All right. Well, hello and thank you all for tuning into another episode of the Professional Pricing Society podcast. My name is Terence, and today we're going to be discussing B2B pricing in the era of AI and data science. Spearheading this conversation is a director at Gap named Vivek Anand. Vivek is an operations research professional with over a decade of expertise in leveraging machine learning and AI and pricing analytics and optimization.
In his previous role as the Director of Science at a leading B2B pricing vendor, Vivet cultivated an impressive track record where he successfully developed analytically driven pricing solutions that led to revenue and profit growth for numerous Fortune 500 companies across various traditional B2B
industry verticals. Additionally, his expertise extended to the niche industry segments such as oil and gas, legal advisories and subscriptions and HR solutions, where he innovative price optimization frameworks for these industries. Vivek, how are you doing today? I'm pretty good. It's great to be on this podcast with you. It's really a leading marketplace for pricing ideas. Yep, great to be with you. Good, good. We're glad to have you.
Let's go ahead and just dive right into this conversation. B2B pricing in the era of AI and data science. So can you share with our audience a bit about your professional journey and delve into the specific ways your expertise and B2B pricing has kind of evolved over the years, just to kind of give everyone a better grasp of your background?
Yeah, absolutely. So currently I'm Director of Advanced Analytics at a Fortune retailer Gap where my team's managed really is to optimize business outcomes. It could be profitability, productivity, actually it's both for use of analytics for the enterprise. So basically my team primarily deals with application of data science analytics in pricing analytics, inventory analytics and buying analytics. And like I said, prior to my current role, I was Director of Science at a leading pricing
solution provider. And in that role I worked with a number of B to B clients across various verticals where I help kind of build or improve upon their existing pricing programs by developing machine learning driven practice optimization solution. I kind of also worked with like you said, niche industries like oil and gas and subscription legal solutions where like you did not find many academic or industrial scientific literature on how to use data science
analytics to optimize prices. And to your second question about like how it has evolved, When I started in this industry like 5-7 ten years ago, it was like slightly different. The data science uses were like not prevalent. A typical price optimization solution like still constitute segmentation, computing, elasticity and optimization. But all of those 3 components have evolved over time, right?
So previously when I started segmentation, attributes were kind of divided decided by the business stakeholders like they would feel like, hey these attributes are important. They will have like a short list of attributes and you use primarily, primarily use like basic or simple statistical techniques to identify like which of those attributes made sense and which one of those did not.
And similarly when you go to elasticity again it used to be like a regression and still is a regression can be log log regression, linear regression. But those methods are not always appropriate all the time. So for example there have been cases where you do not have enough data and in that case you're trying to create a regression. Then like a few months in, you have new sets of data come in and all of a sudden your results are different because like 1
outlier can change it, right. So it's still a robust method, but the way it was used traditionally was not appropriate, it's not optimal if you will. And similarly like price optimization, like there is this challenge industry, but there are many, many places where this price optimization is basically a rules based price setting, right? Like OK, when this happens to this kind of thing, right. So when I started, what I did was I started experimenting with data science algorithm.
So basically to get these things done faster and better, so I developed a new machine learning framework that would quickly assess a battery of segmentation attributes. And then we are not just limited by what the business stakeholder thought, they were important, so we could run data analysis EDA quickly, faster to understand like OK, these are the dimensions on which you can slice the data and then also
assess their attributes. Sometimes like the business, the hypothesis that we get from the business is not visible in the data and we are coming up with better attributes. So this machine learning framework can assess a battery of attributes in a relatively short amount of time. Similarly for elasticities, I brought in like newer and faster techniques. I need to talk more about these like later, like faster sampling algorithms that could converge
and reduce the run times. And scale is a big problem, right? Like 5-10 years ago, you are primarily dealing with a small amount of data on transactions, customers, products. That's about it, right? And now we have a lot of data, right. So how do you go from having a method that was suited for a small amount of data to a large amount of data?
So this is like net net the cultural shift where the clients, the solution providers and the end users are kind of not shying away from using data science and analytics and leaning on machine learning, data science to make faster and smarter decisions. That's like overall tectonic change that has happened in the landscape. OK, awesome. That's cool.
Now, you do have a very extensive background in this, even though things have changed in the past five to 10 years and you've been at the forefront of applying innovative data science and AI techniques in pricing. So, starting with segmentation, do you mind walking us through some fascinating instances where you have kind of harnessed AI and data science to create a more robust segmentation solution? Yeah, yeah, absolutely. And this is, this is such a
great topic. Actually there are many such examples of it. And as we speak, like the one that comes to my mind is a customer where like they have like a traditional segmentation model. The customer had implemented a customer segmentation model for their own customers across 2 dimensions, like how much their customers spent money over a year and what is the price they
were like paying, right? Like a price index, which effectively kind of measures like, hey, this person paid below the average price paid by all customers or above, right. And if you think about these two attributes, think of it as Skype plot and these two dimensions will create like 4 quadrants, right? You break the customer base by spent and then you further break it by the price index, right? So you have 4 bad clusters or or or segments, if you will.
