Cannes Special: Oleg Korenfeld from CMI Media Group Gives with a Case Study on Implementing AI at an Agency

Summary

Marketecture Podcast Summary: Cannes Special with Oleg Korenfeld on Implementing AI at CMI Media Group

Episode Details

• Title: Cannes Special: Oleg Korenfeld from CMI Media Group Gives with a Case Study on Implementing AI at an Agency
• Host: Ari Paparo
• Guest: Oleg Korenfeld, CTO of CMI Media Group and Compass
• Release Date: June 19, 2025

---

Introduction

In this special Cannes edition of the Marketecture Podcast, host Ari Paparo engages in an insightful conversation with Oleg Korenfeld, the Chief Technology Officer at CMI Media Group and Compass. The discussion delves into the real-world application of Artificial Intelligence (AI) within a media agency setting, emphasizing the complexities, challenges, and learnings from their AI implementation journey.

---

Background of CMI Media Group and AI Initiative

Oleg Korenfeld introduces CMI Media Group as a full-service healthcare media agency under the WPP umbrella. The agency embarked on an AI-driven mission to automate and enhance their media planning processes.

Key Points:

• Agency Focus: Full-service media planning and buying within the healthcare sector.
• AI Mission: Automate media planning to handle vast and diverse data sets efficiently.

Notable Quote:

“We were looking to find a way to automate a lot of our processes. One of the processes that we focused on was media planning... It took a lot of time.”
— Oleg Korenfeld [03:09]

---

Implementation of Generative AI in Media Planning

CMI Media Group aimed to leverage generative AI to assist hundreds of media planners by creating a prompt-based chatbot for querying data.

Challenges Encountered:

• Data Normalization: Consolidating four to five distinct data sets into a unified environment was time-consuming but essential for effective AI application.
• Generative AI Limitations: While generative AI excels in text and image creation, it struggled with consistency in number-based data responses. Repeatedly asking the same questions yielded varying answers, undermining reliability.

Notable Quotes:

“Just because you can do something doesn't mean it will always work... numbers based data sets, it was not that accurate in returning consistent results.”
— Oleg Korenfeld [05:18]

“The biggest surprise was that in generative AI, we know it's great at text, it great at pictures. But when it comes to crunching a lot of different data sets... it would come back sometimes with different answers.”
— Oleg Korenfeld [04:00]

---

Transition from Generative AI to Predictive AI

Recognizing the inconsistencies of generative AI in handling structured data, CMI Media Group pivoted to a predictive AI model that operates without user prompts.

Key Adjustments:

• Removal of Chatbox: Eliminated the prompt-based interface to prevent unreliable responses.
• Dashboard Integration: Implemented a dashboard that automatically ranks suppliers based on normalized data sets, ensuring consistent and accurate outputs.

Notable Quotes:

“...we ended up doing a predictive AI piece that automatically produces the results now in a dashboard, rather than being prompted.”
— Oleg Korenfeld [08:56]

“...based on these five or four datasets, whichever was obviously needed for that plan, here's how the suppliers were ranked...”
— Oleg Korenfeld [09:07]

---

Rollout and Adoption

The predictive AI system is live and coincides with the planning season, positioning it as a timely tool for media planners. However, adoption faces hurdles due to the necessity of client approvals, especially given the sensitivity of healthcare data.

Implementation Insights:

• Training: Over two hundred planners were trained to use the new AI-driven tools.
• Client Approvals: Legal and security considerations require thorough vetting, slowing widespread adoption across the agency.

Notable Quotes:

“We need client approvals for any kind of applications of AI on their business. There's security issues... it's very important.”
— Oleg Korenfeld [10:57]

---

Future Applications of AI in Media Agencies

Oleg outlines plans to extend AI utilization into generating insights from performance reports, aiming to streamline data analysis and presentation tasks that currently consume significant time and resources.

Focus Areas:

• Performance Reporting: Automating the generation and packaging of performance data for marketers.
• Continuous Learning: Enhancing AI models with new data and human intelligence to incorporate emerging suppliers and opportunities.

Notable Quotes:

“Insights is huge... a great place that needs to be addressed and we actually are addressing that as part of this kind of whole kind of automation initiative.”
— Oleg Korenfeld [11:54]

---

AI's Impact on Media Consolidation

The discussion touches on differing perspectives regarding AI's role in media consolidation. Oleg suggests that while AI is reliant on historical data, human intelligence remains crucial for incorporating new opportunities and suppliers into the AI’s framework.

Key Points:

• Historical Data Dependency: AI recommendations are based on existing data, potentially limiting the incorporation of new or innovative suppliers.
• Human Oversight: Agencies must integrate human insights to ensure AI evolves with the market and includes novel opportunities.

