Vivek Anand leads a team of data scientists, statisticians and operations research professionals that apply Data Science and Operations Research techniques to build production ready solutions for problems related to Price Optimization, Inventory Management, and Fulfillment Optimization for a Fortune Fashion retailer.
In this episode, Vivek delves into the strength of combining science and machine learning to make informed pricing decisions. He emphasizes the importance of starting small and continuously enhancing strategies through data-driven approaches. It also explores the differences in B2C and B2B pricing, the impact of AI on supporting salespeople, and the challenges of misaligned incentives in achieving business success.
Podcast: Play in new window | Download
Why you have to check out today’s podcast:
- Understand the need for compensation structures and incentives to align with business objectives that support desired outcomes
- Discover how AI is impacting the way salespeople handle pricing and negotiation
- Learn how to leverage both science and machine learning to make more informed pricing decisions
“If you’re new to the journey, start small. You don’t have to go and build a transformer network on day one. Even a simple EDA, Exploratory Data Analysis, can give you a ton of insights than just a guesswork.”
– Vivek Anand
Topics Covered:
01:19 – What paved his way into pricing
03:26 – How credit risk and interest risk influence pricing derivatives
04:37 – Thoughts on quantifying risks
05:36 – Data science in relation to AI advancement
07:39 – AI capabilities in capturing contextual and temporal information and other pricing techniques using AI
11:30 – Talking about AI and extrapolation
12:47 – B2C pricing versus B2B pricing and the significant difference in data availability between the two
18:38 – How AI can be utilized in the B2B world to assist salespeople in pricing and negotiation
23:57 – Vivek’s one best pricing advice
Key Takeaways:
“B2C is mostly, pricing is a lever to generate demand. Whereas in B2B, it’s more of a lever to gauge the willingness to pay from the customer’s standpoint.” – Vivek Anand
“I think data science is a big tent, and AI is a part of it.” – Vivek Anand
“The way I think about AI is like, instead of making it heuristics, it learns from the data or it just comes up with a mechanism that just translates input to the output, more systemic, faster learning and can be deployed at scale.” – Vivek Anand
Connect with Vivek Anand:
- LinkedIn: https://www.linkedin.com/in/va2260/
Connect with Mark Stiving:
- LinkedIn: https://www.linkedin.com/in/stiving/
- Email: [email protected]
Full Interview Transcript
(Note: This transcript was created with an AI transcription service. Please forgive any transcription or grammatical errors. We probably sounded better in real life.)
Vivek Anand
If you’re new to the journey, start small. You don’t have to go and build a transformer network on day one. Even a simple EDA, Exploratory Data Analysis, can give you a ton of insights than just a guesswork.
[Intro/Ad]
Mark Stiving
Welcome to Impact Pricing, the podcast where we discuss pricing, value, and the nuanced relationship between them. I’m Mark Stiving, and our guest today is Vivek Anand. I almost said Vivek Ramaswamy. Vivek Anand. And here are three things you want to know about Vivek before we start. He is the director of data science and analytics at a large, large retailer. He was director of Data Science at Zilliant. It’s a pricing company most of you have probably heard of. And he spent several years in pricing and analytics in the financial services industries. And he just knows a lot about a lot. So this is going to be fun. Oh, and he’s hiked most of the Appalachian Trail. So, welcome, Vivek.
Vivek Anand
Yeah, thanks, Mark. It’s good to be here.
Mark Stiving
Okay. Easy question. How’d you get into pricing?
Vivek Anand
Good question. So, it goes back to my training. So I have a graduate degree in operations research, and I went to school in New York, so that made it easy for me to transition to Wall Street where I was kind of optimizing risk and return pricing derivatives. And then, I was leading a big team of pricing, emerging market derivatives. And then at a certain point, I realized like, okay, this is not my thing because of the business environment and the regulations and all those things. I mean, I felt constrained. So the natural transition was to move from there to pricing, revenue optimization, and Zilliant, like a good place. So I moved from Wall Street to Zilliant. I was there for like a good chunk of time where I spent, again, instead of risk and return, I was optimizing profit and revenue. And then Zilliant, I worked on traditional B2B sector companies, manufacturing, distribution, you name it. And then I also got a chance to work on some niche industries of like, things that are relatively new to B2B pricing, like oil and gas. I worked in subscriptions a little bit. And then, currently I work for a retailer of omnichannel commerce, where in addition to pricing optimization, I also own data science related to inventory optimization and fulfillment optimization.
