This is an audio conversation from the CIO Talk Network podcast. The timestamps in the chapters and transcript below refer to the audio recording.
About this conversation
A CIO Talk Network conversation with Mark Miller (Party City), Prasanna Kumar Burri (Dangote Industries), hosted by Sanjog Aul.
What the conversation established
The conversation established that scaling AI is less about novelty than about discipline. Mark Miller of Party City said there is no such thing as an AI business case, only a business case that uses AI methods, and that the data pipeline is the foundational prerequisite, since models depend on data that staff have already explored, transformed, and validated. Prasanna Kumar Burri of Dangote Industries cited a survey of about 800 organizations in which only 10% saw tangible results and about 15% got a lot done, and said adoption is strongest in customer-facing domains such as CRM, call center and inventory optimization, cross-selling, and chatbots. Both agreed AI is being democratized, with Burri expecting about 70% of businesses to rely on productized solutions over the next two to five years, and Miller urging leaders to set expectations and allow some failures on the way to success.
Host Sanjog Aul talks with Mark Miller, Chief Information Officer at Party City, and Prasanna Kumar Burri, Group Chief Information Officer at Dangote Industries, about the steps to scaling AI from a prototype into an enterprise-grade pipeline. They compare why adoption stays low, argue there is no separate AI business case, and point to the data pipeline as the first prerequisite. They cover where scaling works today in CRM, chatbots, and inventory prediction, how to fund and govern the effort, and the intellectual property questions AI raises.
Contributors
- Mark Miller, Chief Information Officer, Party City
- Prasanna Kumar Burri, Group Chief Information Officer, Dangote Industries
Host: Sanjog Aul, Founder and Host, CIO Talk Network
On this page: Key takeaways · Chapters · Explore more · Transcript
Key takeaways
- Mark Miller, Chief Information Officer at Party City, argues there is no such thing as an AI business case, only a business case that happens to use AI methods. He says the project still has to justify itself on its reasons, technology framework, delivery method, and expected timeline for return, with the business owning the rationale while IT owns the data and pipelines.
- Mark Miller of Party City identifies the data pipeline as the first prerequisite for scaling AI. He says organizations that already staff data exploration, transformation, enrichment, and validation have taken the big first step, because a basic machine learning model cannot be labeled or built until that foundation is healthy.
- Prasanna Kumar Burri, Group Chief Information Officer at Dangote Industries, cites a survey of about 800 organizations in which only 10% were somewhat successful with AI and about 15% got a lot done. He concludes that adoption remains a niche concentrated in customer-facing, revenue-affecting areas such as CRM, call center and inventory optimization, cross-selling, and AI-based chatbots.
- Prasanna Kumar Burri of Dangote Industries says intelligence is being democratized and expects about 70% of businesses to get by on traditional productized AI solutions over the next two to five years. He notes his own manufacturing organization has the ingredients, large-scale ERP and a hybrid cloud, but that mobilizing the business is still a work in progress.
- Mark Miller of Party City says patents are reshaping how AI teams are built, which is why lawyers are now among the key positions. He points to patent applications that name AI technology among the inventors, raising unresolved questions about whether intellectual property law extends to machines that independently develop new algorithms.
- Mark Miller of Party City says AI governance should reuse what organizations already know, applying the same agile methodology, QA process, and product-owner roadmap used for any production system. Prasanna Kumar Burri of Dangote Industries adds that socialization, aggressive funding once a business case is identified, and well-rounded teams of data scientists, analysts, and communicators are the key ingredients.
Chapters
- 0:00 Introduction: scaling AI from prototype to enterprise pipeline
- 2:49 Why AI adoption stays low
- 4:18 Making the business case for AI
- 9:14 The human side and the computerization parallel
- 13:00 Data pipeline as the first prerequisite
- 14:54 Only ten percent see tangible results
- 17:43 Where scaling works today
- 21:25 Funding, data foundation, and skills
- 28:42 Intellectual property and patents
- 34:18 Democratization and productizing AI
- 46:08 Governance by reusing agile and DevOps
- 49:13 Paying for AI without short-term ROI
Explore more
- Steps to Trustworthy AI
- Steps to Maximizing the AI Opportunity
- The CIO as the Conscience of AI
- Preparing Data for AI
- AI in IT for 2024
- Is HR ready for AI?
