AI in Retail & Commerce Media: From Hype to Hands-On. IAB Europe Retail Media Roundtable, 26 May 2026, 48 min.

Guests: Marie-Clare Puffett (Senior Director, Industry Development & Marketing, IAB Europe); Cedric Epiard (VP Product – Retail Media, Mirakl Ads); James Taylor (CEO, Particular Audience).

A May 2026 conversation on how retailers, brands and commerce media businesses could put AI to work — the use cases already delivering value at the time, the barriers that stop teams scaling, and a four-step path from assistive chatbots to a multi-agent workforce.

Step 1 — Assistive: AI lives in its own window. People ask a chatbot for help, then copy and paste the answer into the systems they actually work in. Useful, but the AI has no context of the business and every hand-off is manual. Illustrative example: An AdOps lead asks a chatbot to draft a campaign brief, then retypes it into the ad platform.

Step 2 — Agentic workflows: Agents sit inside a defined workflow with the data and context they need. They are narrow — one task done well — but they no longer depend on someone pasting information in. Illustrative example: An agent reads pacing data and drafts the under-delivery notice to AdOps and the advertiser. PA notifies; the retailer decides on floors, caps and any makegood.

Step 3 — An agent for every person: Every person has their own agent. Each is still narrow, but the agent count scales to roughly the same relative size as the human workforce, so AI support is everywhere rather than in a few pilot teams. Illustrative example: Every account manager, merchandiser and AdOps analyst has an agent that knows their categories, suppliers and campaigns.

Step 4 — Multi-agent autonomous workforce: Agents hand work to other agents — planning, relevance, pricing, pacing and reporting — and operate autonomously inside guardrails the retailer sets, with people setting goals and reviewing outcomes. Illustrative example: A planning agent proposes a supplier-funded campaign, a relevance agent checks every placement against the shopper's intent, and a reporting agent closes the loop — all within retailer-set limits.

Use cases: Optimisation — Campaign and placement decisions informed by live performance rather than weekly reports. Personalisation — Ranking products and ads to each shopper's explicit and implicit signals. Product discovery — Semantic search that understands what the shopper means, not just the keyword. Workflow automation — Removing the copy-and-paste between planning, trafficking and reporting. Decision-making — Bringing shopper value, supplier funding and retailer margin into one decision.

Barriers: Fragmented data — Signals split across search, site, ads and stores. Start from the catalogue and live shopper signals you already own; connect one surface before all of them. Legacy systems — Platforms built before AI, with no way for an agent to act. Pick tools with APIs and tag-based options so agents can read and act without a re-platform. Organisational silos — Retail media, merchandising and e-commerce optimise different goals. Agree one shared measure — shopper relevance, supplier value and margin together. Stuck in pilots — Individual tests that never scale beyond one team. Use the four steps to name where you are and the next step, then scale that step across the team.

Recorded and released 26 May 2026. Update, November 2026: In October 2026 Particular Audience moved into step 4, running multi-agent autonomous workflows within retailer-set guardrails. At the time of the May recording, James described PA as being at step 3.

One working day in October 2026 · internal estimate: about 3 to 3.5 hours of one person's time plus about 9 hours of agent run time produced an estimated 115 to 205 hours of mid-level engineering work (3 to 5 working weeks). Agents: Authentication, DiscoveryOS MCP, Coding runner, Verification gate, Security remediation lead, Cloud infrastructure, Retail media QA. Largest items: DiscoveryOS MCP updates (about 40 to 70 hours, 708 tests passing across 3 runs) and an authentication security improvement (about 24 to 40 hours). Working by voice took about 3 to 3.5 hours versus an estimated 6 to 6.5 typing. A single day, measured against our CTO's own estimate of mid-level engineer hours. Not a benchmark; results will vary.

In retail media (illustrative, no figures implied): Verification gate → Relevance and brand-safety check on every placement; Retail media QA → Creative and campaign set-up QA before go-live; Security remediation → Policy and compliance review of supplier creative; Infrastructure → Feed, catalogue and stock health monitoring; Coding runner → Campaign build and trafficking; DiscoveryOS MCP → AI-assistant and Sponsored Prompts readiness; Authentication → Supplier account and access management. Agents notify AdOps and advertisers on pacing; floors, caps and makegoods remain the retailer's decisions.

Cedric Epiard, VP Product – Retail Media, Mirakl Ads: "I'm the VP Product of Mirakl Ads"

James Taylor, CEO, Particular Audience: "there's a four tier framework"

James Taylor, CEO, Particular Audience: "now level four is a multi-agent autonomous workforce"

Cedric Epiard, VP Product – Retail Media, Mirakl Ads: "on the company side at Mirakl, we are working actively on the level four"

Full transcript (spellings corrected):

[00:00:05] Welcome to the Retail Media Roundtable, the podcast from IAB Europe that dives into the dynamic world of retail media advertising. I'm Marie-Clare Puffett, the Industry Development and Insights Director at IAB Europe, and in each episode I invite two expert guests from across the industry to share their insights, exchange ideas and explore the latest trends and challenges shaping the retail media landscape in Europe.