Notable Quotes:

“AI is built on historical data sets... I need to ingest this kind of human intelligence piece to it to say, strategically speaking, this is a great opportunity.”
— Oleg Korenfeld [13:01]

---

Advice for Product Engineering on AI Projects

Oleg emphasizes the importance of selecting the right AI tools for specific problems and the necessity of meticulous upfront manual work to ensure data accuracy and relevance before automation.

Recommendations:

• Selective Tool Use: Choose AI technologies that precisely address the intended problem.
• Data Validation: Conduct thorough validation of data quality and utility before deploying AI solutions.
• Manual Preparation: Invest time in designing and structuring data and processes to maximize the effectiveness of AI.

Notable Quotes:

“AI is not one monolithic thing, it's a bunch of different technologies built to address often very specific things.”
— Oleg Korenfeld [13:56]

“Upfront manual work... is so valuable to get the automation right.”
— Oleg Korenfeld [14:45]

---

Conclusion

The episode concludes with Oleg reflecting on the challenges and successes of integrating AI into CMI Media Group’s media planning processes. While generative AI presented initial hurdles, the transition to predictive AI has yielded a more reliable and effective tool for media planners. Oleg remains optimistic about future AI applications, particularly in generating actionable insights, and underscores the importance of strategic planning and data integrity in successful AI implementation.

Final Quote:

“Thank you for subscribing to marketecture.”
— Oleg Korenfeld [14:48]

---

Key Takeaways

• AI Integration Requires Careful Planning: Understanding the specific needs and outcomes is crucial when selecting and implementing AI technologies.
• Generative AI Limitations: While powerful in certain domains, it may not be suitable for tasks requiring high data accuracy and consistency.
• Predictive AI as a Reliable Alternative: Automating data processing and supplier ranking through predictive AI can enhance efficiency without compromising reliability.
• Human Intelligence Remains Essential: Combining AI with human insights ensures that new opportunities are recognized and integrated effectively.
• Future Directions: Automating insights generation holds significant potential for reducing manual workload and improving data-driven decision-making in media planning.

---

For more in-depth discussions and the latest insights in advertising and marketing, visit Marketecture's website and subscribe to their podcast on your favorite platform.

Transcript

[Ari Poparo]

This podcast is brought to you by Adelaide Media verification and measurement are undergoing major disruption. Legacy players are pivoting to performance, advertising AI is reshaping brand safety and attention is replacing viewability. Adelaide is leading the shift with au, a new way to assess media quality that scores placements based on their potential to drive attention and outcomes. Before your ads run, think of it like a credit score for media. Finally, a clear view of quality. Before you buy, take the guesswork out of your investment strategy and try Adelaide AU on your next campaign. This podcast is brought to you by audiohook, the leading independent audio dsp. Audiohook has direct publisher integrations into all major podcast and streaming radio platforms, providing 40% more inventory than what could be accessed in omnichannel DSPs. What's more, audiobook has full transcripts on more than 90% of all podcast inventory, enabling advanced contextual targeting and brand suitability. Audio Hook is so confident that in addition to CPM buys, they offer the industry's only pay for performance option, where brands can scale audio and podcasting with peace of mind, knowing they are only paying for outcomes. Visit audiohook.com to learn more. That's audiohook.com.

[Experian Ad]

You know that one friend who somehow knows everything about money? Yeah. Now imagine they live in your phone. Say hey to Experian, your big financial friend. It's the app that helps you check your FICO score, find ways to save, and basically feel like a financial genius. And guess what? It's totally free. So go on, download the Experian app. Trust me, having a BFF like this is a total game changer.

[Ari Poparo]

Welcome to the Market Podcast. This is Ari Poparo. We're broadcasting from Cannes. All week we're going to be publishing short interviews with interesting people here at Cannes and we hope you enjoy it. And we'll be back next week with a Can wrap up in our normal format. Welcome to the Market Extra podcast. I'm recording live from CAN this week and today I have Oleg Kornfeld who is the CTO of CMI Media Group and Compass. Oleg is a longtime friend of mine. He's been around the ad tech world forever in the agency side with former DoubleClick knows a lot of things. I'm really excited to have this conversation.

[Oleg Kornfeld]

Very happy to be here. Thank you Ari.

[Ari Poparo]

So Oleg and I ran into each other maybe a month ago and I was having my usual conversation that listeners of this podcast have probably heard a million times where I was asking like AI, are people really using it? What's happening? And he Had a kind of interesting tale of experience that he had in an agency. And I said, would you like to tell my audience about it? And he said he would. So we're going to hear a tale of complexity and learning and some challenges in real life use cases of AI in an agency. So Oleg, with that, give us the background about what is this agency. Maybe people don't know it very well, what you were trying to accomplish and how you went about it.

