Sponsored by Infosys | Beyond Dashboards: Operationalize AI Agents with Databricks Intelligence

Summary

Adidas and Mondelez partnered with Infosys to move beyond dashboards to autonomous AI agents built on the Databricks Data and AI platform, deploying governed, real-time decision intelligence across core business functions.
Mondelez ran 35 Genie pilots across seven functions and four regions, while Adidas focused agentic AI investments on customer support, personalization, and engineering use cases.
A DRRI (Data Reliability and Readiness Index) framework at Adidas ensures data quality and trust before agents are granted access, creating the measurable foundation required for ROI quantification and margin uplift.

Chapters

FAQs

Why are most AI pilots failing to scale into production?

This video identifies a deployment gap where technology is evolving faster than enterprise adoption, often because pilots lack a clear business value attachment or a path to industrialization. The speakers argue that the next 12 months will be defined by which organizations can demonstrate measurable business value from AI investments.

How is Mondelez using Genie on the Databricks Data and AI platform?

Mondelez ran 35 Genie pilots spanning seven business functions and four geographic regions, using natural language interfaces to give business users access to governed data. This video presents this rollout as an example of scaling agentic AI beyond a single domain into a cross-functional deployment.

What governance approach does Adidas use before deploying AI agents?

Adidas built a DRRI (Data Reliability and Readiness Index) to assess and certify data quality before granting AI agents access to any data asset. This framework creates the trust foundation needed to educate business users on probabilistic answers and to justify ongoing AI investment with quantifiable ROI.

What role does the semantic layer play in Infosys's agentic AI framework?

The semantic layer—built on Genie Ontology within Unity Catalog—serves as the AI agent foundation by providing structured business meaning and context on top of underlying data assets. This video presents it as the prerequisite for consistent, trustworthy answers across both dashboard and agentic AI use cases.

Full transcript

[00:08] So we had we had an exciting morning a lot of interesting aspects that we learned with respect to how data bricks is approaching AI and how they're building on the datab bricks genie product and many more. Um this it's
[00:22] clearly you know so just a quick introduction my name is Ambi I'm an executive vice president with emphasis based in London I'm responsible for a lot of the work we do on data analytics and I focus primarily on consumer goods
[00:35] retail and logistics as an industry segment um if I just look at what's what we've been seeing over the last 12 months a few elements have come out very clearly
[00:47] this is the first time in the history of technology where technology is evolving faster than the adoption which really means that there's a deployment gap which means that there
[00:59] are a number of over the last 12 months there are a number of pilots that are getting launched all of you are probably familiar with AI pilots across the board but very few are necessarily scaling out
[01:12] and industrializing and in many circumstances there that's often because there's no clear business value attached to some of the AI projects
[01:24] and similarly data is becoming the new oil. It's very clear that both structured and unstructured data are critical for our organization and for us to be true differentiators in today's
[01:36] world and from in that context the next 12 months will be all about how do you demonstrate value and how do you demonstrate business value? Well, there are a lot of great technology
[01:48] enhancements that are happening, but how do you get that value? And to that extent, the topic today is relevant to all of us in terms of how do you operationalize AI agents
[02:00] with a true data bricks intelligence platform, any platform for that matter, but more importantly, how do you move away from dashboards to decision intelligence to data intelligence? That's really the true
[02:13] focus now. And it's important that the next 12 months are all about how do we get measurable value for our businesses in order to continue to invest on some of this great technology.
[02:26] And with that context, I'd like to invite two of my esteemed client leaders. I'd like to invite Minakshi, senior director at Manderly Data Product and Engineering. And I'd like to invite
[02:38] Vikalp, senior director digital data and AI from Adidas Group. Please give them a warm welcome on stage.
