Subscribe
Share
Share
Embed
Ever wonder what lies beyond your everyday ChatGPT? Join the VML Canada Innovation Team as they dive deep into the world of AI Agents – specialized, self-enclosed chatbots designed to be experts in one thing. This episode unpacks what agents are, how they work, and crucial best practices for leveraging their power... and avoiding their potential pitfalls.
In this episode, you'll discover:
- What Exactly is an AI Agent? Learn how these focused bots differ from general LLMs, offering unparalleled accuracy, personalization, and the ability to challenge your thinking – unlike "puppy dog" chatbots prone to hallucination.
- The Art of Training: Rick Brown and Jeremy Lenz reveal the critical importance of quality inputs ("garbage in, garbage out") and clever strategies for structuring data to make your agent an expert.
- The Human Element: Trevor Byrne highlights the essential role of feedback loops, while Jeremy issues a powerful warning: "We are still the experts." Discover why the carpenter never blames the tools when it comes to AI.
- Real-World Innovation: Explore how VML Canada is already using "briefing agents" to streamline projects and the exciting potential of building "teams of agents" for complex tasks.
- How NOT to Use AI: A crucial segment featuring cautionary tales from Bell's Fan Expo blunder and Home Depot's facial recognition controversy. Learn why empathy, transparency, and avoiding generative AI in public-facing social content are paramount.
Connect with VML Canada Innovation: Have questions or ideas? Email us at innovationcanada@vml.com
Creators and Guests
What is Why Didn't I Think of That??
Curiosity piqued? "Why didn't I think of That?" is your weekly dose of Canadian ingenuity, brought to you by the Canadian Innovation Team. Join our team for a relaxed, round-table chat where we break down the latest in innovation – from the complex world of tech to surprisingly simple, yet brilliant, creative initiatives. We'll cover news, offer deep dives, and share fascinating interviews, all through a distinctly Canadian lens. Whether you're a fellow innovator or just keen to learn about the exciting developments shaping our nation, this podcast is your go-to source for understanding what's new and next in Canada.
At VML Canada Human First Innovation is at the heart of everything we do. We are doers, innovators, and big thinkers dedicated to leading and inspiring. What's next.
“If anybody came up to me and said that's what the agent told me to do. I think that would be a big red flag.” – Jeremy Lenz
Welcome to, why didn't I think of that? The podcast that explores the most inspiring and innovative advertising campaigns from around the world. We are your VML Canada Innovation team. I'm Rick Brown and I'm a creative director based in Toronto. I'm Trevor Byrne. I'm a managing director here at VML and to my left in person is, yeah, I'm Jeremy Lenz. I'm a creative director as well at VML. And on the on the line we have Tim Voet, technology guy from VML sitting in Montreal remotely.
Awesome. Welcome guys. On today's show we're talking about AI agents. You've probably been hearing a lot about them lately. We break it down for you. What are they, what can you use them for? How do they work? And later in the show, we've got some great examples of how not to use technology.
Jeremy, why don't you start off what exactly is an AI agent?
Yeah, absolutely. An agent is more of a kind of a self enclosed chat bot, so I think we're all familiar with using chat bots like chatGPT an agent is a smaller version of that where we. Basically program it to be an expert in something. The great thing about a chat bot of course, is that, it has endless amounts of knowledge from everywhere, but that also means that it's gonna hallucinate, it's going to go off topic, it's gonna bring you back information that's probably, not quite as accurate as you would like it to be, but being able to program it and streamline it into just being really good at one thing.
It allows it to be very accurate in the way it brings information back to you. You can obviously make it very bespoke. It can be a very personalized experience with it. It learns over time as well. So the more and more that you use an agent like this, the more and more it learns from you. And you can even store extra memories in it as well.
So if as you're going through it, if it's coming back and it's, if there's things that you want to add to it, you can just do that through speaking with the chat bot itself. So it's a really powerful tool. Multiple ways of using it. Tons of different use cases for it. It's also the word of the day.