[00:00:28] So whether you're here for practical tips, expert perspectives or just a deeper understanding of this rapidly evolving space, the Retail Media Roundtable is your go to for engaging and thought provoking conversations. So grab a seat at the table and let's dive in. Welcome everyone to this week's episode of the Retail Media Roundtable podcast from IAB Europe. My name is Marie-Clare Puffett, I am the Senior Director of Industry Development and Marketing,

[00:00:54] and I'm very happy to be joined today by James Taylor and Cedric Epiard. James, would you like to kick off introducing yourself please? And also I'd love to know what your peak of the week has been. Peak of the week? The pinnacle? The pinnacle of the week. Gosh, I'll think about that while I'm doing the intro. My name is James, I'm the founder and CEO of Particular Audience.

[00:01:23] Particular Audience was started back in 2017 to effectively democratize the technology on the Amazon.com front end. For other retailers, we specialize in fairly advanced hyper-personalization tools, recommendation systems. We have a transformer-based search engine, which is significantly more advanced than legacy keyword search engines.

[00:01:48] And we have a full retail media suite built into all of that. And we believe that similar decisioning that powers organic search and discovery should power retail media as well. Introduces better customer relevance, more automation for the advertisers and better yield for the retailers. So the tripartite win. Peak of the week?

[00:02:12] I've just moved house and England is bathed in sunshine this week. And I had a coffee with my soon-to-be three-year-old son this week. So that would be my peak of the week. Great peak of the week. You can't beat coffee with the kids. Cedric? I didn't give him any. That's also okay. Cedric, over to you for intros and your peak of the week.

[00:02:43] I will present myself before. Cedric, I'm the VP Product of Mirakl Ads. And I worked since 17 years on retail, retail media, programmatic. I previously worked for Cdiscount, which is a French big retailer in France. And my peak of the week will be less important than James Taylor.

[00:03:05] It was two days with my team on a training, on a skills communication training on the beginning of the week. And it makes me think that the skills that we need to have, it's not AI. It's to be able to communicate between ourselves. More and more with AI, we need to have these kind of skills. So it makes me reflect, but we need to stand by some time to reflect ourselves. To understand what is matter.

[00:03:35] What are our main objectives? So it makes me think, aside my agenda of meeting on products, etc. So it was really important for me this week. It's interesting that you say that. I've literally just come off of a two-day team off-site, and we also had training on communication. So maybe, like you said, there's more and more of a need for actual interpersonal and communication skills in an increasing world of AI. So that's interesting.

[00:04:05] I think a big problem is we're all sat behind keyboards a lot of the time, right? Like we're all in Slack and Teams. And it's amazing what you can just shortcut if you get on a call. 100%. Yeah. But with that in mind, we are going to focus on the topic of AI today. So I'm going to focus on the topic of AI today. Because there are obviously lots of great stuff. There's a lot of great stuff happening.

[00:04:28] And I really want to get into kind of the hype of AI, but also the hands-on practical applications of AI today. So I'm going to start with my first question, which is we hear a lot about AI transforming retail and commerce media. But when you strip away that hype, if it is a hype, what would you say right now? Let's have a reality check.

[00:04:55] What's the most practical ways that organizations could start using AI today? Cedric, I don't know if you want to kick us off, please. Yeah, it's a complex subject because we have a lot of buzz around the AI. And the AI evolves so fast that I love to use this sentence that AI never sleeps.

[00:05:21] At every morning, you will have news about something that didn't work yesterday or work now on AI. So my advice on what I'm currently doing is to, at each task that I have in my day, I'm trying to, and you need time. You need to invest time on that. If you are not investing time, it will be difficult. You need to check if AI could help you to improve the quality, the speed of your every task.

[00:05:49] And at the end, maybe you will find an autonomous workflow that can work alone. But all most of the time, it's more about finding the task that AI can help you. For example, I will give you an example, a practical example. Now I'm able to have an agent that is my chief of staff. It was a long run to do that because on the first stage, I need to collect all the information of my day.

[00:06:17] Because a chief of staff is like a shadow of you. So it was a part just to say that I need to have granular for transcripts of all of my meetings. If the AI has all my contacts, it's able to get it and I will give feedback with that after that. So it's a daily step by daily step. You find the task that it consumes a lot of time for you. You try if AI is able to help you on a day-to-day basis.

[00:06:45] And if not, you need to repeat it maybe two months after, three months after. Because as I said, AI evolves a lot. So we need to have in mind that the skills that AI has here is not the skill that it will have tomorrow. So it's my personal plan to do that. And sometimes my chief of staff does not work at all. Because Claude relaunched a new version that broke everything.