[02:56] So if you just start off and a question for both of you. Um if you could just give us a premise of your organization and specifically where does data bricks fit in to the overall construct and the criticality of a platform like data
[03:08] bricks in your organization? Vikalp. >> So for for I'm Vikalp and I think um Ambi was kind of rightly mentioning to
[03:20] the fact that you know the pace at which the technology is evolving from from my perspective at Adidas I'm really at the center center of that um evolution that's happening. Uh for us what we are
[03:32] looking at is this uh this uh evolution of AI is just not another technology uh evolution. It's more about changing the way in which we operate. The operating
[03:45] model itself is completely shifting. And why it is important because at Adidas what we have is the level of complexity that we deal with. We are dealing with uh hundreds of thousands of uh consumers
[03:57] across globe. We are looking at thousands of products per season getting launched. We are looking at multiple channels wholesale, retail, ecom, all of these getting engaged. So with that said
[04:10] with that level of complexity involved what we are looking at is a platform which is just not something which collects signals from these sources but something which in future is able to
[04:23] also act as a foundational backbone for the evolution that we are having from simple intelligence-driven decision to intelligence at scale.
[04:36] >> Yeah. Awesome. So for us um I'm from Manderly and I'm sure you guys know our brands if you don't know the name of the company Oreo. How many people love Oreo?
[04:48] Okay. And the rest of you maybe Cadbury or Milka or Tobleron. Um so again we we have a very complex we are a very large uh um you know enterprise present in 50
[05:01] plus markets in the world uh producing chocolates, biscuits, candies etc etc and then servicing all of you. Um data bricks has played a role for us and I
[05:13] really like your framing of data intelligence. Um so we've been on data bricks for around 7 to 8 years now. Um and data bricks has always played a central role in how do we get more value out of our data. Uh so when we started
[05:27] seven or eight years ago it was about data science, machine learning, uh modeling um and really getting that closer to the decision from reducing the data to decision gap. Um and in the last
[05:40] few months we have been really experimenting with Genie. Um and we have 35 pilot use cases on Genie again once again because we have a lot of data and this is the way for us us to bridge the gap between data and action and not
[05:55] insight right really the action. So again data bricks is playing a big role for us on data intelligence. So, so Minachi, if you were just to take that further, it's clear that you're piloting Genie, you've you've invested in data
[06:08] bricks and the 35 use cases that you're taking across, but is there any specific area that you feel needs to be explored or where you're spending more effort on
[06:20] because again you can apply it across a set of business functions. So, are there any specific areas which you feel are bigger bets than others? So a couple of frames to look at answer this question. One of them is uh we are
[06:33] really looking at where do we have good data number one. So if you really don't have good or reasonably good data um and which is not sitting in our data lakes then that's that first step is to really
[06:46] go and figure out what's the source system and fix the data quality. So having passed that barrier that okay I have good data we are actually uh our 35 use cases are across four
[06:59] regions and seven functions. So across marketing, supply chain, finance um and all the four regions as we look at North America, Latin America, Asia and Europe and we are seeing tremendous value at
[07:12] scale across um whether it's finance, marketing, supply chain planning or um you know even manufacturing to look at plant metrics and to deep dive and get the answers. >> Got it. So >> so very broad-based.