Transcript
Introduction: scaling AI from prototype to enterprise pipeline 0:00
Sanjog Aul [00:00:23]:
Today’s topic, steps to scaling AI. So it’s one thing to have an AI concept, dream it up, then you develop a prototype, and then you play with it and eventually take it to production, but then what if you wanted to scale such an effort to build a very strong enterprise-grade AI pipeline? What would you do with the different functions you have starting from the ideation, the innovation function, to product engineering, to product life cycle management that you might be doing in other parts of your business, and even the governance? And maybe there is another discipline which I may have missed out to scale your AI effort so that we have consistent AI-related pipeline being built, and each of the items or the capabilities we put out there for our enterprise, using AI, it always delivers well, reliably, and has a strong performance. So to discuss this, we have Mark Miller, CIO with Party City. Hey, Mark. How are you?
Mark Miller [00:01:41]:
I’m well. Thank you for the invitation.
Sanjog Aul [00:01:43]:
Great to have you, sir, and we also have Prasanna Kumar Burri, group chief information officer with Dangote Industries. Hey, Prasanna. How are you?
Prasanna Kumar Burri [00:01:53]:
Great, Sanjog Aul. Thanks for having me. Hey, Mark.
Sanjog Aul [00:01:56]:
Great to have you as well. So, Mark, let’s start with you. AI, as we know, we are pretty excited about it. Something new and cool we get to do with it, but earlier when we started, we said, oh, it’s just a concept, and it’s almost in the labs, and then we, over time, started experimenting with it, putting some products out there, but then still when we go out to people, they are still kinda fuzzy and hazy about how to take it to a level where it becomes an inherent capability for a company. So what do you see? You talk about your organization or you talk about the industries in general. To what degree has this excitement been converted or channelized into some stable ongoing pipeline building for AI?
Why AI adoption stays low 2:49
Mark Miller [00:02:49]:
Well, I think the adoption’s been fairly low. Like you said, people have a lot of success building these models that they go ahead and they find all of the features they want to model. They pull all of the data in. They build the model, and then they have some success and that’s something they want to implement, and what do you do next? I think a lot of our organizations really aren’t scaled for that type of success. I think also the organizations have a great deal of difficulty building that level of evangelism that you need to really sustain an AI methodology or an AI set of activities within not just the IT organization. This is an interdisciplinary sort of engagement where we need a lot of help from the business SMEs. We need an organization that’s really dedicated towards the procurement of AI assets and the maintenance of AI assets. So I think a lot of organizations are really struggling with that right now.
Sanjog Aul [00:03:56]:
So, Prasanna, when you look at what’s going on with AI, we understand that AI primarily should be a business capability. The outcome of that will also be utilized by the business folks, but they cringe at the very idea of dabbling in it because it looks so far and so alien. So how do people like you deal with this?
Making the business case for AI 4:18
Prasanna Kumar Burri [00:04:18]:
Sanjog Aul, this is nothing new, in terms of making IT business cases. It’s always an uphill battle, unless it’s a very direct technology like CRM. You don’t need to convince many businesses, or a business warehouse. Perhaps, 15 years ago, it would have been an uphill battle. So AI being on the early part of its adoption, it’s only certain leaders who are embracing it more actively. So it is going to be a challenging business case to say the least, but the most important thing is in cultivating the ground, what are we trying to really accomplish? Is AI something totally different from any of the IT tools, solutions that we have been promoting in the past? Not really. This is another way of doing things better, rather radically better. So the key thing is that we get the socializing of AI, machine learning, data mining, process mining, all these concepts, the business owners must be acquainted with. Fortunately, business is more savvy than before. That’s a good sign, but, again, it depends on industry. If you are in banking, retail, FMCG, the chance of adoption is much smoother. It’s much more tangible. Other industries, for example, I’m in manufacturing, and we also have a bit of transportation, FMCG, a bit of different things. So we are looking at some really impactful business cases. Even so, it’s new. It’s always that gap that we need to bridge. I would say, while we take the traditional approach of making business cases, value propositions, getting the analyst reports, and so on and so forth, the social aspect of delivering this technology is quite significant.