[00:07:10] And we need to work with AI and it's a bit disturbing when you have these kind of days. I love that. I love them. I'm already thinking about how to create like a PA or a chief of staff with AI. So my brain is worrying. James over to you. I can give you advice how to do that and how to automate it. But it's easy. I really know it's easy with Claude Code on Claude Pro Work. It's really easy to do that. So if you need advice, I will give you one.

[00:07:39] After that, I'm sure that you will give me back other feedback on other advices on your site. Perfect. Cool. Cedric, how do you find the accessibility of this agentic chief of staff to all of the systems that you use and databases and SaaS tools and things that you would need a real chief of staff to access? How do you handle that? Yeah, yeah.

[00:08:04] It's it was my fear because I was previously working for a company which was really close in that kind of stuff. I think we will discuss in the next couple of questions. But at Mirakl, we are very open to open the tools to be able to catch information and to work on because the main difficulty is just to have an LLM that answer any question. But I have no capabilities to do anything. So you are very happy to have a big brain, but you can't do nothing with that.

[00:08:33] So at Mirakl, our founder, Philippe, asked us to use AI at every day basis. And if we need some things, which means unlock a tool or to buy a license, there is no problem at all. So we have all the conditions that help us to try, test, learn on that subject. So it's really easy for me because I didn't have to ask something. Yeah. Yeah, I am. Sorry, go on. Sorry. There you go.

[00:09:02] I was just gonna say, does this resonate with you? And how are you like, what would be your kind of most practical tips to start using AI? Maybe with some retail media thoughts as well. Yeah, yeah. And so I think also important to caveat when we're talking about the transformation that retail media may or may not see. There is the consumer and their AI journey.

[00:09:29] There is obviously the advertiser and their application of AI, and then there is the retailer operationally. But I think that everybody in that, and also for agencies and basically anyone that's a professional to Cedric's point, you know, how are they using AI?

[00:09:47] For me, the biggest barriers are it being able to authenticate and access and execute across the platforms that I would expect a new hire to work on and work with. And I have been hoping to have a pretty wild Saturday night this weekend.

[00:10:10] And actually, I'm a bit late to the party on OpenAI because I sort of wanted to be a follower rather than a leader in this and see who else, you know, jeopardized their system security or deleted all of their data or whatever it may be that might have gone wrong. I feel like it's had some time to mature to a stage where I'm going to.

[00:10:32] I've actually got a device here, which is going to be my OpenClaw MacBook Pro, and it has already got email accounts, Slack accounts, CRM logins. And let's see. So ask me next time how it went.

[00:10:54] Because like you, I would love to have a always on virtual chief of staff, EA support rep type agent. Excellent. How that corresponds to retail media. So I think, firstly, you have currently manual workflows when you're booking advertisement.

[00:11:23] This is not limited to commerce media. And there is a vision that if you could have agents that are able to accurately interpret reason and effect change, that you will have a more efficient or less hands on and arguably more efficient from a metrics and performance standpoint. You'll make more money from your advertising campaigns.

[00:11:51] Now, there's been some sort of counter discourse around, okay, but if I have an agent and an agent and they're both optimizing on two different sides of the same metric, you know, where to where to head's butt. But I don't necessarily think that's too much different to a buyer and a seller in any marketplace ever in the history of human humanity. So I think that there are lots to do. I think there's a trust issue still. I don't fully trust AI. I've had bad tax advice from AI.

[00:12:21] I've had bad plumbing advice from AI, both of which have cost me lots of money. I wouldn't necessarily want an autonomous agent to go and cost me lots of money accidentally on the advertising side of things. So I think that the systems that enable this for advertisers need to be a ask for permission rather than forgiveness first policy.

[00:12:51] And then perhaps over time, when you start to just be approving absolutely everything in the list all the time, you might switch off approvals. And I think that the dexterity of which approvals you can switch on and off in a platform is going to matter more and more. None of that is to stop you having an agent that simply monitors and communicates with you. Right. Which I think is something that is possible today and with the right APIs.

[00:13:18] In fact, we've actually just relaunched a whole bunch of reporting APIs because we want to make sure that they are accessible via MCP or directly through agentic workflows. You know, stuff that didn't need to be a reporting API previously, but fueled something in our dashboard. I think increasingly people won't necessarily need the dashboard if they can be talking to the AI.

[00:13:44] From a retailer standpoint, same thing operationally, there's a bunch of gains to be had. Same thing with, you know, the yield management and performance management monitoring and potentially future change. But I think really where it's and, you know, dare I say this is not always agentic.

