areas over years.
Computational power is a moving target. Moore’s law is still in effect. We’ve had to completely upgrade our systems several times. We’re constantly adding and improving new systems and taking down old ones.
The three keys to moving closer to our AI vision of one-to-one selling are teams, data and computational power. Over the last three years we’ve had to work with limitations in each area.
We didn’t have the right people. We had to build the team. Thierry from Vendo talked about the progression of analytics teams as companies grow in the “analytical teams” session. Not everyone we hired was a good fit. That’s how we learned to give them detailed tests at the beginning. Several of the people we hired recently have biomedical backgrounds. There’s a lot of crossover between the two areas. The guy who built Facebook’s analytical team is now working on finding cures for diseases. Try look there for your next analytical hire.
The data we had at the beginning was practically useless. A friend of ours named Luca is a statistics professor. In addition to his research he works with the E.U. to do election monitoring around the world. When you see on the news that some country has had elections and “international observers determined that the results were free and fair” then chances are that he was there. One of his mantras is that “90 percent of statistics is getting clean data.” That’s been true in our experience. We weren’t collecting the right data at the beginning. Or we were collecting data that didn’t matter, that just created noise. Or, we were limited in how we used the data. We created hierarchies that didn’t make sense. Matt from MindGeek talked about the frustration of not being able to get the data you need in the session on “analytical tools.”
Here’s a challenge we had, for example: Rank order the importance of these variables and their interactions as they relate to risk …
- Postal code
- Time
- Day
- Site
- Niche
- New vs. Returning
There are more variables, too, but just imagine how you would rank the relative importance of each one and interactions that can include all six, that’s 46,656 different configurations. Not only is it difficult to imagine how you would do it. The variables change in importance constantly. At 2 a.m. local time may be the most important variable. At other times of day the actual time may be relatively less important.
Laurent from Kehpri talked in the “testing” session about tools that go beyond human interaction and into machines taking both the decision and action. We discussed tools that close the loop.
Next we had challenges with our computational power. We had to work with limitations. There would be a way of calculating that seemed perfect but would use all the computing power in the known universe. Since we couldn’t access that much power we had to look at other ways of calculating. We also have to look at costs. We make a fraction of what the end user spends and we have to run a profitable company. We can’t over spend to generate results. Those are real constraints that we have to factor into our decision making. Google, which has no shortage of cash, is famous for creating their own cheap servers to power their search engine. We’ve also created custom server configurations to enable us to efficiently process mountains of data.
Next we needed to tell the world about it. We’re using AI throughout our company. We think of our company like a cyborg, part human and part machine. It’s here in pricing, it’s there in risk and it’s heading into other areas. It’s our present and our future. How do we talk about it?
We chose pricing because it is the clearest, simplest example. Now, because of Visa’s decision to lower risk thresholds our AI approach to risk is extremely relevant.
As we get closer to one to one selling we’ll continue to face challenges in all of these areas.
We’ll be trying to figure out what the hell is going on…and building a culture that can respond well to changes. We’ll be running our company and encountering more analytical and leadership challenges along the way.
Thierry Arrondo is the managing director of Vendo, which develops artificial intelligence systems that allow merchants to dynamically set prices for each unique shopper.