Mark Stiving
Nice. Okay. I’m going to ask a really off the wall question that I didn’t think I was going to ask.
Vivek Anand
Sure.
Mark Stiving
But since you used to price derivatives, it’s all about risk and return. That’s like a really quantitative area that I get completely. If you were selling cybersecurity products because I tend to find a lot of companies that are selling, I’ll say the word insurance, right? It’s, hey, we’re reducing risk or we’re making it so that this may not happen, whether it’s in cybersecurity or some other type of risk reduction. How would you go about pricing those? And do you think there’s enough quantitative data to do this quantitatively?
Vivek Anand
Cybersecurity is slightly different. but generally, if you think about the kind of derivatives I used to work in, the two sources of a risk were like, credit risk and interest rate risk. Credit risk is like, think about you have a hundred dollars to lend, and I say like, hey, Mark, give me that hundred dollars and I will like, a year later give you 105 bucks. And then JP Morgan comes to you and says, Mark, give me that hundred dollars. I’ll give you 125. You will choose JP Morgan, right? Not me. But if I told you like, hey, Mark, I’ll give you $125 instead of a hundred. And JP still says like, okay, I’ll give you 105, you’ll probably choose me. So that $20 difference in interest is the interest rate risk. So credit risk that you have, like, hey, this guy will probably not run away with my money, and if you run away, I have recouped enough. So that is credit risk. And then there’s an interest rate on, like, how interest rates are changing, and based on that, how the present value of cash flow changes. So that was like how derivatives used to be priced.
Mark Stiving
By the way, I love that, Vivek. But what I hate about it is that it’s formulaic, right? We understand the problem and when we move into other types of risk, we don’t understand the problem. And that’s really what I’m trying to get at.
Vivek Anand
Yeah. I mean I don’t really have a good answer to it, but I read a book, so from someone some time ago about risks, the system and those kinds of risks like, hey, what is the risk of I don’t know, like a nuclear warfare? I forget the name of the author, I’ll probably send you offline. And like, how do you quantify those risks and mitigate those risks? Those are not necessarily science similar to what we do in pricing, but there are ways to kind of quantify it, but there is more art than science, if you will.
Mark Stiving
Okay. I’m totally with you on that, but I just thought I would ask, because at least you lived in this risk and reward world.
Vivek Anand
Yeah, totally.
Mark Stiving
Okay. Another slightly off the wall question, but you probably have this one in the back of your head. Think about data science before AI and data science now. So, is AI making data science less necessary?
Vivek Anand
I have a different position compared to most people. I think data science is a big tent, and AI is a part of it, is the way I think about it. So, I mean, there are algorithms, there are techniques that will like to use data science. The way I think about data science is like, hey, you’re using data to make informed decisions, intelligent decisions, right? How do you make those decisions? Could be like, hey, you build a heuristics or rules-based process, which uses data in a certain way to come up with a certain outcome. Now, the way I think about AI is like, instead of making it heuristics, it learns from the data or it just comes up to the mechanism that just translation of input to the output, more systemic, faster learning and can be deployed at scale is the way I think about it.
And there are different components to it. Like, again, in the pricing, while you’re using our example, if you have enough data, you would know that, okay, around Thanksgiving my holiday sales increase, right? Or like, if I’m looking about summer, my AC or air conditioning cooler coolant sales will increase, right? So that is something that you know because there existed data on sales transactions to tell you, right? But if you are talking about something that goes over a very wide horizon for you to understand every nuance or every variable that kind of impacts it, but you don’t know what you don’t know, right? So there might be some underlying pattern that exists in the data that to you and I, it’s very hard to see. But AI has made it easy for us to see those patterns. Machines can learn those things by themselves and are able to translate that input or signal into a more informed outcome, I would say.