I don't think we have a gong or anything like Peewee Herman, but we hear it a lot and I don't think a lot of people actually know what they're great for yet.
Yeah. Jared, you touched on it a little bit, but if there's anybody that's listening that isn't completely in the know, in terms of what these agents are, what they do.
What would you say is the biggest difference between our standard interface on an LLM that is not an agent and then after having built and trained an agent, what the biggest difference is there between those two? Because I think a lot of people are still not necessarily using those agents, but they are using the traditional or the easiest way in on some of these platforms.
I guess the biggest difference I've noticed is first of all, you know how smart it actually is compared to a regular chat bot, I think. Once you make it an expert in something and it's not a generalist like regular chatbots are, suddenly you are, you're competing with a colleague almost.
They wanna be your friend. They're gonna tell you what you want to hear. That's gonna lead to things like hallucinations. It's gonna lead to things like giving you the wrong answers, just to please you. It's like a puppy dog. It's gonna do bad things, but it's gonna be really, it's trying to be your friend.
Exactly. Exactly. Because it's really cute the whole way around you, so you just want it to be right. Whereas I find an agent can really challenge you. It can push back on you, it can, be confident about its information and really challenge you to get better and to dig deeper into into any data that you're going into.
What are some of the parameters that you can control to help that agent behave that way?
The one, the, obviously you can do that through, a lot of the tone of your sources. So the tone of the sources that you input the way that you write the chat bot, it's gonna pick up on those things.
You can specifically ask it to take on a tone. So if you have examples of the tone that you want it to express itself in, you can do that. But the most powerful thing, and maybe, I don't know if all agents, agent building systems have this, but if they have an ability to change its personality and it seems like the default personality is always factual, analytical, deep thinking. That's what as the default options in there. But when you start playing around with them, I like to think of how can I get somebody to really push my limits? And it's almost like makes make a chat bot somebody you don't really want to have a long conversation with because that type of chat bot is really gonna push you into.
Into bigger domains. And what I mean by that is by, make it confrontational, make it very critical of things, and you'll love it for it because it will, it'll make you better at what you're doing.
Yeah, maybe it's worth spending a second talking about the inputs or the training methods in order to get to that audience and the importance of the quality of those inputs.
Rick, maybe expand a bit on that, just so people understand the importance of that input. It's like garbage in, garbage out for sure. But take us through that a little bit.
Yeah, that's exactly what we've run into. So whenever we hear somebody's building an agent for something important, we immediately wanna see it.
We want to take it apart and see how they organize the sources of information. Because when you upload files to an agent, that's its world. That's its knowledge base, and it's critical to make sure that it understands the information and can use it to give you useful responses. I remember the first time I built an agent, I dropped in brand guidelines as a PDF for a client and expected that the agent was now an expert on that brand.
And it, it really wasn't. It's important to simplify the documents, those source documents that you're using, and make sure that. Everything is explained in a way that the agent will understand. Jer goes even a step further and creates a spreadsheet with all the links to the various sources. Can you tell us a bit more about that, Jer?
Yeah. I always like to give I like to think of an agent as a lot like my own brain sometimes where I need to be able to scan something before I take a deep dive. So I need to have kind of that high level at my fingertips type of knowledge to keep quick. It certainly doesn't wanna be slow either.
Which is actually one of the reasons why it screws up a lot is that it sometimes favors speed over accuracy. That's another conversation. So you create these quick Cole's notes, version of all of your sources, little brief descriptions of what's in each source. I even like to reference file names with the agent so that the agent can.
Pick out a file name when it needs to reference a certain source. And then in that supporting document, that's where you can have the 500 pages of Word doc to really have the knowledge that you want to make it an expert.
And after you've done all that, it's super important to make sure that you test the agent to make sure that it's gonna give you the kinds of results that you're looking for.
So if you have a complicated chart in there, you wanna make sure that it will explain it to you and get it right. We do that. For all of the documents that we upload to make sure that it understands each of the. Source files. And then as Jer mentioned, if you have that spreadsheet and you point it to that as a starting point, it knows where to look for things because we found that in some of the agents, if you had a really detailed document that it was like stopping after page 10, so you weren't getting really proper results.