[00:14:02] In fact, it's, you know, seldom agentic, but with the predictive AI and how they can use the search and recommendation systems to automate relevant and high performance targeting and remove a lot of the manual configuration flows for their advertisers. I think that's been possible for years, but the majority of the incumbent legacy commerce media technology platforms are still doing manual keyword targeting and manual placement targeting and manual segment targeting.

[00:14:31] And that's just not the future, right? And I probably speak on behalf of Mirakl Ads with this as well. You know, we see a very much a different, a different world there. And then the consumer. Well, shall we speculate, Cedric? What do you think? Is the consumer going to relinquish all of their shopping behavior to an agent or not? It depends on the customer, the type, the kind of customer.

[00:14:58] If it's an AI enthusiast like me, yes, it could be. But at the end, what we saw with the OpenAI tentative about a checkout on this system is that they come back to reality. It's really complex to manage a retailer side. You need to manage the logistics. You need to manage your support. You need to manage a lot of things.

[00:15:21] And after that, if your retailer have all the information about you and understand also what is your last order, etc., it could also give you the right advice on the right product. So I'm sure that on the customer side, we will have more and more check out directly on the LLM for sure. But I think that it will be more app into the LLM.

[00:15:49] For example, on StackArt, you have the app on OpenAI. You will be able to ask OpenAI to have information and to make an order. But in reality, it's an app into OpenAI that is making the job. So I think we will be more in that subject tomorrow. And it will be for tools, for objects that you don't care.

[00:16:14] Not you don't care, but you don't want to think a lot if it's a good product or not. You will have all the discovery in the LLM. I'm sure of that because the discovery will be more intense. I want to make hiking with my family. I need to find the right product to do that. I didn't know nothing about hiking. So let me understand what I need to do and what I need to buy.

[00:16:41] And after that, you will spend time on a lot of different retailers. We are not thinking that LLM will take all. It's not a good work. Yes, I think hiking is the best example, right? You go into an area that you know nothing about or you want to know more about. I shopped for Niche Vermouths last night, right? With the help of AI, right? Yeah. I love Negroni and I want to... So it really helps you through that process.

[00:17:10] I spoke on the panel at the Retail Innovation Forum that Dunhumbi put on this week up in London. And one of the questions was, what's AI good for today in the shopping journey? And my answer was filtering. Yeah. Right? It's just another way to filter through lots and lots and lots of information. Right? And the same way that recommendation search systems have always been methods of information filtering. Filtering is a great use case for us as consumers.

[00:17:40] It saves us time, right? Yeah. It's a more effective way of... Yeah. It's time saving, which I don't know about you guys, but in today's, you know, when everyone's so busy and it just, it's invaluable. I think I just want to move like beyond just the shopping and the consumer point. I think it's really important point. And, you know, maybe we leave plumbing advice for plumbers and not AI just now. And tax advice for accountants. Well, and like actually any advice. Or any advice. You sort of...

[00:18:09] Everyone's got advice to give. You don't have to take it. And I think there's no difference with AI. It's fine to read it now, but, you know, I didn't buy every vermouth it suggested actually. Yeah. So, what is the major challenge on that? Is that... Sorry, sorry, Marie. No, no, carry on. That is the major challenge is that you will be able to have an expert of each subject with an AI. If you are speaking to your friend, you will have an advice, but maybe not an expert advice. Not an expert. Yeah. Yeah. It's...

[00:18:38] I think it's the CEO of Anthropix said that you have a room with all doctorants and very expert in each kind of knowledge you will have in your pocket now. So, it helps to have good advice, I think. Just other use cases. So, beyond kind of... And James, you touched on a little bit like yield management for retailers, et cetera.

[00:19:08] What are you actually seeing or where are you seeing the most meaningful AI use cases actually working today? We've talked a lot about kind of things that might work tomorrow, but today what do you think is working and working well? Maybe there's something that's working, but working not so well. Yeah. So, I think this conversation is going to be biased to on-site applications.

[00:19:36] It's where Cedric and I probably know a lot. And where it comes to in-store and Omni, maybe we know slightly less. Although I don't want to speak on behalf of Cedric. I'm sure he knows a lot more than the Omni. You're right. You're right. So, I think... First of all, what is yield?

[00:19:55] Yield is the most possible value from each individual pixel in an on-site context. Now, to a retailer, it's not just retail media. It's also the sale of the item. So, you have this... You know, let's just simplify a formula here. You know, you have something about conversion rate to impression. Right?

[00:20:25] And then you have, is there an opportunity to add advertising revenue within that without degrading the ultimate sale? Right? So, ultimately, we're looking to maximize that profit per pixel. Now, if you break it down, how does AI help us to do that? Now, it's definitely not just PCTR. It's definitely not just predicted CTR. Right? This is...