[07:25] >> So but but it's clear that the focus is on getting value. >> Yes. And to that extent the question that gets often asked is should you do it through a dashboard? Should you do it through a co-pilot AI agent? It's not
[07:37] one sizefits-all. So how do you how do you decide within Manderly what to go for? >> Um so recently we have passed a moratorium actually that we will not build any more dashboards like in the last few weeks we have said no more
[07:50] dashboards. Um having said that uh the way we decide is that you know if there is a execution like you just need to run and repeat a certain action every day then I would say go for basic automation
[08:02] because you also need to manage your costs and you know tokens right so go for basic automation if it's a known question and a known answer like this is how the P&L I want to see it every day in this format
[08:14] >> or I want to see market share like this then if it's a known question and a known answer go with the dashboard right and then that's That's where an agent comes in because many times you when you see a dashboard it's not that it answers
[08:27] your question. You have 10 more questions in your mind by looking at the dashboard and you want to answer all of those 10 questions in those five minutes when you started thinking about it. So for me it's like can it answer me at the
[08:40] speed at which I can think right and then if that's where the agent comes in because then I have answered all my questions like why was my forecast bad it was bad because this week to this week there was a drop where did it drop
[08:53] in this particular cracker in this particular customer so I have got my answer in 5 minutes with the agent which I would never get with a dashboard with a dashboard I'll have to go to five dashboards I'll have to extract data
[09:05] finally into Excel I'll have to get some other data, I'll have to call somebody. Right? All of these steps are replaced with the agent. >> It's interesting with the fact that you have a a complete ban on future dashboards. Uh
[09:17] >> I mean it's a temporary pause. We are it's a temporary and a smart pause. We are not saying it's a complete ban for future. It's a pause till we determine what's our strategy because the reality
[09:29] is we have tremendous amount of data. We have too many dashboards and we are still not getting the answers to the questions fast enough despite having the data and the dashboard. So there is definitely a problem that we need to fix
[09:42] which we are hoping Genie can fix. >> Absolutely. And Vik from that perspective very similar Adidas is at the pinnacle when it comes to sports apparel and footwear
[09:54] with the millions of consumers that you are engaging with often near real time. um what are the use cases you think are relevant for you know a data bricks genie concept or agentic AI and how are
[10:06] you framing those problems into solutions so uh I I'll pick up from uh her and answer the question as well so uh take a guess while we were putting
[10:18] across dashboards on the system and I think lately you know one of the the guys sitting here shared with me the statistics we had close to 30,000 micro strategy dash dashboards sitting on the system more than 30,000 plus and while
[10:33] the use cases that you mentioned that are evolving over last one year the pilots have popped up as mushrooms but as an organization beginning of 26 we took a decision that there are three key focus areas where we see the
[10:46] intelligence to the action the pace which agent brings in we're going to focus on that the first space uh for us at Adidas is uh customer support which means churning whatever information that we have captured capturing from the
[11:00] quadric forms that we are collecting or from the customer service setup that we are selecting and so on. All of that and then generating insights out of that. The second one use cases around personalization journeys. How are we cross- leveraging that and feeding that
[11:13] back into how we can leverage the uplift in a campaign and last but not the least is the pure play engineering acceleration the entire crashing of the SDLC. How do we use agents in that
[11:25] direction? These are three key strategic decisions coming from the board that we'll be focusing on. >> And in the all these these all great consumer centric examples. >> Um do you think we at a stage where you
[11:39] know often companies think this is too risky or you know these are not the most obvious of examples but clearly Adidas is you know at always been at the front when it comes to consumer action. So
[11:52] what's your view versus risk versus reward? Yeah. So some of the what we have clearly seen is the the reward the rewarding use cases right now ambi are more on the lines of these three plus on the lines of how you can leverage
[12:04] trading how you can leverage campaign assessment the business performance. So the reward of that compared to what you're investing is quite high. some of the cases which are on the lines of u uh
[12:16] the the strategic decisioning architectural decisioning or even to the matter of fact you're feeding in an auto autonomous way pricing back to your ERP systems these are the use cases which
[12:28] are still considered as comparatively risky but but for us at Adidas the decision- making is more on the lines of it's not that AI uh the co it's not the cost of AI which is blocking you it's the cost of making the wrong decision