But this way works around that.
And on that, I think it's also important, and it's something I've experimented a bit with as well, is. Making sure that feedback negative and positive feedback is a loop that is on. I mean that for us and the tools that we use. Yeah. With VML Open, it is something you can toggle on and off, and I have found it to be incredibly helpful, if not essential.
Yeah. In building an agent because. As soon as I had awful results with something, and it is, it happens, it learned really quickly and I found that the next query was bang on. It was totally right and I, I literally had to say to it, I don't ever want that type of response again. When you pull that information, I want you to ensure that it has come from a source like and X, Y, and Z.
They're actually like fairly deep linked destinations on a website. Where there was very critical and timely information and it only searched for and sourced that from that point forward. So it was very helpful. And I'm so curious to know too, maybe more an ethereal question in it is how long will those memories last?
Does it, do the LLMs memories degrade over time? Do they forget?
You can find them so they are actually editable. So memories, and I believe this is in all agent building software, memories exist as a. Basically a text document. I've gone in myself and created a memory just manually.
Okay. So I have seen that as part of the interface.
I didn't realize that was the exact reflection of what we're just talking about,
huh? Just done differently. So working with the chat bot, you have to create the prompt as if you were creating an agent. It just all has to be part of the prompt, so you can go in and tell it these things and say, from now on, consider this as a fact.
And it'll be as localized to that chat history of Gemini, Chet clothes, sonnet, whatever. It won't persist across chats because it's a new instance, whereas our agent is the one instance we're working with. But you can, and I've got a few going on Gemini and Chachi pt, where it's like in this context, and this is where Jerry, I love your example.
If you wanted to be the most critical person against you. I'll run ideas by both instances, the real versus the angry or critical one. And those instances continue to persist as long as you're within the same chat.
But then as soon as the chat's closed I guess as soon as you move on to a new chat, then it's basically you're clean slate, right?
Yep. I know, I remember before making the agents, one thing I would do is always ask the chat to output everything that I would need to input into a new chat to make it exactly like the chat that I was in. So I actually, I think I still have these documents around of chats that gave me basically the instructions to create another version of itself.
That brings up a really important point. Whenever you create an agent and it's working and you're happy with it, it's really important to back that up. So we at the moment will copy and paste every single setting that goes into it. Into a, just a ba basic, plain text document and then have a folder with all the source files and that way you can zip it up and send it to somebody if you wanna share it.
Or you can recreate it again in another project very easily.
It's like keeping your jewelry in a safe.
Do you think that like the issue of IP, I think it's a I think it's a really big topic. I still, I don't have an account anymore with Midjourney and I did have one for a long time and I do remember, 'cause it has a public.
Facing gallery of things that people have created. And sometimes you see things in there where you're like, oh, I think that's for something that I should probably not see. You know what I mean? Because I can get where they're going with this and understand who it might be for. So I think that's, obviously I think most agencies and most maybe, larger.
Public agencies are probably looking for ways to secure their IP under something walled like this. It seems so long ago. It was wild days of us playing around with Midjourney.
Yeah. It was like 14 months ago. Yeah, exactly. I actually remember being in one of those chats and watching somebody work their way through a prompt getting results for it was a really bad meme.
That I actually ended up seeing the next day in the news of this AI thing is showing up. And I remember seeing it in Midjourney, the guy working through the prompt to actually make it. And it was a weird kind of watching the sausage get made situation.
Yeah, maybe this is a little bit of a midjourney reality show, somewhere in there.
Tim, is there anything that you wanted to talk about in terms of agents, anything we haven't talked about yet?
No, I think what you were saying before is it's only as good as you're willing to put out there. And if you're looking for something to placate you you're not gonna get the quality you're looking for.
You have to be going in and feeding it the information you want to get out is otherwise you're just gonna get, yeah, that's a great idea. You should go with that.
It's just I think it's a really important watch out to, to understand that, it's not a person, it's not, it's an entity that we've created.