Sanjog Aul [00:06:46]:
Mark, how much should we consider AI? The way Prasanna mentioned that it can be seen similar to what another IT project. Or anything new that you put out there, business always cringes because they feel, man, this has come out of nowhere, and now we have to deal with it. So they say, okay, folks, IT from IT, go ahead and deal with it, but this one, can they afford to stay away from it even though they cringe? And can IT just take this as yet another new technology? Because this is kinda different. What do you think? How would you categorize AI?
Mark Miller [00:07:25]:
I think to add to what Prasanna said, there really isn’t any such thing as an AI business case. There’s a business case that may employ some AI methods, but the gold standard still remains focusing on what are the reasons for the project? What are you doing it for? What sort of technology framework will you utilize? How are you going to deliver this project? Are you going to deliver it leaning towards agile or towards a waterfall methodology? And what’s your expected timeline for return? I would agree that 15 years ago, maybe you would have to do some evangelizing around the benefits of AI, but I think most companies that are successful in this really have the business own it, and it’s not really an IT owned asset. We own the data, and we own the pipelines, but the business really owns the rationale around the roadmap for what AI is going to do to bring value to business.
Sanjog Aul [00:08:32]:
Prasanna, when you look at this whole AI-related capability, one of the biggest sticky points for many was it’s a replacement for human beings doing certain processes using their intelligence and their ability to do some labor. Have we gotten over this? And as a result of it, have you been able to scale it? Because the more you scale, there is a concern among humans, the very humans you want to have participate in the scaling effort.
The human side and the computerization parallel 9:14
Prasanna Kumar Burri [00:09:14]:
Yes. Sanjog Aul, if you recall, back in the 1990s, when banks started adopting computing, computers, desktops, UNIX mainframes, there’s a lot of pushback. In some countries I’ve seen, there are a lot of unionized activity to resist the computerization back then. We can see similar concerns, especially around ethics and the use of data. It’s a slightly different shade. It’s not identical, but it’s similar. So, I think we have to still address the questions around how AI is going to be deployed, how it’s going to impact the human dimension. I think it’s something that needs to be addressed. How these automated activities, there’s some intelligence, it’s doing some situational response based on certain criteria. So we need to make sure that whatever AI we are trying to deploy is, first of all, for ease of adoption, we should pick business cases in such a way that they are pretty much no-brainers where AI can make an impact. Areas like logistics orchestration. To have AI do those things brings a sigh of relief to many line managers, which otherwise relies on a bunch of manual and system-centered tasks. In my view, the human side cannot be understated, but I think the innovation side of this, people are getting pumped up. The largest search company that we all know is emphasizing AI technology significantly, and they have been demonstrating certain business use cases around how AI can speed up the diagnosis of certain types of diseases that would take days. Sometimes you may not even get an accurate diagnosis from traditional medical practitioners. So there’s some impressive things. Each one is a unique case, unique scenario. We have to factor in the human side of it, both in IT as well as in business, but I think it’s just another new phase. People will find business to let out, just as we did, albeit in a slightly different way. The change will come, and we will survive.
Sanjog Aul [00:12:19]:
Mark, just to let you know, we have been covering AI while it was in the labs. These similar questions were asked, and some of those problems still stay. They are chronic or they are tough to tackle, and that’s why they still exist. Since the topic is steps to scaling AI, which means, yeah, we got the fundamentals in place, hopefully, so that we are starting to crawl, and here we’re talking about scaling. If you were given an option to list out prerequisites which will be absolutely needed before someone can even think about scaling AI in their company, which ones would those be?
Data pipeline as the first prerequisite 13:00
Mark Miller [00:13:00]:
Well, I think first and foremost, we can’t underemphasize the foundational aspect of what we’re doing here. So for me, it’s the data pipeline. It’s whatever you’re building right now for your data pipeline for things like data warehousing or doing data lakes or just the basic health and munging process of your data. So if you already have people on staff that are doing the data exploration and the data transformation and the enrichment and the validation of a lot of your data, then you’ve automatically taken a very big first step to making that data available for the data scientists to begin to use. You’ve allowed them to really start the first step of, in the case of, say, a basic ML model, they can begin labeling because you’re working now with SMEs that have already been through the first step of making sure that that data is of the highest quality. The next logical step before you begin to build the model is how do you label it? How do you build your features? If you don’t have the foundation, if you don’t have those datasets available, you really have to go back to the beginning and make sure that the foundation of your data is healthy.