[12:43] are you equipped to do that? So that's what is stopping us and if you find that risky let's not step into that use case. >> But that that is a segue to another important question when it's about trust
[12:55] and transparency. You know what is good enough? What's success? What's failure? That's always a big question mark because often business team members want 100% accuracy. M >> so the question remains how do you build
[13:08] accuracy and trust in trying to push some of these concepts across >> yeah u what what you also mentioned earlier mini was that you know whenever you evolve a use case the biggest
[13:20] challenge that comes out is the quality of the data or the outcome that comes through the two dimensions of that how we are approaching it at a first dimension is from the d from dashboard to the agent journey that we are having if this agent is giving me what evolving
[13:35] ing to why and then evolving to how can I address that as an action. Uh we are looking at a uh data quality behind the scene which is the strong foundational layer at Adidas which we are focusing on
[13:47] which is the semantic layer that we clearly built out. We invested really heavily on that in the last years. Uh so that's one part of it and what we have evolved is a entire metric system called data resilience and reliability index
[14:00] DRRI uh which is one of its kind that we have from an industry standpoint where we are looking at various dimensions of data and making sure that we have uh the uh the quality of the data reliable. But the second dimension of it that comes in
[14:12] is once you are giving the feedback to your user that okay why did my campaign for this market did not work the it the the agent would throw back some some response and for those of you who were
[14:24] in the uh in the uh keynote today morning you saw that number 24 coming out from Kim in his keynote he said oh that was it just hallucinated how can you trust that number so what we have
[14:36] done is at the back we have also evolved a judge layer along with what the agent gives you back. And this judge layer is a complete independent authority which based on a certain predefined rules is
[14:48] independently auditing the confidence level of what what we are demonstrating back and and and that's something these are two mechanics which we deploy to kind of make our users confident on what we are recommending is it reliable or
[15:00] not and so on. Yeah. I mean actually you any any inputs you have how are you driving trust and uh adoption in such scenarios? Right. So I think couple of things one
[15:12] is um you know like I think what um what he mentioned is
[15:26] cases we we are still using AI but we have human on top of the loop u to to judge these and we don't have a judge layer yet but that's a great concept but we have human who is on top of the loop
[15:39] who knows this stuff as the judge so it's not like we have built a direct integration between AI and pricing in SAP or promotion in TPM etc. So there is a human in the loop and then obviously
[15:52] there is a there is a constant improvement of these models and you know forecast and drift monitoring and all of that that we are doing. So that's the critical ones which where the decision
[16:04] cost of going wrong is very high. Then there is the other ones right where you are trying to generate and get insights on let's say market share and market share losses for example. Um that's where we are giving the models much more
[16:16] freedom like especially because we are testing Genie right now to really test the limits of what kind of insights can we get out of it. Um and we are we kind of are thinking about a two-pronged approach. One is let's say finance data.
[16:30] So the bar for trust is not the same. It can be very high in a finance use case uh where I also need access control more accuracy and all of that. And then there is a more of a lesser bar for trans for
[16:44] accuracy um where you are really exploring and trying to get to the insights uh because there are a lot of things that we learn from the agents that we don't know ourselves and many of them are accurate. So we don't want to
[16:56] curb the insights that we can get from the agent by trying to make it 100% accurate. So we are also trying to educate our users to be comfortable with probabilistic answers not deterministic
[17:09] 100% accurate answers and that's where we are saying that you need to know your craft really well to be able to use agents. So we need to upskill ourselves not in technology but in just knowing
[17:21] what we are doing. So if I'm using a uh agent to build my data product and semantic model then I need to know really well how to build this so that I can leverage the agent to its fullest
[17:33] right so it's not it's still not human out of the loop it is still human in the loop but with the right level of guardrails. I I love what you said. It is actually human on top rather than
[17:45] human in the loop because in the loop means you're just part of the information whereas >> I I do believe you need someone who is controlling and running it end to end and helping improve the entire activity. But then that comes to another very
[17:58] relevant topic around data foundation because yes you can build trust but how do you set the entire platform? What's the right operating model? So um what role does a govern data platform play in
[18:12] getting ourselves agent a agent enabled >> um so that and that's been the entire journey that we have been having with data bricks by the way correct so uh