[00:20:53] It's not a complete solution. That is an ad tech company saying that they have smart, intelligent approaches to solve for ad yield, but ad yield is not the full goal of the retailer. Right? So, yield is the whole thing. Now, then there are other things. Like, we talk about pixel conversion and potentially ad profit as well. But even that's not the full picture. Right?

[00:21:22] The top line sales of an item are not the profit, the bottom line to the retailer. It gets really complex. You know, you have private label that has higher margin. Okay? You have third party brands. You might have certain trade agreements where if you hit a certain sales volume for a certain item, you're going to get a better margin on it because they're going to give you, you know, more free units on the next order or a discount off your next order or some sort of bonus. Or repayment or credit back.

[00:21:51] You know, there's this sort of nebulous fabric of incentives that retailers can get through certain items. And then on the flip side, what about like recovering lost money? Like I've bought stock and inventory and it's just not selling. So my cash is stuck in this mountain of t-shirts in this warehouse that I can't ship. So I need to reduce the price on that. Do I, or can I increase sell through on that? And even worse in grocery, you might have spoiling inventory that's about to go off.

[00:22:19] Like you need to move that stuff out the doors quickly because you don't have the luxury of time. So when we come to yield management, I'm thinking like the entire value chain there. Um, ultimately, what is it going to mean to a customer, to a retailer, to have more cash on hand and have moved stock through quicker at the highest profit. Right. And I can literally help us in every step of that journey.

[00:22:44] Um, to give you a feel for that, you know, where our recommendation systems sit, there's a cascade of algorithms and you can AB test an algorithm in the slot. That can tell us the difference between a highly merchandised filter set sort of rich algorithm versus the, the, the bare naked algorithm. What happens with the click through and the click conversion rates of those? And how does that correspond to an overall advertising engagement?

[00:23:13] If you're doing a CPC campaign, you may also be doing a CPM campaign or a CPA campaign, or even a fixed tenancy campaign. So there's just all of these like multi-dimension problems all the way down each of the goals that the retailer wants to have. Is it helpful to have AI to reason through that? Absolutely. Yes. You need the measurements of each of those tiers in the first place. I'm so agree with that.

[00:23:40] Um, because I was in the shoes of a retailer maybe one year ago, so I'm totally agree with that. It's the main problem that is that a retailer has. So I can jump on it, but on the seller on advertiser perspective, because we are, we have the same, uh, on that. So for example, a seller need to manage a multivariate problematic, which is sure. It's what it was a long question that I have in mind and no AI can handle it.

[00:24:07] Um, should I promote with a coupon? Should I, uh, below my price, decrease my price or should I, uh, make advertising? Or maybe it's a make of that. With AI, this very difficult problem could be solved. And you said, um, the retailer can have a goal to, uh, uh, to clear his inventory because he needs to have more space in the, in the, in the warehouse.

[00:24:35] So, uh, it's the same for sellers and same for other advertisers. So you, you will fix the goal, but you have, you fix your rules. And after that, you fix the AI will be able to handle the complex equation, which, which is also a real time equation. It's not possible to do it manually because at each hours your inventory stock level is different.

[00:25:01] And your, the competition is also moving at real time. That's why AI and algorithm are so important these days because the pacing will be more and more important. The decisioning will be more important. And at the end, the main, the main things that we need to have in mind is what are our main goals? Are we clear with that or not? And we need to spend time on these subjects. What is the vision?

[00:25:31] What we need to achieve on the, the, the, the AI will be, do all the, the stuff with the, the hard stuff for us. It's interesting even there, like, and then it's, we talk about pacing, right? Pacing is this like legacy thing of advertising where for some reason we want to stretch a budget over a campaign period. Um, why pace?

[00:25:55] Um, ultimately you probably should be pacing in line with your ability to ship products, right? Like pacing should almost be related to, uh, how much you can physically get out the door. And if you advertise and it leads to sales, you know, that's almost how pacing. Or it's one argument of how pacing should work. So how do you think about pacing? Because it shouldn't be this arbitrary thing. Just spend this equally over the next 30 days, right? Yeah.

[00:26:25] Um, I mean, between it, it, it will depend on your objective. For pacing, we are, we are working on a pacing feature on our product. Um, why? Um, why? Because we are seeing that sellers on advertisers have a smaller amount of bullet daily budget. And they just delivering advertising on, uh, zero o'clock to nine o'clock the morning. So you are not fitting the right hours. What are you, that CVR is, uh, is a high peak.