Sanjog Aul [00:14:26]:
So, Prasanna, when you look at where we are today, we used to do maturity curves or we used to measure the maturity of an organization in a given context. What would you say about the maturity levels for most organizations without grossly generalizing them? Are they ready to scale their AI efforts, or is there some ways to go?
Only ten percent see tangible results 14:54
Prasanna Kumar Burri [00:14:54]:
Yeah, Sanjog Aul. I was recently reading some survey by a leading management consulting firm. They have surveyed about 800 organizations and they found that only 10% were somewhat successful in getting tangible results out of AI deployments, and out of that, about 15% really got a lot done. So the success rate, the adoption curve is still on the rise, but it’s not where it should be. To me, it is a serious niche as of now, and it’s also industry specific. Where you are looking at is mainly customer-facing domains. Things that directly impact revenue are the ones that get a good deal of traction. AI-driven customer care is a very early adoption business case. Call center optimization is interesting. Inventory optimization is great, and then cross-selling, upselling, AI-based chatbots. These are typical ones that we see where a lot of business are getting value. As far as other traditional nuts and bolts business operations, it’s still on the low side. One thing that is really pumping up businesses is competitive adoption. We see the competition embracing certain AI-driven technologies, and they see them in action. We find them coming back and asking, hey, how did they do it? Why can’t we do it? Why aren’t we doing it? Those are typical conversation starters for AI. I think that’s where the business should focus on to begin with.
Sanjog Aul [00:17:08]:
Mark, what have you seen that has worked for companies? Because, of course, Prasanna, you gave that, small percentage of those 8 or 10 companies who were polled, they found that they were not as successful. Here, we’re talking scaling. So, yes, there will always be those companies who have not been successful, but that should not make a doom and gloom situation. What would you say have you seen that has made you successful? Where are you, by the way, Mark, in your journey to scaling these efforts? And what have you seen working, and what’s turning out to be a damper for you for these initiatives?
Where scaling works today 17:43
Mark Miller [00:17:43]:
Well, we see a lot of the same things that Prasanna just mentioned. It really is dependent upon the organization. Where we’ve seen the most penetration, where we’re seeing the most maturity and scaling is around CRM, AI-enabled chatbots, and moving very quickly into using AI to help with the predictive modeling around inventory. Those are the use cases that when you’re trying to manage your risk and you ask other companies what are you doing, where are you successful, you’ll find that CRM and inventory management are the two primary areas where we can use other companies as an example of how they scaled that and how they maintained it. I think also, typically, in the world of CRM, our data models are more mature. We’ve already done a lot of that foundational homework that may not be present in other areas of businesses. From an adoption point of view, especially around the evangelism and the enthusiasm we get from our business units, CRM marketing and ecommerce are leading the fight.
Sanjog Aul [00:19:10]:
Prasanna, how about we talk about the very foundational elements that we spoke about. We spoke about data infrastructure, the models, and everything that you need for AI to be successful, but are organizations willing to spend that much upfront with the confidence that they’ll be able to scale? I’m sure they’re beyond the experimentation stage, but to what degree are they investing, and how are they trying to build that capability? Are they building, buying, leasing, outsourcing? What are they trying to do?
Sanjog Aul [00:20:46]:
So, Prasanna, we need strong foundation, and data, of course, is going to be key. The type of infrastructure for computing and the models that you would need for AI to work and deliver the results we expect. How are organizations approaching this capability building without truly knowing how much investment will lead to what ROI given the newness and the flavors that keep popping up on the way you can deliver value using AI? So are they building, buying, leasing, outsourcing? What are they doing? Or what makes sense at different stages?