[18:24] that backbone that I was talking about earlier uh the the first phase of that is making sure that all the various signals that we're looking at that gets collected the phase two of that is the
[18:37] the insight that it generates is made democ in a democratic way available and the third phase of it is making that insight at an operational scale using agents. So but in order for that to
[18:51] happen the foundation layer that we are building needs to be something which speaks the same language across the board. A simple example, we recently launched uh an agent which um which called diva and I'll be talking about
[19:04] that tomorrow in another forum with anthropic. But in this agent, we consolidated all the direct to sales data information together enabling the analysts to make sure that they are able
[19:17] to speak the same language through that same semantic layer. But then uh my counterparts in uh North America came back and say I like that semantic layer but I also want to add a local flavor to that because the inputs that are coming
[19:30] in I would want to associate that with your semantic layer the inputs on feedback I'm receiving from the local markets or on the campaigns or the local signals that I'm getting from web scraping and so on to consolidate that together and then use that. So creating
[19:44] that foundation layer in a way which becomes plug and play is something which becomes super crucial and what excites me is about the announcement that was made from Ali's side and the theme around you know the ontology layer. So
[19:56] the entire work of semantic layers that we have been doing now that's taken to the next level with the ontology layer coming in not >> um so yeah I think like it's data
[20:08] foundation is the backbone I mean if you don't have a at least a basic data foundation I don't think you can really start doing anything with AI or agents we've been building data foundation for the last eight years in mundus and I
[20:21] think that's where we are able to now bring on AI on top of that and I also Like I like to say that while it's a foundation and you bring AI models on top once they are together they are actually making each other better. So we
[20:35] are heavily seeing that AI can help us identify data quality issues which can then get fixed in the platform which then helps the models um you know to get better. Similarly, we have we are just testing building our entire semantic
[20:48] model with Genie code for example, right? And as long as I have my markdown file of what are the metrics, what are the KPIs, I can just use that and build it as a foundational layer and then build context on top of that which is
[21:01] more like on the edges or the local you know insights that the business teams are getting uh on top of that canonical sort of data model which is still a governed layer because I want to ideally have one definition of active customer
[21:15] not 100 definitions of active customer >> and and to that extent I think the question then comes how do to balance central versus market driven because it's a fine balance. You let it go out, you'll end up with a lot of different
[21:29] elements. So, how do you govern the central data governance element and yet ensure that you're driving speed and agility in your markets to basically make benefit of the platform that you're
[21:42] delivering centrally? this uh the the way the concept in which the semantic layer functions it kind of forces you to go in that direction. So the first setup is collecting all the
[21:54] data set but then the next phase of it is having that harmonized view of the business KPIs which you're having and the the canonical uh KPI the metric view registry that uh you know Mini was
[22:06] talking about that kind of is where the governance actually comes into play. So that's where you're looking at okay these are the set of KP these are the supply chain KPIs which are there uh where does the accountability of that
[22:18] sits and this is by the way the operating model shift as well. Yeah so this is where the ownership sits can we align with local markets to see do we speak the same language and then enabling them for the local tweaks
[22:30] within the local markets itself. So that's how we kind of enable governance in this setup. Yeah. And I think I I also want to challenge one uh premise that exists, right? That when anything
[22:42] is done centrally, it is slow. I mean that's the standard um uh point of view that a lot of people carry. And I would like to really challenge that. Maybe that is true before we had AI, before we
[22:55] had clouds, before we had you know ability to process and compute as much data as we can do today. But today with uh with AI and with the ability for us to compute I would say just put
[23:07] everything into a markdown file from all the markets and let's process it as fast as any market can process it. So I don't think it's no longer bound by like whether is central or being done locally. It is our ability as human
[23:21] beings to collaborate which is actually getting really tested as we try to kind of find one way of working or n ways of working you know and we like to call it as common as possible and as a
[23:33] differentiated as needed right so it's really having those honest conversations coming up with who is who is going to govern which metric because 10 markets ideally would like to have everybody would like to have their own point of
[23:46] view right and I think the human part is really The hard part it's not the technology and centralized versus non-entralized. It's about defining the roles and responsibilities clearly and having that view of what each KPI is and