[Oleg Kornfeld]

Sure. So CMI Media Group is a healthcare media agency. We're a full service agency, planning and buying. We're part of wpp. And last year we were on this mission. We were looking to find a way to automate a lot of our processes. And one of the processes that we focused on was media planning. Media planning is this kind of holy grail of opportunities for automation, specifically on the agency side of things. And when we thought about all the data sets that we've had that our planners use or do not use when they build plans, we figured, you know what, this is a great place to figure something out. We have four or five different data sets from campaign performance to analysis to clients data that we tend to use when we plan media. First and foremost, we had to normalize all those data sets because to write any kind of algorithms to be able to apply to the answers, the data had to be normalized in one environment. So we've done that. It took a lot of time. But then the question came up as often does, where to apply generative AI? Does it make sense? What would it mean to create some kind of a prompt chatbox where in theory, hundreds of planners could start asking questions? And we've tried that because again, generative AI was on everybody's lips. Everybody was trying to figure out, and still are trying to figure out where it's applicable. And what we've learned fairly quickly is that just because you can do something doesn't mean it will always work. To me, the biggest surprise was that in generative AI, we know it's great at text, it great at pictures. Right? It can write your essays, it can create new imagery that never existed before. But when it comes to crunching a lot of different data sets, numbers based data sets, it was not that accurate in returning consistent results. So we could ask the prompt the same question multiple different ways or the same ways, it would come back sometimes with different answers. So from the standpoint of us giving this power into the hands of hundreds of planners, that became an issue.

[Ari Poparo]

So are we talking about hallucinations here? And is and When I would think some media planning questions have, you know, correct answers, like, what's the average CPM we paid on Yahoo last year that has a real answer, right? Is that what you're talking about?

[Oleg Kornfeld]

Yes and no. And what I mean by that is there's multiple questions that are being asked at the same time, right? Sometimes it could be, hey, what's the average CPM for this supplier? What are the right suppliers? What are the right data providers that should have been used? What is the client's data? And when you start to organize all those data sets and you expect certain types of answers and you don't get them, then becomes an issue.

[Ari Poparo]

So let's step back one second. What were your goals for the project? Did you think you were going to offer a tool that might optionally be used by the media planners, or are you looking to fully automate some of their work? Or somewhere in between?

[Oleg Kornfeld]

I would say somewhere in between. We were building an assistant media planning assistant, basically for a planner to have all the opportunities in place to evaluate all the data sets. That often is really hard to do in the time constraints that they have to go through all those data sets to come up with the most optimal plan and come up with certain set of recommendations. So we needed to provide an element of that to be done through automation, just to sift through all those data sets, not to leave anything to the side of it. So purely from a standpoint of automating it, we were able to achieve it. So our engine works and it can predict the right suppliers very accurately. But once we put any kind of a chat box on top of it, say, like, hey, why don't you planner, ask the right question? We cannot guarantee that the answer will always be correct, at least not yet.

[Ari Poparo]

All right, so the data was organized. The LLM or the generative AI knew what it was doing to some extent. It had somewhat predictable results. But the humans were the problem.

[Oleg Kornfeld]

That's an interesting way you phrased that question, Ari. I don't think the humans are the problem. I actually think the journey of AI is the problem here. And again, I think it's a short term situation. I really think it will be resolved over time with enough data flowing through it, with enough questions asked. But when you're asking me to put something this powerful in the hands of hundreds of planners who are responsible for so much media investment, if I cannot be 1000% sure the results will be accurate as part of the recommendation, that becomes a question.

[Ari Poparo]

All right, so tell me how the rollout worked. Did you catch this in beta and you started seeing some scary stuff like take me through the timeline.

[Oleg Kornfeld]

Yes. It was never used live. So obviously we caught it in design and development. We, like I said, first had to normalize all the data set and we started to train the models to come back with answers. And it got to a point where we had to pre write the questions and keep asking the model the same questions over and over to see if it's going to come back with the same exact results. And sometimes it would, sometimes it wouldn't without changing much. Right. So at that stage we realized, hey, it's not worth it. Sometimes you kind of look at it go like, just because we can do it, is it worth what we're trying to solve for? If the goal of this whole project was to automate part of the media planning process to help a planner sift through so much data that they often don't have time to do that may not be the right path to it. So again, going back to like, do we need to use specifically generative AI or just some kind of a predictive which. Exactly. We ended up doing a predictive AI piece that automatically produces the results now in a dashboard, rather than being prompted.

[Ari Poparo]

Okay, yeah, take me through how you re architected based on this learning. So was it just removal of prompts and just automating the kind of the data flow or was there more to it than that?