Mark Stiving
So let’s use the example you just gave, and let’s pretend for a second that we didn’t understand seasonality very well. We didn’t know that we’re going to sell more air conditioners in the summer. And AI goes and figures this out. Do they tell us that? Or do they just tell us, hey, change your price in the summer?
Vivek Anand
Actually, this is more recent, so like everyone has heard or used chatGPT. So, I was hoping I would not drop this word, but here we are. The architecture behind chatGPT is transformers, right? And what transformers do, if you think about it, they have a sense of attention mechanism as we call it, or they have a sense of temporal variables. So for example, when I say attention mechanism, like think about it. When we are reading a book or like someone is telling a story, and so it’s like, okay, the dog went and ate the ice cream; just making stuff up. Traditional data science modules we use would read the dog ate the ice cream. They don’t see the relationship between those words. But what the transformer mechanism is doing is providing that mechanism where it has the attention mechanism.
It realizes the focus is on dog, ice cream, dog, and each part of it, everything else is like constructor, if you will. So that relationship, essentially, the question that you asked was like, hey, can AI put context to what we are saying or what is seen that is part of the new methods, the transformers that we are talking about are able to contextualize that information. They’re able to take the sequential information, they’re even able to take temporal information. Like, to your question, like the older methods that we have, like if you have a time series model, it’ll just say like, hey, in summer the AC sales will increase. That would be an old time series model. Now, there is a temporal component where it says like, hey, when the temperatures start to increase, then the AC sales will increase. So if the temperatures become very hot in March, AC sales will probably start increasing in March. So this kind of temporal information is the AI infrastructure that we have right now that is able to capture that, whereas compared to the previous models that we used to have, they used to just look at the trend and like to be able to quantify the trend or fit the trend in some ways. Right?
Mark Stiving
And so what, I guess what I’m asking is, I get it that AI can capture that. Does it tell you the answer? So I’m always worried about the black box where AI says, hey, go raise prices, as opposed to AI saying, hey, it’s actually related to temperature not related to the month.
Vivek Anand
Yes. So that is a good question actually. So I came across new methods of pricing where instead of like, think about B2B pricing, the way it used to be done previously is used to create decision trees, now use segments and create segmentation trees and things like that. Then there are newer ways people are trying where they fit new networks to it, right? A pricing is a target variable, and all the variables are like independent variables fitting a giant regression. Right? Now, if there’s a pricing that comes out of the neural network, you’d think that, hey, it’s a black box. I cannot explain it. But again, there are some data science machine learning techniques that are able to explain that prediction. So there’s something called Shapley analysis, which essentially tells like, okay, here is your base price and here are the impacts of different variables that would move the needles above the base price or below the base price.
So that decomposition of what are the drivers within that underlying deep neural network that kind of explains what is the outcome, that explainability, thanks to techniques like Shapley or PDPs, we are able to do that now. I mean, again, this is very early stages, but like I have built solutions where we are able to use Shapley to explain like, okay, this number doesn’t make sense, but here are the factors that are driving it. And then when you start looking at those factors, it starts to make sense. But again, doing that at a scale for every prediction is a challenge.
Mark Stiving
So talk to me a little bit about AI and extrapolation. So one of the things we always learn about in statistics is the data’s only good for the range of the data that you’ve collected, right? So, it doesn’t go beyond that. In AI are we doing a better job at extrapolation?
Vivek Anand
I would not say that. Essentially like, I mean, if the data has not been shown or the relationship between the data points and the outcome that we are trying to predict or like trying to assess has not been there, it’s going to be hard, but still it’s going to be.I mean, it’s not a part of AI, but data science techniques where you can kind of do simulations, you can do causal inference where you can try to get to areas where you have not been and come up with a probability distribution. Like, okay, the data has, let’s just say, using pricing as an example. If you have not sold something beyond $60 and you’re thinking like, what happens if I go from 60 to 65? There are copper models, or you could run simulations where you can say like, okay, what is the probability distribution of volume when I go from 60 to 65? So there are probabilistic methods, but I would not call it AI. A lot of stuff that is like traditional statistics that has existed for almost 60 years is being passed as AI. It’s just the computer that has changed and made it possible for us to compute it. But a lot of this has been around for a long time.