[24:00] what's the formula for each KPI. >> So Maki I'll challenge you this um that's all great but if you look at some of the concepts that are being launched like you know being able to eliminate
[24:12] dashboards and run it in a natural language processing construct. We are still seeing a challenge in scaling some of these concepts across and I'm not sure if it's is it a
[24:24] technology problem, is it a data problem, is it a governance, is it a culture or trust but it hasn't scaled at the at the pace at which we should expect considering the power it has >> and the examples you are quoting are
[24:37] truly very powerful and very impactful. >> So where do you think the challenge lies? So I think um you know I like to say that um if you want to be AI first you have to be human first. You cannot be AI
[24:50] first and human last. I think most of the times we forget that we are people and we are dealing with people in our organization and they need to be part of the whole transformation. There is too much focus on technology for all right
[25:03] reasons but I think we don't talk enough about operating model. We don't talk enough about all the uncomfortable questions that we need to answer like you know what is going to be my role now that this a genie has come in right
[25:16] what's going to be my role um now that I don't need to govern this metric anymore and you have created a semantic model which refreshes every minute what what am I going to do because normally I was creating reports I was creating
[25:28] dashboards I was maybe producing rep uh word documents right so I think for me it's that we keep trying to push this a bit under the carpet Um and I think that therefore there is a lot of resistance in the system to not
[25:42] let AI go out as fast as possible because it cannot be just good for two people. It has to be good for everybody and everybody has to see a benefit in it. Um and I think that's why in my opinion we are not scaling as fast as we
[25:56] could because we are not imagining the impact of the change on people and thinking about how to kind of you know make it work for all. So for you it's more a change a massive change management
[26:08] >> change management completely different operating model accountabilities what are the new OKRs like do I now need to start
[26:22] it's a very different way of thinking which I think slowly people are doing it but it's not as fast as we would like it to be >> what's your view
[26:34] >> I think I in fact started with the same thought process where uh AI is just not another technology revolution for us at Adidas. It's the way in which we are looking at the entire operating model shift and uh this
[26:48] operating model just uh involves not only the governance part of it, not only the accountability part of it, not only the roles and responsibility, the change management part of it but it stretches
[27:00] more wider than that. We I think briefly we were discussing about the new role on 6th of June which um which SAP launched the forward for forwardlooking product
[27:13] manager who sits with the market or something like I I'm still thinking about the name of that but yeah so they launched a complete new uh new new gener of um uh how a product manager would
[27:27] look like and this product manager is not only equipped with hands on driving pilots but also sitting with the market he's just not only responsible for collecting requirements and creating the road map of the product but handson is
[27:41] able to drive that and I think that's the next level of operating models that we need to start to look at how the boundaries between the various layers involved which is both from a product standpoint from a technology standpoint
[27:54] from a market from a business standpoint how that would evolve but that said as as you know you rightly mentioned involves people technology every two weeks we have a new model uh but people are not evolving learning at the pace at
[28:07] which uh technology is changing and I think that's one of the one of the biggest biggest blockers um that that we need to kind of address >> that was probably why there's a massive deployment gap at this point of time
[28:19] >> I think that that's the point that you mentioned in the beginning as well there's a massive deployment gap because the nature of skills that we need just 3 months down the line I mean you as you opened up you said next 12 months are
[28:31] going to be how you generate value out of what what we have been equipped with today in the in the keynote. Correct. But do do we actually on ground have the skills to comprehend that and then uh
[28:44] deploy that in our operating model. So I think that would be one of the challenges that we would be having as we go through the next 12 months. So clearly the technology seems to be something we get we have a handle of but
[28:56] actually you talked about change and the ability for people to feel secure and have a path for them with respect to OKRs because you've talked about the the lack of skills and the need to have these skills within the organization but
[29:09] when you look at the operating model in itself who and when you talk about a AI agents who owns them who governs them who monitors who's accountable It's again we
[29:21] have varying sets of views. What are your views because that that probably will define the success at least in this year based on how you govern agents in entirety.