Funding, data foundation, and skills 21:25
Prasanna Kumar Burri [00:21:25]:
Sanjog Aul, in my conversations with some of my CIO friends and industry analysts, personally, I have to say that we are on the adoption curve. We are working on a few business cases. We don’t have any active AI scenarios that business is using today. We use some very large scale ERP solutions. We are dispersed on cloud. We have a strong hybrid environment. We have the ingredients. We have the foundation, but the mobilization of business is still a work in progress that keeps happening. My observation from other businesses that have done it in a reasonably successful manner is they are really innovators. They are aggressive. They really set the example. You have to exclaim, wow, you can do that too. Aggressive funding is really important. Once you identify the business case, you have to really ramp it up very quickly. Strong data foundation. Those that have built good data warehouses, good databases, good master data management, they’re able to quickly leverage and minimize the confusion with AI. Another very important dimension is having well-rounded skills: data scientists, application innovators, business analysts, communicators. Those are really key ingredients to have. Once you start, what I’ve seen is that it’s quite agile. Don’t try to build the perfect solution on first run. It’s going to take some iterations. It’s going to take some learning. We’re talking about machine learning here. To me, it’s proper identification of business cases based on the industry segment that you’re in. There are quite a few templates. Two years ago, we had to really do some serious research. Today, we see quite a few product offerings from many leading providers. Many of them have deployed AI, machine learning, RPA frameworks in their cloud offerings. If you are doing some colocation or platform as a service, you have options to plug into off-the-shelf APIs that do AI and machine learning. They are also beginning to provide more and more templates. You go to any of those AI solutions and you see a few sample scenarios. If you have to really ask yourself what can we do with AI, someone savvy can do a little bit of reading on plenty of analyst information content, podcasts like this. They can give you plenty of insight on what gets started. I think that would be a good starting point and socializing with the business stakeholders, carry them along. Walk them through what you’re trying to accomplish. Give some visual of what others are accomplishing with AI. To me, those are some key steps that we can take to begin the journey of adoption.
Sanjog Aul [00:25:21]:
Mark, would you have any ideas about what would prevent unnecessary bruising and pitfalls in building this capability, or are some myths that we see that are hurting people?
Mark Miller [00:25:37]:
It’s very hard. This is a cultural change. You’re trying to advise the people that make decisions in your business to move really from an experience-driven model to a data-driven model. There’s always a little bit of distrust there because people are moving away from something that they’ve been doing for years and putting their trust in something that they don’t fully understand. So from a cultural point of view, we really do have to begin the journey with explaining what the advantages are with a properly trained model in providing a lot of data-driven analytics. That needs to start well in advance of any company that wants to go down this route and do a whole large scale adoption of AI. On top of that, from an organizational perspective, you need to build a center of excellence around this capability, and this isn’t just data scientists and data engineers and IT DevOps because you, of course, want to productionize everything. These are business units, domain experts that come together, and the product owners that really own the roadmap of what are the key measurements that I want this AI model to derive. From an HR perspective, what is the effect of AI on the business and how do we deal with people? We spoke earlier about people potentially being nervous around this model removing them from their livelihood. Sitting down and really beginning to talk about why we’re doing this, why we need access to better decision making processes and how this is helpful for everybody to perform their jobs and make good decisions, I think you just have to start there before any technology gets laid down.
Sanjog Aul [00:27:53]:
So, let me ask another question, Mark. We want AI to be utilized wherever possible. Automation can lead to cost savings, which is, of course, powered by AI. If you build some good models and it gives us the insights, we can also use that as a competitive advantage. How should an organization build this capability without losing the properly guarded intellectual property they had been building in the past, which they may put in the machine, which then is not available openly but still it’s in a digital form now? Or how do they use AI to build intellectual property, period?
Intellectual property and patents 28:42
Mark Miller [00:28:42]:
I’ll answer the first part of that question. For me, when you read a lot of the opinion pieces on how to properly build an AI team within your organization, one of the key positions they mention are lawyers. That really is because when you look at some of the patents that have been filed, this is the first time we’ve seen several patent applications where the applicant has actually declared that AI technology is part of the group of inventors. I think that’s opened up a whole discussion around does intellectual property law extend to machines that are building new algorithms that are finding better ways to do things, or can they be patented at all? What happens if a machine in its purest form of an unsupervised learning model develops something independent that potentially impacts or interferes with somebody else’s patent? How do you defend that? How do you actually patent that? That’s a discussion I think we’ll be having for years, because it does really bring up the question: are humans the only ones capable of inventing patentable things?