Mark Stiving
Okay. Let me step back from AI for just a minute and ask a very different question. Since you’ve played in both B2B and B2C pricing, what are the big differences between B2B and B2C pricing?
Vivek Anand
Very, very good question. So B2B basically, if you think about it, is very custom, kind of customized for every customer. The sales cycles are along the contracts are negotiated, and no two pricing situations are the same. And in that case, what you are trying to do in some ways is target the willingness to ping in some ways. Like, hey, what is the price where I don’t lose the business? But at the same time, I don’t want to be pricing too low so that I leave money on the table, whereas B2C pricing is more like a business lever as I kind of see it. Basically, like for us, the pricing problem is more downstream compared to B2B, like pricing where we have like a demand forecast, we have assessment of like, okay, here is how much of a demand I expect, and then if I want to increase my revenue, then I need to prop up my demand.
And that is where we play the pricing lever. The other pricing lever, especially if you think about fashion retail or things that are perishable, right? Then there is a hard cutoff or when you should stop selling that stuff like meat, dairy, which is why you’ll see all of a sudden the price has cheap discounts when the product goes to expiration, right? So there the pricing is the lever to clean the shelves to bring the new stuff, right? So B2C is mostly, pricing is a lever to generate demand. It’s the way I see it. Whereas in B2B, it’s more of a lever to gauge the willingness to pay, if you will, from the customer’s standpoint.
Mark Stiving
Okay. So, I want to share my opinion of this and it’s kind of similar to what you just said, but I want to hear your thoughts or maybe we can figure out how to tie these two together more tightly. One key difference between B2B and B2C is that as you pointed out, we often have direct salespeople negotiating with individual customers. And so therefore every customer gets a different price. And what we’re focused on is for that customer, what’s the willingness to pay. However, in B2B, we also often sell through distribution or even on our website. And so we do have some of this, I call it ToLi, take it or leave it pricing where we set a price and people either buy or don’t buy. We’re not negotiating individual deals, it’s just, hey, let’s go set the price. Where almost all of B2C is ToLi pricing, right? Take it or leave it pricing. We’re going to set a price and you buy it, you don’t buy it. That’s just the way it’s going to be. So I see that as one big difference. And then the second big difference is the amount of data, right? B2C, there’s just a ridiculous amount of purchase data, and in B2B we don’t have that. And even if we had it, I don’t think we captured it well enough to be able to say, oh, here’s why this person paid this much more than this person.
Vivek Anand
Oh, totally. I mean, just filling a neural network or a transformer, it’s so much easier for me to do in B2C compared to B2B because of the data, right? so I mean, to your point that there are some B2Bs, which are kind of like running this like ticket or leave it pricing. I have worked with some of those in the past. So they have a good deal of data and you are able to come up with a good assessment or forecast of like demand, or you are able to even simulate like, okay, what is the impact of pricing on this set of customers? The other thing that I think what happens in the B2C and B2B world is like, for this reason, data disparity is why we use different kinds of techniques in B2C compared to B2B.
B2B might sound more rudimentary, like because, oh, you’re not using a fancy deep neural network. You don’t need a tank to kill a fly. Right? But yeah, to your point also like the data elements, what is driving a negotiation situation is also not well captured. And that’s partly driven by the fact that companies have conducted business. I knew there was someone who was trying to maximize profit, but they did not have a mechanism. The incentive for the salesperson was tied to how many deals you made, not how much money you brought. So your objective and the incentive, if this is misaligned, then you cannot kind of come up with a smarter way of predicting because what are the drivers that do it through price? Well, it was a salesperson incentive. There’s no way to capture that.