So I think that a very clear watch out, I think right from the beginning is that we are still the experts. Even though we're building these machines that can help predict and help bring new insight to the table, we have to be the ultimate deciders. We're the ones responsible for our own decision making.
We can't move that liability onto the agent itself. I think that's going to get us into a lot of trouble later on. If anybody came up to me and said that's what the agent told me to do. I think that would be a big red flag for me on whether or not that person is capable in this position because I think that we all need to have our own agency when it comes to our decision making.
You had a great reference point when we've talked about this in the past, that the carpenter never blames the tools. We are the humans still controlling and I think that it's very important that we all look at it that way.
Yeah, definitely. You just reminded me when you mentioned tools.
One of the most useful tools that we've been using these agents for lately has been a briefing agent. We do it on every project.
I love the briefing agents that we make 'cause I think that even just the fundamentals of what are the deliverables for this project, I can get in an instant, which saves me from having to text somebody at, two in the morning.
I think it's one of the best Cases that I have heard just from a convenience standpoint in a while, and I would imagine, I don't know if you guys have done it already, but building an agent that you can swap in and out between briefs no matter what the brand is. They can still perform all the same tasks or would you use the same agent and just ask it to repeat the task with a new set of information in and hope that it compartmentalizes it?
I've been creating bespoke agents for each Right project. Hopefully it's a different brief every single time. We're not just copying and pasting briefs. Yeah. So I would expect the agent to also act a bit different for each brief.
It's just the sources would be different and the brief would be different.
But you would, you could absolutely create your agents. To Jeremy's long explanation before is I want you to behave like this. I want you to challenge IDs, I want you to do this. Is that then becomes the, Hey, this is my new agent with this new set of sources and this is the brief. Analyze it and review it.
I think that the way that I would actually build it would be more or less, I would have different agents that do different things. So I'd have persona agents, and we're all using the same personas over and over again. Strategy type agents as well, like shopper agents, consumer agents, things like that.
But then I would actually consider building the agent that speaks to all those. I would actually consider making that unique to each project because that's what's gonna take on the tone. That's what's going to use those sources in different ways based on the project. So I would almost consider creating like a suite of.
Of supplemental agents that we could then plug into one bespoke briefing agent each time.
And if I'm not mistaken, there is a. There is a way of organizing that as a team of agents. So you've got the ring leader who would be guiding all of that, and then your sub-agents for each of, potentially for each of the different brands or maybe coming with a little bit of a different perspective that you've built for each of them based on a client or a industry.
I think the possibilities for agents is endless, honestly. We've been doing a number of training sessions and the first slide that we have as people are coming into the room, it says, I want an agent to do. Blank. And at the beginning, maybe people don't necessarily know what an agent is or what it's capable of doing.
By the end of the training, they're coming up to us and they've got all these ideas and I absolutely love that. So hopefully this discussion has inspired anybody listening to the show and given them some ideas for some agents that might help them.
Okay, what's next?
Why don't we jump into some of the hot topics and news?
We've got two stories basically masterclass and how not to use technology. And the first one is actually from a technology related company. Bell got blasted for using AI at a photo booth at the fan expo in Toronto. Jared, what happened there?
Yeah, it, I think this is a really good example of how we have to be really cautious about when we're using ai, especially as marketers.
I think, it is a flashy new thing and we love to get it out there. We love to see what it can do with with a base of people. But I think Bell had a pretty big misstep when it comes to ai. They showed up at Fan Expo. They had a booth the booth. Used ai, like it took a pic, it took your picture, and it used AI to change you into, a superhero and anime character, different things.
And it actually sparked a bit of a protest, a backlash from, a lot of the people that were at fan Expo. I think what they failed to realize was that fan expo is actually a really big collection of artists. Whether they're cosplay artists or they are illustrators or comic artists. They're writing comic books.
They're making comic books, and there's a lot of artists that show up there and their sole kind of purpose to be there is to actually do exactly what that Bell Booth is doing. Yeah.