Sanjog Aul [00:30:17]:
That’s a pretty interesting angle that you bring up. So are you saying the attorneys will be richer than the companies who build AI?
Mark Miller [00:30:24]:
I would say that having read the latest feedback on this, I’m going to make sure our legal department is always present as we introduce new technology around AI.
Sanjog Aul [00:30:39]:
And with that said, Prasanna, a very logical question would be if anyone is trying to talk about intellectual property and they are also then trying to bundle AI as easily consumable products, how the productization of AI and piggybacking on the previous response from Mark about the intellectual property, what would make sense? How do we productize AI? Should we only start moving towards eventually building AI into a product so that it becomes commoditized, even though it is used only internally? And all this while, how do we handle these other software issues which are related to intellectual property and the ongoing product management, etcetera? It’s a loaded question, but I’d like you to take a stab at it.
Prasanna Kumar Burri [00:31:32]:
I’ll try, Sanjog Aul. Indeed, it is a very complex issue. A lot of ethics and privacy and compliance borderline issues. As Mark said, this is an evolving topic, and it will go on for quite some time. Given where we are, in my view, the traditional IP protection and patenting approaches would still apply. You bought an AI solution. Of course, you have to also look at the licensing, whether you built AI from scratch or you are leveraging a productized off-the-shelf version. If it is off the shelf, then you have to look at the implications, whether there are any exclusions. For the most part, you have to have ways to protect your intellectual property. Each unique thread, for the most part, can be patented to some extent. To what extent are we going to patent? Ultimately, we’re going to be looking at outcomes. The nature of technology is that the competition will eventually figure out a way to achieve similar outcomes. At some point, unless you are a technology company or a patent attorney, you’re not going to spend too much time trying to go after competition doing similar things. It will be some concern, but I would say eventually the focus is going to be on the outcomes, on the deliverables that you are trying to achieve from a deployment of AI. It’s an open-ended discussion. That’s how I see it.
Sanjog Aul [00:33:46]:
So, Prasanna, do you see productizing AI as the simplest recipe to scaling AI efforts? Because if you can do that, if you can just focus on productizing anything that you’ve envisioned, then you can use the age-old mature principles which are used for product lifecycle management and other related things that you can do with it. You could apply them, and you’ve essentially put that in a box and keep building boxes and keep scaling. Is it easier said than done?
Democratization and productizing AI 34:18
Prasanna Kumar Burri [00:34:18]:
It’s getting easier as days go by. Two to three years ago, AI was more on the innovation side. Now it is getting into mainstream adoption. We’re seeing subtle but ever-present application of AI. You call your airline, you chat with your booking agent. Many places you are seeing AI-based responsiveness and decision making that is happening. We are not noticing it, but it is there. I see a lot more productized AI happening. It’s not every organization that is going to set up a dev team and data scientists. There’s a lot of democratization going on. That’s foundational: intelligence is going to be democratized. In the past, you had to rely on the smartest guys to give you the insights. I think it will be more democratic where AI-delivered outcomes are going to be adequate for most businesses. There will be a few leading edge innovators who will always have stronger in-house teams, centers of excellence that do more cutting edge things, but for 70% of businesses, they’re going to get by using traditional productized solutions more and more in the next 2 to 5 years.
Sanjog Aul [00:36:06]:
So, Mark, good old days when we used to talk about COTS – custom off the shelf. Then people said we want to build something custom to us because that’s our source of competitive advantage that allows us to streamline our functions. So are we going full circle with AI? Are we saying, again, we will leave the art and the science of building AI or converting AI into a product to somebody else, and then we’ll wait for that to mature for us to use it, and all this while, since we are not planning to hire developers and the data scientists, we will just stay put. How would you do with your organization? Would you be willing to pause your efforts because productizing is the only way to scale it and products are not available and so on?