[29:33] >> I mean in in typical organ how we have been evolving. We have not been I'm not saying that we have been able to nail down this question. We also have our internal uh assessments and changes of that. But then typically how the
[29:45] organizations have been evolving is you have a team which is pure play focused on platforms and this is the team which is kind of equipping you with the best technologies that are possible and then you have use cases on top of that which
[29:58] are then correspondingly owned governed by uh the product owners. uh and I think these are two key big organizations which are going to going forward define the way in which how how the the KPIs
[30:11] with respect to say for example accuracy latencies all of these are the ones which from a technology skill set standpoint I'm I'm foreseeing that this would be owned and then when you look at that product layer and slash the market
[30:24] layer that boundary there is going to thin out because the product owner of the future is going to be somebody who is going to be technical as well needs to know that it needs to be comfortable leveraging using AI. So I think these
[30:37] are the two spaces where the overall governance for this entire agentic ecosystem uh would be sitting. Uh that's my personal opinion. >> Any inputs? >> Yeah. So it includes quite similarly
[30:49] right I think that the accountability will remain with the person who is actually making the decisions and driving the outcome. So let's say if I'm a media manager and I'm responsible for
[31:02] building the content for Oreo. Let's say tomorrow there may be a AI agent and a product that help enables me. I'm still accountable because if I don't believe in it, I'm not going to use it to make
[31:14] any decision. So I'm still accountable for the that AI agent and it could it would be owned by the product owner of that particular product whatever the that product is, right? media agent
[31:27] product um who would ensure that you know it's it's getting all the right data it has all the right tech etc and then there is a platform team as Vikul mentioned which is responsible to ensure like we have the right platform and all
[31:41] the processes in place to enable the product owner and then who is enabling finally the person who is making the decisions who who is accountable at the end of the day because if that that's the final customer for us
[31:54] >> so that's really driving we how do we drive AI now or how do we drive some of this decision intelligence >> the the last question that still remains is how do you measure value
[32:06] >> and till last year value was you know if AI has certain accuracy or if you have even managed to get it to go live it's a success >> but that's clearly changing with all the token economics and all the other parameters we're talking about so in
[32:18] your context as you're shaping use cases or how you when you're defining business problems how are you measuring value So any any business case that we start off with so we start off with what
[32:31] decision needs to be taken and then work backwards on that with respect to what is it from an first of all does it even requires an automation if yes then what would that automation be so that's how it is being tracked. There are three or
[32:44] four dimensions against which this is looked at. one is uh the the traditional which is like adoption that has been one of the key metrices which have been you know played around with all the various mushrooms of pilots that we have had but
[32:58] then it's slowly evolving towards the AI leverage that we generate and AI leverage is more on the lines of how fast are you how uh efficient are you how less costly are you are you able to do it faster are you able to do it less cost or with less people I think these
[33:12] are the key uh a few of the key drivers that from an efficiency perspective. And the last but not the least of course is the revenue uplift that to take the same decision uh or the same uh leverage that
[33:24] you get from a decision um margin uplift or it could be campaign margin uplift or anything. How much investment are you making in order to achieve the same volume of uh benefit?
[33:36] >> Yeah. >> I mean actually in your view any business KPIs that you focus on when you build these? So I think productivity is still very much around and I think sooner or later we will have to answer the question on productivity. So for
[33:50] example um and I think everyone feels that when they start using AI that and in data engineering alone I can have 3x or 4x productivity for example right and
[34:02] then even now when we are testing genie like the question is is the answer complete is the quality good what other way would you have got this answer and how many hours would you have taken to get to this answer even if you manage to
[34:14] get this answer right so I think productivity we still have to deal with it I think and we have to solve it. Second is in term second which is a very complex right if I have to start looking
[34:26] at revenue uplift or cost reduction there are it's a complex space because there are many variables at play and it's not just AI that is at play right there is the product them what we do in store all of that so we have been trying
[34:40] to do some AB testing in order to really nail what could be the revenue uplift because we are doing AI in this store and not doing AI in the other store and then trying to link it with adoption as you said and correlate the adoption
[34:54] along with the benefit in the stores where we are driving AI. So it's not a very easy answer I think. Um and I also want to quote that I was in Sapphire few
[35:06] weeks ago or maybe a month ago or something right and I was actually very very surprised that in the keynote there were people from JP Morgan Levis etc etc etc and there was all talk about 5,000
[35:18] agents being live this that and the other but I didn't really hear about value like nobody spoke about value in that entire keynote and these were the CFO talking and even then they didn't speak about value right so I think right