[00:26:54] So it's, it's not worth for them to do that. So we are imaging pacing for that. And we also saw, uh, and I also saw in my previous expense, that, uh, on the, um, each days of the month, you will not have the same CVR. On the beginning of the month, people have money, of course, but at the end of the month, it's more difficult to spend money. So you need to invest more on the beginning of the month. So I think, yes, I mean, between with this thing, we can provide the tools, but at the

[00:27:20] end it's need to be understood by the, by, by, uh, by the sellers, by the, the brands, et cetera, also by the retailers themselves. I think that's a really, really good point. Right. Which is let's say that payday happens at the end of the month. Right. So people are flush with cash at the start of the month. Right. Now, a sufficiently accurate AI solution, uh, may be able to sufficiently forecast when

[00:27:47] the best time it is to run your limited ad budgets at a time where you're likely to get the best performance from those things. Yes. Um, so like, this is a really good application of how the legacy pacing for pacing sake, I think we like need to just completely get out of our minds when it comes to retail. Cause we're so close to that point of conversion. So retail media shouldn't be pacing for pacing sake. They should be pacing because it's going to be a more efficient way to spend the money. Yeah. Yeah.

[00:28:18] And we need to, as a retailer or technology provided for the retailer, we need to open the box to let the AI work with this kind of data. We, we, we, we need to be transparent with that. So that leads me nicely onto my next question. I've really, I found, no, no, it's fine. I've found that, that conversation about kind of the AI and the algorithm and the pacing, like the, the kind of potential that I feel like we're probably just on the surface of

[00:28:48] that right now. It feels like, um, I don't feel that's kind of operating at any scale. I could be wrong, but AI has actually been pretty bad at math. Uh, generated there has been pretty bad at math so far. So I think that even then, you know, we verge back to statistical approaches and it's like, even in a forecasting, let's say that you're, you're taking a, uh, just to put it simply, right.

[00:29:13] Let's say we're taking like a long-term trend data set and we're layering over that and moving average of a more recent period. Right. And you're looking to come up with a prediction that takes into account both that short-term and that long-term prediction set. You know, we get into an area called temporal or ordinal models, which is all about prediction, right? It's, it's, it's forecasting based on long-term short-term that there's literally an algorithm called that.

[00:29:40] And, um, is that generative AI? No, it's not. Right. It's, it could even be statistical moving back into some sort of forward into machine learning. Um, I don't know if generative AI has an absolute solution on it yet, apart from the ability to be able to look in to the results of the forecast, look at the weightings and reason through those.

[00:30:05] Um, and that's where you're having like, like what is the type of AI that we're specifically talking about for each of these functions? Cause it is a multi-layer statistical through machine learning and then generate on predictive AI to generative AI. Like what are we calling what, but it's usually a combination of technology. Totally agree. And at the end you will have a mix of technology. Uh, you will have the, the LLM will be orchestrating all the, the goals.

[00:30:33] And after that you are using algorithm to achieve, really achieve at, at a cost, which is acceptable because AI is really costly technology. So that, that the kind of leading me into my next question, which is retailers, retail media networks, brands often struggling with kind of legacy systems, legacy data, um, setups, fragmented data silos organizationally and, you know, internally.

[00:31:05] They're obviously going to be blockers. I would have thought to AI adoption. How does that, how does that kind of pan out in reality? Uh, and what, what kind of, what's the biggest blocker right now you, you see, um, in adoption of, of AI, um, tools? Human. The humans, people. Yeah. People, people, people with these men, the first main blocker because, um, I saw more

[00:31:32] and more people that are enthusiasts with AI or doomers. Um, the, um, some, some people are in between, but, uh, we need to, to convince people in organization, in the retailer side, on one side is the same at, at the end, but yeah, it's not a danger. It's a, a tool that helps to accelerate, to, um, elevate yourself, to do other tasks, to, to think about, uh, other subjects.

[00:31:59] So you, you, you need to have all people thinking in the right direction. The second is security and scaling because you need to be sure that on the security, it's, uh, it's a two phases. The first is to be sure that the wrong data will not spread over the internet, of course. And the second is security because there is always, we have already a lot of securities

[00:32:26] in our companies that is used to block the access of some information. And if you want to have, as we said on previously on the discussion, if you want to unlock, uh, uh, AI, I mean, smooth potential, you need to give it access to the, to the information. I think the, the two main topic that we need to handle first, uh, is security on human, uh, behavior.

[00:32:55] James, would you agree? Is there anything else? Uh, definitely agree. Um, I think that it could be maybe easiest to think about it within a framework. Um, and I, it's sort of how I, I, my mind tends to try to work.

[00:33:17] Um, so, um, has anyone seen this news about X, uh, X AI, SpaceX, uh, basically buying an option to buy Cursor, the, the development agent? Right. Like, I think it's like 10 billion now with an option to acquire the 60 billion within a year. Right.

[00:33:46] Now, um, Cursor is a, it's an IDE. It's what engineers can use to put code together and develop systems. Um, and it basically gives you the option of which, uh, AI model you want to support, um, that, that coding and development work. Um, Cursor is rolled out to a hundred percent of my engineering team.