Mark Miller [00:36:54]:
No. I would agree with Prasanna that we’ve seen a lot of democratization. If you look back to the early days before AI was really prevalent and we tried to codify a lot of this advanced work through rules engines and things of that nature, we really started at the low level of maturity with building and finding people within our organization that can build those rules-based engines. Then when it became more mainstream, you could buy that, democratize a lot of that and remove the need for a lot of really high-end, very expensive rules engine developers. I think we see the same thing especially with supervised learning models. A lot of it you can go out and buy and to the point where you do need to wrap it into an existing product. We’re not doing anything differently to manage this product than we would any other traditionally written product. I think about the only difference in maintaining a model-based product like this would be things like model drift. If you’re building in an agile, test-refactor type of methodology, you’re always going back and retraining a model. If it’s drifted off its intended results and you’re making changes to it, I think where you still see a little bit more investment in internal resources is really on the unsupervised model because unsupervised models are very data hungry. They want a lot of history. They need a lot of compute power, and those aren’t as readily available to buy off the shelf as a supervised model would be.
Sanjog Aul [00:38:59]:
Prasanna, since you mentioned products, how about talking product engineering or product lifecycle management and those similar functions which are very mature and actually will perhaps get there. So either we take the ownership or we collaborate with other companies who want to take this as, say, a technology company or a consortium who would take these AI capabilities, put them in a product, turn around, and give them to you for you to use. Whoever ends up doing it, whether you are collaborating or you’re giving guidance or you’re the one consuming it, what skills, mindset, and collaboration communication needs to happen fundamentally so this ecosystem is developed?
Sanjog Aul [00:40:45]:
So, Prasanna, product engineering, product life cycle management. There could be a lot of different functions which are mature and could be utilized to formalize the very process of ideation, prototyping, operationalizing, versioning, maturing. There are a lot of them out there, and that could all be utilized very well as we have done for other type of products into building the AI pipeline, assuming AI can be productized. So whether you do it, you get a partner, you collaborate, you co-source or outsource, still some capabilities will be needed, and if this is so new, not sure if all capabilities exist. So what skills, what mindset, what collaboration, what ownership needs to be in place for us to see this dream come true.
Prasanna Kumar Burri [00:41:36]:
Sanjog Aul, in my view, whatever good things that we’re doing with DevOps and agile only get an accelerated kick with doing things like AI. Many AI projects are 3 to 6 months. You don’t see any of these projects lasting for more than 6 months. If we are doing like that, then I think you’re on to some serious innovation research. The traditional productized approach in our manufacturing organization we take for our solution: we have internal applications that we developed which we figured that nobody could do better than ourselves. Then we have a bunch of off-the-shelf products that we use. Generally, we take the product approach for the ones that we are building or doing some major adaptation. Assuming that it’s going to be a hybrid, I think there is going to be a decent amount of building involved, especially at this point in time. A product-centric approach requires a product manager who is not necessarily an AI geek or a data scientist but a business analyst or a project manager with an edge on AI and product management principles. That’s the kind of profile that can drive AI-based product delivery. Productization of many of these newer technologies is like a benchmark for success. You have to make best use of what we have learned over the years with mature organizations that have DevOps and agile becoming part of the culture. We know how to leverage development platforms online as well as offline. We have evolved in quality assurance and testing and user adoption approaches. It’s a combination of all these: a bit of cost management, a bit of expectations management, plenty of change management and UX sensibility. All these ingredients come together with some agile scheduling and iterative delivery. Those are the wanted features for delivering productized solutions. The product can also be an external product that you’re embracing off the shelf. Even that can use some of this. What makes us different is the data part of it, as Mark mentioned. As you evolve data, how are you going to build it into some site products? Sometimes things keep evolving with AI. How do you interconnect different existing data platforms? I think productized approach, inside-out now combined with mature DevOps and agile is, to me, a key ingredient for delivering it.
Sanjog Aul [00:45:45]:
So when a product or a process becomes mature, then you know what it takes to run it the way it should be run, but since AI is a moving target, Mark, how on earth could someone come up with a governance structure without knowing what that maturity level looks like?