[Oleg Kornfeld]

So yeah, so instead of having the prompt screen as part of RFP process, in the stage where we're selecting suppliers, it would be, hey, click this AI button here for recommendation of suppliers based on data that you've provided. And the output would be a dashboard where we'd say based on these five or four datasets, whichever was obviously needed for that plan, here's how the suppliers were ranked and we obviously, as part of normalization process, we've ranked the entire data set and the output would be that automated.

[Ari Poparo]

Okay, so, and I assume this has worked, it's in production. Tell me what the reaction's been and how usage has been and anything else you can share.

[Oleg Kornfeld]

So it's live and we're right now in planning season. So this is perfect. This is just the right time to, hey, you're going through the next season, then plan everything. Here's another assistant for a planner to do their jobs. The rollout actually has little to do with the technology because we've gone through training, we've trained a couple of hundred planners in our company. We're fairly big agency where the holdups are right now is actually approvals because when it comes to AI, we need client approvals for any kind of applications of AI on their business. There's security issues. We're a healthcare agency. The data that we work with obviously is sensitive. So there's a lot of actual legal pieces that we need to go through before we can allow our planners to get access to those solutions. So we have the feature built, it's in our operating system, it's part of our RFP process. But some planners have it turned on, some don't, depending on where we are with legal.

[Ari Poparo]

That seems crazy because ultimately the client's going to approve the plan anyway. So why does the client care to approve what goes into the plan that they're going to approve anyway? I like, is hesitant to answer that.

[Oleg Kornfeld]

Question because I'm not a lawyer, right. My job is, my job is to build, to build the most secure possible platform. And then because the vertical we're in, because every one of our clients have their own legal teams and because they have different policies around AI, we need to go through those steps to make sure everybody understands exactly what is happening, why and how it's different from what was happening before. That's why actually it's very important. And I think actually removing the generative AI piece right now at least makes things easier because we're not creating anything new with generative AI, right? There's opportunity to create brand new data set that doesn't exist before. Now when normalizing a data set, we're just automating a manual process using AI to crunch through all the data sets. So I'm hoping actually because of that it will make things easier with the legal teams.

[Ari Poparo]

What are the next surface areas that you see as optimal for using AI to reduce costs, improve quality, et cetera?

[Oleg Kornfeld]

Insights. Insights is huge, I think a lot of time and effort and hours that are spent to generate performance reports and outcomes and make sense of them and package them up in a way that could be presented to a marketer. I think that's another great place that needs to be addressed and we actually are addressing that as part of this kind of whole kind of automation initiative. And if you ask me like Derek, which one is kind of will move the needle, I think that's the big one.

[Ari Poparo]

So I've heard as it relates to media planning and AI, I've heard two different points of view this week at can some people think it's going to cause consolidation because the AI will not be influenced by, you know, salespeople and basically media will just go to whatever works the best. And then someone else is saying, well, AI shouldn't care about, you know, spending money on more sites because it's all automated. So it'll reduce the cost of spreading your money in more places. And it will be anti consolidation on the media side. Any thoughts on that?

[Oleg Kornfeld]

I think it kind of depends on the history of the data you're working with. Because from the standpoint of an agency, our job, again, is to bring these new opportunities. We need to constantly find our audiences in different environments to deliver our messages. AI is built on historical data sets. How do we introduce new suppliers? Let's say today in Cannes, I met with some great new company that I never heard of before, and I want to apply it somewhere to my business. My AI will not be able to recommend it. So I need to ingest this kind of hi, human intelligence piece to it to say, strategically speaking, this is a great opportunity. And now this new supplier becomes part of the learning curve.

[Ari Poparo]

Okay, so for anyone listening who's more on the product engineering side, what advice do you have about when taking on a big new AI project? What have you learned?

[Oleg Kornfeld]

The analogy I've been trying to use and know how great it is, like finding the right hammer for the right nail. And because AI is not one monolithic thing, it's a bunch of different technologies built to address often very specific things. Understanding what you're trying to solve for and what your outcomes are. Actually, the manual work upfront to design exactly what you want it to do and don't just trust the system to do it is critical in programmatic. In our world for decades, we spent years trying to get accuracy in data. So right now, we shouldn't just rush into AI without validation of what is this data that we're using? Is it valuable? So upfront manual work, I would say, actually is so valuable to get the automation right.

[Ari Poparo]

All right, well, Oleg Kornfeld, thanks for sharing your story. It's really interesting.

[Oleg Kornfeld]

Thank you very much. Thank you for subscribing to marketecture.

[Ari Poparo]

New interviews are added every week at.

[Oleg Kornfeld]

Markitecture TV and your favorite podcasting app, Sam.