Now the salesperson used to come to them and say, like, hey, I don't really like this pricing because I feel like these two customers are selling situations are similar. I feel the price guidance for one is very low and we are leaving money on the table, whereas others is very high. And we are kind of like at the risk of not winning the
business, right. And this was a problem that like kept on repeating and and that sort of also erodes the trust from the end user, the person who is using the pricing solution. So the customer like the business sponsor came to us and asked like, hey, like this is what we are hearing. And I kind of agree with what they're saying is in their argument like, can you like make
sense of why it is happening. So again, like looking at the transaction data, it's very difficult to kind of like have a scarab water and eyeball and say like, does it make sense? So that was like one of the first use case of data science, like years ago where I just like ran a unsupervised crossing algorithm. So basically instead of telling the system like how many clusters of segments of data you need to create, go figure how many should be there.
And the machine came back and said like, OK, there are optimal 5 clusters instead of the four that the traditional segmentation tree was creating, right? And then we did some further analysis and we found that there was a cluster of customers or transactions that were not the top spenders or not the bottom spenders. They were right in the middle. And then there was another set of customers that were like not paying the top dollar but also
not paying the rock bottom. So they were like right in the middle. And you partition that middle cluster into four segments, the way the traditional segmentation was done, 2 customers with a very similar selling situation, one goes into a high price cluster, other goes in a low price cluster, right. And that was the source of the
problem. So we did some further analysis and work with the customer to kind of give them like hey, instead of like doing this customer segmentation across these two attributes like in the segmentation tree maybe we have one attribute that like clusters customers and you have those five clusters more a data science driven cluster segmentation and that like low behold the problem was kind of solved, right. So that was one good example.
Another good example actually goes back to customer specific price, B to B, like as you know, it's very negotiated bilateral conversations, right. So segmentation will give you a segment price, right? But again, there are situations within the segment that there are customers who pay the rock bottom or below the floor price. Like in typical price guidance, you have a stock price which is like the best price that you
expect to get. There's a floor price before below which you don't want to go and then there's a target price right in the middle, right. And there will always be transactions that are below floor. There will always be transactions that are above
stock, right. And we were dealing with a customer who was like typically into selling raw materials and stabilizing products for food and beverage industry and they were like, hey, I want to build customer specific pricing for my customers because I feel like hey, we are losing business or we are not increasing prices fast enough.
And a typical implementation of customer specific prices is like people look at like, OK, what is this customer within this segment has paid and where are they related to the segment prices? They are like way below the floor price, maybe slowly increase them to floor if they're between floor and target, increase them close to target, right. So it's again a rules based pricing, not a super scientific way of doing it, right. And the customer was like hey, can you come up with a more
scientific way. So like what we implemented was like a very hyper focused segmentation based on not the segmentation attributes that we use to segment the business, but instead of like within the segment itself, what is excuse me a hyper segmentation that we can implement. And I used like machine learning methods for that hyper segmentation that I'm talking about. And we were able to leverage like their purchase behaviour, who are their end customer to assess like how much is their
willingness to pay. And in this example we learned that all things equal you have 3 customer, 1 customers. End user is like adult food industry like you and I, right? Eating like packaged food and then similar customer is making baby food and similar customer is making pet food. The willingness to pay for baby food manufacturer and the pet food manufacturer was significantly higher compared to adult food manufacturer. Oh wow, that's interesting.
And and in that case what we were able to do is like OK, instead of going by like where they are in the pricing segment, let's go by their willingness to pay. So we were able to give higher price increases to the baby and pet food manufacturers compared to like adult food
manufacturers. So like these are like hidden patterns that are like not visible or even like the traditional segmentation methods will kind of like fail at identifying these trends and patterns in the data which is like machines are very good as this. Sure, that's really cool. That's a good point.
You make leaning on machine learning for these algorithms and for these hyper focus segmentation components to further study the willingness of customers and how their their willingness to I guess purchase certain products. You know there's there's a lot that goes into segmentation. But can you highlight a few other instances where you effectively utilize, you know, AI or data science to tackle
complex pricing challenges? Because it kind of would be great to hear about the diversity of problems you you can address in this particular space. Yeah, absolutely. So I mean one thing I alluded to like customers, end users willingness to lean on data. So people are harvesting a lot of data and like previously like I said, you would have some transaction data, you would have some customer data and then things like that.