Mark Stiving
Yeah. If you think about it, every time a sale is made, odds are really good that person was willing to pay more. We probably didn’t have it priced at exactly the right price. Every time a sale is lost, there’s some amount you had to discount in order to win that deal. And, it’s never just a penny or a dollar, it’s some number. And we just don’t know what that is. And so what’s funny is in B2B, we’ve picked a price, we negotiate a price and we win or we lose, but we have no idea where we were in the customer’s real willingness to pay. In B2C, we get a lot of examples of people saying, yes, I buy, no I don’t buy. And those are people who say, I’m either above your number or below your number in my willingness to pay. And so that gives us a lot better data set to make these decisions.
Vivek Anand
Yeah. And I think I advised a startup or like a small company on how they should set their price. Not like how they should sell their SaaS, but how they should set their subscription price and it was in that area, because if you are running surveys, even with your potential customers, like surveys have this impact of low balling, right? . So yeah, it’s willingness to pay is like a hard nut to crack. We were able to do a different kind of survey to solve that problem, but still, I mean, it’s much, much, much harder to point. Yes. Totally.
Mark Stiving
Yeah. Okay. So now I want to bring it back to AI for a second. And let’s not talk about the ToLi part of B2B . How could you use AI in a B2B world where we’re teaching salespeople what the right price is or how to negotiate deals?
Vivek Anand
Totally. First thing like, I mean you have very likely done it probably have done it. So basically like the B2B pricing, the way it works, the whole point of segmentation, if you think about it, right, is to create or group similar selling situations, right? And this is something that I have done at one of my previous clients. So what we did was we built a solid science, but there was a big amount of pushback from the salesperson. So what we did is where in this situation, if you have something exceptional, like a request for exceptions, what we used to do is we created a business intelligence tool like Power BI or, or Tableau or something. And virtually like what the manager who you had to kind of, alright, the exception used to say like, hey, here is a similar selling situation, like a segmentation, and here is the range of prices.
What you’re asking is like way below where it is, what is driving it. Just having that conversation kind of makes it much easier to train people on like, okay, here is the range of prices and here is a typical price that you need to do. And that kind of solved a good deal of their problems. Like salespeople are coming less often with requests or exceptions and they kind of democratize this like, okay, here is how we are coming up with this recommendation. And if you think about this democratization and having a business intelligence tool, the underlying of it is science, right? Where we kind of did smart segmentation we just did not only rely on what the business guys were telling us, like, okay, here are our attributes. You go build a segmentation tree. We came up with our assessment of attributes and then this is where machine learning kind of played a bit of a role.
So, if you have a ton of attributes and you are trying to iterate over one after another and trying to build a solution, you are setting yourself up for failure. So what we had done was we were able to put that whole array of attributes that the business deemed important, the salespeople deemed important, and what we saw in the data in our EDA through like a regressor or decision regressor, and we were not using it for like building a regressor but like a prediction engine, but more like understand the feature importance, what are the attributes that create the best split in the trees. And some of the attributes that we found were not something that neither the salesperson nor the business sponsor had given us. We were able to find that in data using machine learning. And if we were assessing one, two over time, it would’ve taken us forever, probably would’ve never gotten that. And once you have that segmentation, it’s relatively easier for people to kind of explain like, hey, here is a similar institution. So I guess I’m repeating myself here.
Mark Stiving
No, that makes a lot of sense. And so then once you have it, I’m going to go back to AI again, could have the AI answer these are the similar deals that we’ve won at these price points.
Vivek Anand
Totally, like for example, like, to your point, a previous point about data capturing. So I was working with a paint company and they had their customers classified as builders, contractors and DIY and so we created customer segments, builders, contractors, DIY, like people like you and I. And there were a ton of cases where they were like, okay, this doesn’t make sense, right? So we took a step back with like, the fundamental question was like, hey, are there just three segments in this universe of customers that we have? And do we want to take objective decisions on where we are? And a simple run of affinity propagation told us like, oh, okay, there are five clusters, so there are five distinct customer profiles in the universal data we are looking at.