[35:32] Now
[35:45] the to the audience here where would you see the fastest payback for a company that's probably done very little? >> Would it be internal productivity use cases? Will it be customer propensity use cases? where where would the
[35:58] quickest value come based on the variety of examples that you have considered in your organization? >> So I think for me one is clearly in data engineering. I think there is tremendous
[36:11] potential in data engineering and and how we monitor in data ops overall uh and that really benefits the data quality right and makes you ready for AI. So if you have very little money
[36:23] spend on that I would say because that will just start the flywheel of AI with having good data at a lowest possible cost. Second I would say is really about uh at least in our industry it's about
[36:36] pricing promotions there a lot of money gets spent on these and I think some of lowhanging fruits are very much there in pricing and promotions and to some extent personalization and when I say personalization I don't mean like yours
[36:49] but at a store level at a store level what do you need to do in a Walmart store in this geography versus a Walmart store in another geography um so that that's something we are still testing we have not established lished that there
[37:02] is value but we believe that there is value um in doing personalization >> and from an ad perspective what we also believe is that there are three spaces engineering being the biggest one I'll
[37:15] split engineering into two parts as well one is crashing the entire SDLC which means right from the point where you are gathering the requirement to the point in time we are test uh we are uh
[37:27] identifying what would be the backlog to the point you're testing the code, generating the code, testing the code and then deploying it. So that entire SDLC is one of the biggest bet uh at least from our side that we are putting
[37:39] in and we're creating agents which kind of automate those workflows end to end. The second bigger bet is the entire cycle of operations. What you saw today um really excited me was the zero ops
[37:52] part of it. Yeah. And I think that's one of the biggest biggest uh place where I believe we would be able to save money. So that's from the engineering standpoint. The second one is on the personalization journey. When you look
[38:04] at an e-commerce journey, it is more about the realtime personalization and how the signals that you're getting from a click stream, how your user user is behaving on the website and how quick
[38:16] can you get a decision back to him or her saying this is the next best shoe that you could buy. So for me that is the second biggest bet that we are having and the third biggest bet that we are having is on capturing con customer signals from the customer support setup
[38:29] and tying it back to the consumer data ecosystem which feeds back into our experience how to make the experience much smarter better at the same time ensuring that it feeds back into the product creation when I say product
[38:41] means the creation of the next generation of shoes as well. So these are the couple of spaces where we believe there's one uh big bet brand organization. So we are split into sales and brand. Uh brand organization from
[38:53] our side is also putting a ton of focus on how can I uh at lower cost generate uh images for products and and campaigns so on. So now using the green screen and
[39:05] the models and so on how can that be more flexible. So this is another big bet that we having. >> Super. So again within Infosys also we run over thousand such projects and clearly engineering is one with the
[39:17] quickest adoption and now clearly front office back office is also gaining >> uh a lot of intent. The last question for the audience. Um, if there was one takeaway, one lesson in terms of
[39:30] operationalizing AI agents and moving away from dashboards over the next 12 months, what would be what would be that one single advice?
[39:42] I I would go for uh start with the use case, work backwards, identify the business value behind that and then start looking at what could help which agent or what kind of agent would start help automating that faster which would
[39:54] help start automating the workflow faster and then tied back to how you you evolve that in your strategy or tie it back into your strategy. uh what we have seen today I think has a ton of
[40:06] potential to your point uh next 12 months would be to see the the ontologies the genies of the world how would we leverage them to kind of accelerate these journeys >> and actually what would be your one quick advice to the to the group
[40:20] >> yeah so I I would say um like uh start with a narrow scope uh instead of if you have really not experimented then start with a narrow scope where you have a
[40:32] right to in really collaborate in the organization with the people who are making the decisions or closest to the outcome and cocreate along with them. Uh don't get too impatient and try to go
[40:45] too fast because then you would leave everybody behind and the pilot will never scale. Right? So sometimes less is more and slow is fast. Um in this case because if you're able to build the trust and the confidence then that
[40:58] flywheel will move faster rather than trying to run faster and get the perception that you know I'm doing a lot. Sometimes it's okay to just hold on and be a bit slow to go fast.
[41:10] >> Very good. And I didn't realize we are out of time but >> it's been a it's been a very interesting conversation. I've learned a lot through through this session and clearly the next 12 months look quite promising and
[41:22] uh hopefully in a year's time we would have made a lot more progress and I look forward to seeing what we have delivered over the next 12 months. So thank you. A round of applause to Viking.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.