Governance by reusing agile and DevOps 46:08
Mark Miller [00:46:08]:
Well, we have to look to what we know already. Like Prasanna had mentioned, we’re already building production-capable systems on our own. Whether you are doing integration between a couple of different ERPs or whether you’re in an organization that’s gone ahead and said we can’t find anything off the shelf and we’re going to build it, the same governance that you used to build any product you would also use to build a machine learning model. What it comes down to is when you’re leaning towards an agile methodology, the business is really responsible and subject matter experts are responsible for identifying what are the things that we want to measure and how we want to use them. When you build a governance model around that, look at what’s the responsibility of technology and what’s the responsibility of the business. The business still has the domain experts as it would in any other agile project, and they have their product owners that are responsible for the roadmap. Those are the people that are basically telling you this is the type of data-driven decision making that we need. From that centralized technology hub, you have your data scientists who are essentially your architects designing the models and handing them over to your data engineers as they’re building the models. Those are essentially your developers. Once those models are built and proven and tested using the same agile methodology and the same QA process that we’ve done in years past, that moves to the DevOps team to begin to build. So we look to what we know already, and we have embraced agile as a methodology to allow us to productize different parts of what we thought needed to be codified and made part of our business process.
Sanjog Aul [00:48:29]:
So, Prasanna, fundamental question is that most organizations don’t say that their AI initiatives have really been bearing huge fruits. Sometimes the outcome and the value creation may take longer than expected, and all this while, you also are saying, let me scale the AI efforts, and let me put governance in place. Where’s the money going to come from when we cannot show ROI short term, midterm, and the only thing which we see is long term?
Paying for AI without short-term ROI 49:13
Prasanna Kumar Burri [00:49:13]:
Well, I think the key requirement here is socialization. That’s where the leaders of IT, the data organizations, technology leaders in the organization come to play, to educate the business owners and keep them abreast of what is happening if they don’t already know it. That communication aspect is really key so that we have a common vision and goal. Ultimately, we are an IT, a technology organization. We’re not doing it for ourselves for the most part. Ninety percent is directly for business outcomes. Setting that vision goes back to the same old conundrum with convincing the business to adopt certain pieces of technology. Interestingly, we end up becoming the torch bearers in many cases because of proximity to some of these technologies, at the expense of appearing like nerds and cool kids trying to buy all the new toys. The key thing is selling the value. It’s driven by the value and by what competition is doing. People have to make a choice: do we want to be left behind? Do we want to be in the race and be in that elite club of adopters who can make the best out of this technology? I think we don’t need to look very hard for business cases. Two to three years ago, perhaps that was the case, but now it is much easier to adopt and sell, there are some ready-to-corner use cases. You need to start building a serious data science team, analytics team if we are looking for something really bleeding edge. I see a significant change from my vantage point that a lot of this AI technology is getting democratized and it’s much easier to sell. You show some of those large hedge companies’ videos on how they’re showcasing some of the consumerized use cases for AI and what goes on behind the scenes, the power of big data at work and automating it with machine learning. That algorithmic power is impressive. We don’t really have to struggle as hard. Many times, AI business cases make people exclaim, wow, is this possible? It’s not too hard once you’re able to catch the pulse of the right business problem and put some PLCs. To me, it’s a bit of traditional plus a bit of more socialized internal social engineering, if I may say so. Sanjog Aul, you’ve been asking us so let me ask you: how do you see this going? What have you seen especially in the last 2 to 3 years? And where do you see this going in the next 2 to 3 years?
Sanjog Aul [00:53:03]:
Well, people are going to use AI as the best way possible to make their life better, their business better, and they will try to navigate through the haze like they did with digital. There is promise ahead. It’s more about scaling and doing it methodically, and that’s where the topic was brought up. I have only, I guess, 30 seconds to ask you, Mark. What would you do as a leader to be better prepared for scaling AI? Thirty seconds or less.
Mark Miller [00:53:38]:
I think setting expectations. We talked at length about the business case being a pretty traditional business case for AI, but it does have some additional challenges and complexity because there’s some unpredictability that you can’t explain away. Most agile organizations are prepared to fail fast at times, and failure becomes part of our learning curve. For me, just explaining to people that there are going to be models that we build that don’t do what we want them to do or that require a lot more work than we had initially forecasted, and that it’s okay to have some failures on our way to success.
Sanjog Aul [00:54:31]:
Once again, thank you so much, Mark and Prasanna, for sharing your insights about how organizations can take confident steps to scaling AI. Thank you.