And then you build a model Right now with more and more data coming in and more and more products being available to the customers, the scope of data has been really exploded, right. And if you continue using the traditional methods, then you do take a significantly longer time to find pressing solution or recommendations. Like there are two challenges with the advent or or the adoption of the data culture if
you will, right. One is like the scale of the problem has become much, much, much bigger and the scope has become much, much, much bigger. So like there was a time when like I had a customer who had like millions of segments because they were talking like millions of skews, right. And with customers like those, what you have is typically you have 8020 rule pretty much everyone actually where 80% of your revenue comes from 20% of your products, right.
And if you were leaning on traditional segment, sorry, elasticity methods, right, let's just use elasticity as an example. We don't want to talk segmentation. And if you're trying to compute elasticity for those millions of segments, you're talking about running million regressions, right? And those are really time consuming. And again, like I talked about, they're very prone to outliers, 80% of your business, like I said, also 20% of products.
Still, that means that almost a large chunk of your business doesn't have enough transactions to accurately estimate elasticity. In that case, what happens is like if you're running the regressions, first of all for a thin or sparse segment where you do not have enough data, your results can swing from one way to other with like injection of
newer data, right? Like we can estimate like, hey, for this particular SKU, oh it's not elastic and then you have like few months of transaction data coming and all of a sudden there is few outlier and it goes from inelastic to super elastic. That is not good. You don't want any of those. So. And the second thing is like if you are having like too much products, compute elasticities like millions of regressions. It's time consuming right?
So to address these challenges we brought in like 2 methods, right? One method was to be able to work with less amount of data. So most of the data science algorithms are data hungry. They need a lot of data to give good protections. But there are certain algorithms that like deal with like less amount of data. So for example a segment doesn't have enough data.
Instead of like collecting data or climbing above the product hierarchy and computer regressions or elasticity at a product, some are higher in the product hierarchy, right? You could like look at the closest queues that are like similar to this queue we are dealing with which doesn't have enough data. So there are scientific hierarchical methods which will go up and up until it reaches a certain data density and is able to compute LSCS accurately.
The good thing about these hierarchical regressions which are like Bayesian in nature, you can have robust results which doesn't change with like advent of a few data points, sorry, injection of few data points in the model. So stable robust results. So that was one set of algorithms that we implemented like hey algorithms that are like less data hungry and still have robust results. The other question or problem was about scale, right.
So previously what you you would do in traditional methods is just rely on accurately estimating elasticity of top or faster in skews and just like give some elasticity number at a higher level for the remaining skews, right. So for you to be able to run that regression, you need a method. So like traditional older methods that and this elasticity is relied on algorithms that used to take a long amount of time. Let's just put like in some cases 50 hours for a million
regressions. Yeah yeah. And we implemented newer algorithms sampling techniques like. I mean, the new robot sampling technique is called like a new no U-turn sampling NUTS, NUTS is able to converge that same 50 hours of work in less than five hours. So it's almost like you'll see 90% reductions in runtime. And it's basically data science and machine learning algorithms that learn things faster and converge to a solution faster. So you can solve the problem of
scale. You can solve the problem of accuracy being affected by few data points using these techniques. I can imagine that comes in. That's very convenient in today's today's world as far as the amount of data we have in certain industries. Now you've been instrumental in assisting various companies including Fortune 500 companies and building and refining their pricing programs. How do you even measure the
tangible value of your solution? Like how do you measure what that, what that brings to the table? Are there any specific metrics or analytical frameworks you can share with us? Yeah, absolutely. I mean and honestly like in my experience, not much attention has been paid to this topic in the bTB space. Typically companies implement pricing strategies and don't really have a scientific method to quantify what is the benefit I'm getting out of this pricing program, right.
So, and if you don't measure what what you're, what you're working with, what is the benefit? There is no way, no good way to improve the that process, right, The pricing strategy and in essence of which you are just like you cannot continuously improve, right. The continuous improvement relies on effectively measuring it. And this is a problem I had to deal with multiple times. So I leaned on my operations research background to and to this background to solve for
this problem, right. So I leaned on traditional statistical techniques to develop store tests. So basically what you could do is, the crux of this method is, is that you can select a relatively small amount of products and a relatively small amount of stores to be statistically representative of the whole fleet or the whole business that we're talking about, right? And and then implement a pricing program or a pricing strategy in that narrow set of products and store combination.