One row over what you had artists sketching out people as superheroes, as anime. Characters and doing that exactly thing as part of their livelihood.
Yeah, it's the whole notion of authenticity as the final frontier for us. And unfortunately, when a brand misaligns like that is what hits the headlines.
Yeah, I think it's a really good example of how we need to be very careful about when we're using ai when it's outward facing. I know that we.
We've been given license to use AI in public facing spaces, and that's very enticing for a lot of people that they could, we can create what looks like a finished product and send it out. The reality is that if we don't, if we don't think about it and be very empathetic about how we're using it, we are gonna turn people off and we are going to become more pariahs than we are going to be champions of really great art.
One big watch out that I tend to give to clients who think that AI is gonna help them is don't use it in social. The second you use AI in social, it gets recognized as ai and you will be called out immediately.
And you're saying specifically from a generative, yeah. Image and graphic.
Generatively driven outputs and use in social,
yeah. And video now too.
AI still has a look and feel as well. Yes, we like to think that it's exactly like a real photograph, but it's not. And we can detect these AI images whether we know we're detecting them or whether it's more of an instinctual thing, but they don't look real.
And we have to realize that even though we can convince ourselves that they look real, most people are gonna know that it's not. But think about how far it's come in a very short period of time. Sure. I think about Google's VO two versus VO three. Right night and day. Yeah. And then even some of the new image generation models are getting better and better.
So I agree with you. And still none of them can seem to figure out hands for some reason or faces. Sorry, Tim, you were gonna say something? Sorry, Tim, I cut you off.
No, I was just laughing because hands are the most common thing that they mess up. Like they, extra fingers, not enough fingers. Inverted hands, that one comes up quite a bit where fingers are bending the wrong way off of a palm.
I dunno if you've seen that one. I've seen that one a lot in AI generated images.
Yeah. I have, so my partner my, my wife Courtney is a trained illustrator. That's what she's made her entire career out of. And she has said over and over again that the thing that she challenged herself with in terms of evolving as an artist and getting better.
She said at any times she wanted to challenge herself. She would draw a hand. And so it's interesting that the generative models have such a difficult Yeah. Difficulty with it as well. They should have started with that. Yeah. Yeah.
Cool.
Have an illustrator train it. No, don't.
Alright, let's move on to our next story.
And this one is not AI related, but it is technology related. Tim, what happened?
So Home Depot is being sued for. Not disclosing that some of these self-checkout caches were also doing facial recognition. Now what were they doing with it? We have no idea. Now, conspiracy theory in me can go nuts and say, Hey, they're using it to and benefit, like prevent fraud that you're using the same card with the same face and all that stuff.
But to the extent is we don't know what it was being used for. It could be used to map. Your purchase history to your card, to your email, to your face, and now target you wherever they wanted. Because once you exist on camera tied to something, you are available. But they're being sued because there was one, there's no camera disclosure saying, Hey, you are on camera.
Much like the Walmart stuff, checkout counters where you know you're on camera. 'cause like it's up and in your face. But the Home Depot ones like you couldn't see it. It's just part of the screen. It's one of those little dot cameras. On the checkout screen and it was scanning your face, taking pictures and or video and somebody found out,
I'm sure they have a sign somewhere that's in six point type in the back of customer service area.
As you walk in, you may be on camera. That's right. Which people assume as security cameras not tied to a purchase. 'cause anybody who's purchased a Home Depot knows that, Hey, do you want your receipt by email into your email? And the second time you come back with that same card, it still tracks your email association to that card.
So like you're getting purchase history, you're getting. Patterns, visits and add face or facial recognition. You have a marketer's dream there, buy this data set and you'll know everything about people.
If you'd like to talk to the hosts of this program or any of the leadership at VML Canada, email us at Innovationcanada@vml.com.
Check out our show notes for even more details and links to the cases featured in this show. Please remember to follow and rate us on Apple Podcasts. Spotify Audible, or wherever you get your podcasts.
The views and opinions expressed on the show are solely those of the participants and do not reflect those of VML, WPP or their respective affiliates or employees.