[00:34:11] Um, I think that, uh, AI adoption has been the most, uh, it's been most forthcoming really with the, the engineers and it, it's a really good application of, of AI because it, um, you know, code is language, right? Code is language and code is typically available on computers. Uh, and therefore it's very accessible and readable, um, by AI.

[00:34:41] Um, so AI has become a phenomenal coder. Um, you know, I, maybe 70% of the code we write today is, is not human written. Um, it still is human reviewer and, and remember that because that puts us at level three, potentially of, of the framework I'm about to share. Um, so, uh, there's a four tier framework.

[00:35:12] Tier one is where you might speak to a chat bot, right? You know, it's a separate system. You, uh, you may have, uh, we'll call it assistive, right? You, you may have a conversation. You may, you may not take that. You may copy and paste things across. Um, level two. So, um, this is introducing agentic workflows.

[00:35:39] Um, so this might be very, very, very narrow, right? Like you might literally have some automation, um, around how you're using it. Um, ideally the agent has some context that it's not just like a sort of blind Zapier based automation. It's, it's something that is at least interpreting an inquiry. Um, but it's, it's very narrow. Um, level three is where, you know, every person has an agent with them, right? So all my developers have Cursor.

[00:36:08] They have, they have, uh, they're, they're using AI. Um, but, but again, it's still relatively narrow, but you've scaled it to a point that is roughly the same, uh, scale as your human workforce. Um, now level four is a multi-agent autonomous workforce. Um, so is Particular Audience doing this yet? No, we are not right.

[00:36:38] Um, do we want to possibly we want to do this? Um, how would we tackle it? Um, the, the best analogy to this is if you hired a graduate or someone straight out of school. So young, not, not especially professionally experienced. Um, but they do have access to all the knowledge in the world. Right. What would you let them do?

[00:37:04] Um, and I think that we have to think on this, like multi-agent autonomous workforce. In the sense of if we hired one, 10, a hundred graduates tomorrow, what would we let them do? They probably start quite narrow individually. They probably have their own roles. They probably have limited access and permissions, and you'd have to orchestrate that workforce in the way that you would do with hiring a whole bunch of humans. Right.

[00:37:33] There is no, I've just turned on AI and it runs my business now. Um, unfortunately. Uh, so I think that this, and by the way, level four has only been possible in the last three months. Yeah. Right. It's that new. Um, so it's something that we are knocking off things bit by bit. Well, can we do this? How can we do this? You know, I think it has to be a very sequenced, uh, carefully done thing.

[00:38:02] And absolutely back to Cedric's point. Therefore the human is still the bottleneck. Um, yeah, I'll pause there. Good to review that. Yeah. That framework is a very kind of practical, real way of thinking about AI today. Um, I've, I've just frantically made notes, James. Unfortunately, everyone should be in level three.

[00:38:28] And I think that if a listener to this podcast is, is in a business and they're not at level three, um, there's not really too much of an excuse as to why not. Um, apart from big sluggish old businesses, not having the transformation agility that they should have perhaps lots of rules, but I would warrant that even in those, uh, large companies,

[00:38:58] individuals are still using AI on their personal or the personal level. So I think it's, yeah, it's fascinating. And, and I'm thinking about, you know, kind of my own role and my own team and how we, how we make our way through the framework and where we sit. So it's a very practical way of thinking about it. And hopefully everyone will listen to this and take that away and take it back, uh, internally and, um, and do a kind of internal assessment, I guess, on the framework.

[00:39:24] Um, we are, we are kind of coming, coming to the end, but just, uh, we always like to finish with, um, very quick fire round. Cedric, you kind of touched on this to be fair at the beginning, but the first question is how are you currently using AI? So how do you use AI today? And maybe, you know, you could tell us where you are in this framework. Um, and what would you like AI to do in the future? Is it to get to level four here or is it to do, to do more? Um, Cedric, do you want to, do you want to kick us off?

[00:39:54] Yeah. Thank you. Um, as I said, um, I want to have my chief of staff or professional usage, but also for my personal usage and, um, James, I have also OpenClaw installed in my computer. So yes, uh, I want to practice to have the level four, uh, on my personal side, on the company side at Mirakl, we are working actively on the level four. Um, it's difficult because we need to link the, the, the dot, uh, because at each point,

[00:40:24] some, someone is doing automatically something on those are some things, but you, you need to link, you need to evaluate if it's okay or not on this. Is it scalable or not? So, um, my ultimate goal is with AI, because I'm a AI enthusiast, as you, you see on my speech is that I will not, I don't want to spend time to, um, duty tests, but I want to spend time about my main objective, my main goals.

[00:40:50] And if, uh, the AI is able to, uh, help me to not spend time to, uh, writing an email on, uh, to, uh, to, uh, to resume a discussion, I will win time to spend on the, the, the key things on, on our life at the personal stage and also on the professional stage. Yeah. It's about to that time efficiency again, isn't it? Which is always.