Look at the benefit against another similar set of stores and products to see how much of A lift that you are getting out of it. And what happens from this is the key objective of this design is to achieving a kind of balance where the selected products and the selected stores are representative like I said and small enough so that if something goes S, they're not affecting the whole business,
right. And how you do it is you want to select products that are and needed to ensure equivalence across like multiple dimensions, like business divisions. Like if you have like multiple business divisions, you want to have equivalence. You don't want to select products that are top revenue across and a business division is not represented. Similarly, you want to select products across product categories, different product velocities, product profitabilities and metrics like
margin markup, even seasonality. So you can like slice the products based on those dimensions. Similarly stores you can slice the stores and select stores from based on region, traffic, profitability and and performance metrics like store labor, size of the store location, volume that that that store deals with. And then you can kind of form with those pieces of information, you can formulate the optimization problem like, hey, with these dimensions of products and excuse me stores.
Find me a combination that is representative. Why fleet in terms of volume, in terms of revenue, in terms of representativeness across the dimensions that I talked about and and and and it will not be a perfect match. Like for example, let's just say your PROC Division One is 50% of your business. These methods can get you to like 48 percent 49% representativeness right and and and once you have these you're able to get like a near
representativeness. And typically with these kind of tests you can simultaneously test like multiple pricing strategies and like in statistical speak they're like AB test. But the design of it is like a little more nuanced to achieve equivalence.
And yeah. And once you have these design, these tests, you can like come up with multiple pricing strategies, implement those at a smaller scale, understand the benefit out of it and then like depending on which ones are working versus not, you scale up right and quantify the benefit of it. And if it's not working, you can continuously assess like, OK, does my pricing program kind of needs a refinement because the benefit that I was seeing over a
period of time has kind of stagnated or like going down. So it just gives you a continuous read into how your strategy is doing and and then and you can continuously keep on refining it. OK. Now let me ask you this as well in correlation to what you just said, the selecting products that don't aren't too heavily involved with the revenue. You know if it doesn't sell too much, you know it's not going to make make the company go belly up. How often do you run these type
of tests? Actually you can run it once in a while like for example if you have a new strategy you can run this test like as a one go see like hey one time six weeks, 12 weeks test and you need to have like the product store and the duration and it's a statistical formula that feeds that and you'll tell, it will tell you like hey if you run these tests for these long number of days with these number of products and stores and you expect to observe X percentage of lift.
Here is what how long it's going to take. You can run these tests as a one off or you can have a continuous like a as they call it like a control study where you always keep a small set where you're not putting any of your current pricing strategy and then compare it and continuously keep on checking. That is what I was talking about. Like you can have these tests that can be one off to test the strategy or these tests.
This test could be like continuously running over a time horizon and you can continuously refine it. If there is like a product that is not included or is a new in your assortment, you can refine the assortment mix and then keep
going. OK, OK. And that's more of like a controlled testing that makes sense now as it pertains to adoption, you know, driving adoption especially within sales teams, it's often a challenge when introducing a new pricing strategy as many of us may may already know are aware of what are some strategies or experiences you have successfully navigated this hurdle and ensured seamless integration within teams. Yeah, this is just a good example actually.
Good question actually and a few that come to my mind and some of my customers have extensively spoken about it on public forums. So basically if you don't ask anyone, I mean there is always a balance between trying to get the highest price or not losing the business, right. So you have to incentivize those instead of incentivizing volume. For example, if your incentives or adopting the price is like, hey, how many transactions you can close, it's a race to
bottom. Everyone will go to the lowest price that you're recommending and try to close as many deals as possible. However, if you're able to incentivize like, hey, if you are like selling at a stock price, your incentive is going to be widely different than if you're selling at a floor price, right?
So we had some customers who were trying to optimize gross margin dollars while incentivizing number of deals done and that was like counterintuitive defining your sales compensation program is a good way to kind of Dr. adoption. The other is simply asking people for like why do they need a price exception. So for example, we had a customer where like they were seeing a lot of price override. And what they did was they added a simple widget where you have to select from like 5 or 6
regions. Why you think the price is not right and why you need the exception And you cannot leave others that you cannot leave blank. You have to give a reason. And those are concrete reasons. And all of a sudden the adoption increased because if you people can select others or not give a reason, they will always like oh right, right. And yeah, And I had one more
customer actually. They were like pretty good in the sense like what they did was basically when you have a price that is displayed for the salesperson and if you are not feeling confident about, there is a widget to explain, you hit that explain button. It's just like hey, your transaction is similar to these transactions that have happened in the last six months or a year and this is the price they have
realized. So it and that kind of gave confidence to the salesperson, like hey, the price that I'm going to offer is not going to end up in the lost business. So it was more like building confidence. So the multiple things people can do, but what I've seen is rewarding the right things and not rewarding the wrong things like other things that are not aligned with your business or pricing strategy is the most effective way. That's interesting, that's good to know.