And then when we took each of those clusters and assessed them, it turned out that there were some builders which were big homes like Pulte homes or Lennar or whatever. And then there were small builders who were just building one Z two Z houses. So they were getting classified as a builder, getting a sweetheart deal despite not being a big builder. And that was something that the data was not capturing. We said like, okay, you need to start capturing this data. So AI was able to tell us like, okay, here there are five types of customers, you are only looking at just three here. And it gives us the directional sense for us to go find those additional two.
Mark Stiving
Nice. And that makes a lot of sense too because if you think about builders, there’s a huge difference between a developer who’s building a hundred or a thousand homes and someone who’s building one at a time.
Vivek Anand
Yeah. Totally. And same thing, like there were contractors who are not painting contractors painting our houses, they’re contractors who are painting stadiums, so they’re not sensitive to prices. So, and that was like, you call them contractors, they’re paying you a top dollar and like you’re not capturing that just by tagging them as a contractor. Data was able to tell us like, okay, these are like painting contractors. They’re not painting contractors, they’re like contractors within a different world where they’re not sensitive to price stages of the paint.
Mark Stiving
Right. Nice. This is excellent. We’re going to run out of time here in just a second. So let me ask the last question because that’s going to prompt me to ask you more, I’m sure. Which means it’s not the last, but we’ll pretend like it is. So what’s one piece of pricing advice you would give our listeners that you think could have a big impact on their business?
Vivek Anand
I mean, the one piece of advice is really, I mean, if you’re new to the journey, like you are not using data, you’re still using Excel workbooks or whatever, I mean, start small. You don’t have to go and build a transformer network on day one. Even a simple EDA, Exploratory Data Analysis, can give you a ton of insights than just a guesswork. I mean, I don’t know how many times I’ve seen business people say things like, hey, this is a very important attribute that does X, Y, Z to win business. We do the EDA, we show it to them like it doesn’t happen. And they’re like, okay, I buy you. So start small. The second thing is just the…
Mark Stiving
Before you move on, I just got to say what you just said reminds me of the book Moneyball where we’re convinced we know what it is, but once we start looking at the data, it’s like, nope, that’s not what it is.
Vivek Anand
Yeah, totally.
Mark Stiving
Okay, go ahead.
Vivek Anand
Yeah, the second one really is, align the incentives to your objectives. Right? The previous example that I gave, I see this happening often enough. If you’re incentivizing a different objective function and you are running your business to a different objective function, they’re going to be misaligned. If you’re compensating your salesperson based on the number of deals and not on how much money they bought in, you’re setting yourself for a failure where there will be a race to the bottom. So one is behavioral, the other is just like, start small and like there’s always a better way of doing things using science. So I’m a big proponent of science, let’s just put it that way.
Mark Stiving
Nice. On the incentive one, I often talk, when I’m teaching companies, I often talk about the example in the book Freakonomics where the realtors who are selling their own home leave it on the market longer and sell it at a smaller discount than if they’re selling somebody else’s home, which just demonstrates that when you’re you’re compensating someone based on a percent of revenue, it isn’t aligning the right incentives.
Vivek Anand
Yeah. I did not read the recent settlement for the National Board of Realtors, but I got emails like, okay, not emails, actually physical mail where they’re settled with the US government saying, to your point, inflating the price because that maximizes your objective.
Mark Stiving
Yeah, exactly. Anand, this has just been, or I’m sorry, Vivek, this has been so much fun. If anybody wants to contact you, how can they do that?
Vivek Anand
Yeah, I can definitely send you my LinkedIn. LinkedIn is probably the best place to contact me. And once we have a basic introduction, we can kind of just take it over to phone or email.
Mark Stiving
Perfect. And we’ll have your LinkedIn URL in our show notes, I’m sure. Vivic, thank you so much for your time today. To our listeners, thank you for listening. If you enjoyed this, would you please leave us a rating and a review? And if you have any questions about this podcast or pricing in general, feel free to email me, [email protected]. Now, go make an impact!
[Ad/Outro]