[00:41:15] Yes. Time is accelerating with AI. So we need to be more and more on the, the, what we want, what we need. Yeah. Yeah. Yeah. I think, um, yeah, I think just to say like almost exactly the same thing, but I guess to position it in that,

[00:41:37] the opposite way of like, well, what am I not doing with AI? Um, which, which maybe is increasingly going to be the, the question, um, you know, uh, to sales, right? Sales is, is human. Um, people buy from people.

[00:41:59] I used to have very large team of business development representatives that would do prospecting and cold calling, and they would book, uh, six meetings a month per BD rep. And we would convert 12 and a half percent of those meetings within six months for an average contract value of, you know, five, six figure, uh, contracts, right? It was a really nice predictive model.

[00:42:27] Um, there was a huge proliferation of BD reps at tech companies, all the VC funding that happened through COVID. And that became a very congested channel. Um, much harder to get cut through. Everyone was getting cold called from a hundred BDs a week. Right now AI has almost made that problem, uh, worse because there have been companies that have spun up sales agents.

[00:42:55] And, you know, I, I get the worst, uh, outbound. It's like, Oh, I see you follow X venture capital company. I do too. Should we talk about business process? What are you, what are you doing? It's an, and it's so bad. And I think that that's only going to push, uh, go to market. And this doesn't just apply to me as a tech vendor.

[00:43:21] I think it applies to retail media networks. They'd sell to their suppliers as well. You know, I really would shy away from any actual client B2B client touch points with AI. You know, I think that is more and more human. We're investing more in events and being physically present and participating in, uh, you know, industry bodies like IAB than we ever have done before because the human relationship is just so, so, so important for that.

[00:43:49] Um, now I, I mean it when I say I am not looking at firing any of my workforce because of AI either. Right. I truly believe they are positively, uh, augmented. They are, they are, it's a compound win, right?

[00:44:09] They are that much more valuable in their roles with AI. Um, and I'm not alone in this. Um, you know, I think some of the big tech companies that have let people go have sort of almost used it as a market signaling and we're so forward and actually they've overhired in the past and unfortunately they needed to reduce their headcount.

[00:44:28] Um, I really like, uh, lovable the company out of Sweden. They were asked that valued more than Klarna now. Um, AI, uh, basically website developer that you can work with. It's really good. You should try it out. Um, they were asked in the press after raising all of this money. Um, how come you, if you're an AI company, why is your headcount growing?

[00:44:50] And the founder's response was, was very simple. It was, well, the more people I have using AI, the more productive we're going to be as a business. Right. And that's, that's perfectly correlated. Um, so I hope that that compound, um, multiplier starts to increase more and more with AI in a way that's very safe. Right.

[00:45:16] Much like Mirakl Particular Audiences powering many, many, many, many large retail websites. We cannot afford for something to break in production. Right. So there are very real constraints on where we can apply probabilistic technology like AI as well. So I understand there are constraints. I love the human to human thing. And I don't think that's going to go away. Um, and I think that we can use AI to use AI.

[00:45:45] To make us all, uh, probably more time rich. Take Cedric's, um, conclusion. I think that is, you've both summed it up extremely well. So I've taken, you know, time is accelerating with AI. You're right. It is. I kind of took away a formula, James, more people multiplied by more AI equals more productivity. And I think that, that is something we should all, all, um, bear in mind.

[00:46:12] And I think the other thing, you know, listeners should take away and think about is what am I, or what is my business or what is my team not currently doing with AI? Um, bearing in mind the need for human interaction, the, in actual fact, the increased need for human interaction and communication. You know, Cedric and I have both been on training courses, workshops about communication this week.

[00:46:38] How can we use AI to give us more time to make those human connections and that communication process better in actual fact? So I think there's, there's so much opportunity, um, both with using AI and, and the time and the efficiencies that gives us to do other important aspects of, of our roles.

[00:46:56] Um, we've talked a bit about, you know, the use cases within retail and commerce media, yield management, using AI to, you know, uh, better predictability in terms of shift in stock, um, and all of those things. So I think we've had a bit of a reality, reality check as well as to where AI is currently kind of, um, helping us, um, some of the blockers. It's been a really insightful discussion, particularly for me. I've learned an awful lot. You're both experts in this.

[00:47:25] So thank you so much for taking the time to be on our podcast and speak to me and, and our listeners and, and, um, yeah, hopefully see you both again soon. So thank you very much. Thanks for having us. Thank you. Thank you for tuning into the retail media round table. We'll be back soon with another episode, so make sure to subscribe to stay updated.

[00:47:49] In the meantime, you can visit iabeurope.eu to explore the latest retail media resources and insights to help you navigate this dynamic space

Listen on Spotify · IAB Europe episode page · AI in Retail Media Town Hall