Now I want to move briefly aside from adoption and just looking at beyond you know things as aspects beyond pricing. Are there other areas within the realm of B to B commerce where you foresee the integration of data science and AI playing a pivotal role? Any emerging trends or opportunities that may be catching your eye?
Yeah, Yeah. I mean, certainly and then like I have seen that happen with some of my customers and like a good example is just staying with the segmentation team, like you could use machine learning to cluster your customers into different profiles, right. And I'll give you a more
relatable example. So think about it like your customer has like builders, manufacturers kind of thing and then one of your customer base, you're able to identify like, hey, these guys seem to be doing transactions that make it look like they are floating contractors, right. So you're able to create personas of different customers that you have and machine learning algorithms. There are a bunch of those that can accurately classify your
cluster, those set of customers. Now think about just one customer, like floating contractors. Sure. Now in the floating contractor, what would a floating contractor buy? Will buy tile, will buy you reducers, transition pieces, stair noses, all the good things, right?
So you have a persona that hey, of these customers within my flooring contractor, a customer base, pretty much everyone buys 40 to 50% of their sales is, is on flooring type, 20% is reducers, stair noses, accessories, whatever you will. And then once you have that you can use again use machine learning algorithms, sort of even simple analytics really to say like, OK, are there any anomalies in these clusters?
Like is there a customer who is buying pretty much the same number of floor tiles or same percentage of floor tiles that other contractors are buying, but is not buying reduced transition pieces, accessories, right. And you can identify areas of growth there. So that's that's like a very good use case of like clustering similar customers together and then within that cluster identifying anomaly to find areas where you can grow your business that's. Cool.
Yeah, yeah. And then there's another example. Actually, I worked on this one with another customer. So I had a client who stocked like standard parts and also built custom parts like electrical panels and and things like that, right. And many times like customers would come to them and say like, hey, I'm looking for this part XYZ part right.
And that would often be the case that they didn't have the exact same part but they had like a perfectly good substitute of it in the stock and it was very hard for them, manually cumbersome and error prone to have that mapping. Like hey, I have this product, here are the 15 other products that are a perfectly good substitute and can be used in
its place. But for them to manually maintain that mapping across multiple products and as their assortment grew or the product offerings increased, it was becoming harder and hard for them to keep that mapping or like maintain the list of substitute products that could go right. So like I worked on like building a machine learning solution for them that will create this mapping by kind of identifying product attributes. Like you have a free form text description like hey, this is X
world, switchboard, whatever. And then you have product attributes, descriptions like 3 phone texts like I talked about. And it would create a similarity score based on like, OK, these products are similar to this product and this is how, like you'd create manually create a mapping but give that rule to machine to be able to faster and accurately figure it out. And that would create like a good list of substitute products and in this case, what they have to do.
I mean of course you cannot take human in the loop, like human out of this loop. Sure. But the task reduced from like I define products to like looking at, hey, here is my machine's recommendation. Yeah, 90% of these are fine. These percent are not right. They introduced their workload and it was like a good use case of machine learning and making your product offerings more robust, right. OK. Yeah. That's good.
Now Mr. Vivek Anand with Gap, I want to, I want to close out and thank you so much for your time today. Before I let you go, you you did provide a lot of key insights, a lot of knowledgeable wisdom if you will in the AI and pricing and data science realm as regards to bTB pricing lease for those who are listening and are are one are interested in learning more about you and you know your resources, what you stand for maybe where can they go to learn more about you.
That's a good question. I mean, like, I'm available on LinkedIn, so like a good place is to just connect with me on LinkedIn and like research to me with your questions that you have. And I'm more than happy to engage. But yeah, I mean I also am looking to like create a forum for myself. Like I like I mentioned previously, I not only deal with pricing, I also deal with inventory like buying and like Internet issues which are again like pretty prevalent in the B
to B space. And I'm going to like I'm planning to actually. So I don't have a name yet. I have a a domain of myself where I will be like sharing these ideas, but LinkedIn is probably the best way to reach out to me or communicate with me. LinkedIn and and kind of keep up with you on the latest updates about this project you're you're working on personally that that might be really cool. So staying, staying to know, I'm sorry, go ahead. Yeah, I think that, yeah,
absolutely I can. LinkedIn will be the first place where I'll like talk about this project where I have like these things like codes and and use cases and things like that. OK, OK, cool. Well, for those who are listening, stay in tune because he has great things in store for you all in the times to come. Until next time. We'll see you all later. Bye, Bye.