• by Joe Mandese  @mp_joemandese, November 15, 2023

Less than two months after IPG Mediabrands rebooted its data and technology operations into a centralized Kinesso unit, it is rolling out one of its first new products, and not surprisingly, it is a generative AI chat assistant dubbed "My Bot."

While the rollout is far from an industry first, Kinesso Chief Innovation Officer Graham Wilkinson describes the technology more as new user interface to help MediaBrands' 18,000-person workforce become familiar and comfortable working with an AI assistant as part of a longer roadmap of next-generation technologies transforming how agencies plan, buy and service their clients media and marketing.

You can hear Wilkinson describe Kinesso's vision for My Bot's evolution over time in the brief audio conversation above, but the gist is that version one leverages Google's PaLM 2 language model, but is a proprietary IPG Mediabrands tool designed to optimize the organization's internal workplace productivity and, importantly, employee workstyles.

Importantly, IPG Mediabrands is not prescribing how and what employees use the assistant to do, but according to Wilkinson, they have provided them with recommendations about initial do's and don'ts.

Media3.0: With the initial version of My Bot, what can people do and what can't they do? Can they conversationally say, "Here's my client's brief, write me a media plan?"

Graham Wilkinson: It could do that, but actually supplied a ton of guidelines with it and that is one thing we are suggestion not to use it for.

That's because these models essentially don't do math. If you give one of these models, two really massive numbers and ask it to multiply them together, it may give you a number that looks very similar to the correct answer, but it won't actually be the answer, because it's not technically doing any math. All it's doing is predicting the characters that it thinks would make up the answer to that number.

And if it's not really doing math and you asked it to do some media planning that involved math -- whether that is predictive math, or "Here's a budget, spit it out" -- we don't want the assistant to do those things. And we already have a bunch of products that do do those things using appropriate machine learning to make sure it's accurate.

So that's something we're definitely saying in our guidelines and that we are reiterating though education that is not something we want them to use it for.

In terms of things we are advising people to use it for, it's obviously good if you have a lot of information and you need it summarized and have key highlights abstracted from it to save you time. Or something like email creation or meeting planning. I use it a lot for setting agendas and doing follow-ups to meetings.

It's also useful for things like, "Hey, can you tell me who the competitors are for this brand?" Or, "Can you tell me where people might go to buy this brand?" Do they predominantly buy in-store?

But when you get deeper into insights, we would still ask them to defer to the sources that we typically use for those types of consumer and competitive insights.

That's right now, because I still think there's an evolution to this.

Media3.0: It sounds like this initial version does a lot of valuable things -- crunching massive amounts of information to make it more manageable for people -- but that maybe you're also just indoctrinating people to get them feeling comfortable about using the interface and the technology and that more advanced applications will come.

Wilkinson: Absolutely. We have a responsibility to our people to help them build a familiarity with this technology and the behavioral change that will come with it. The behavioral change is important, because we all learned from Google and other search engines over the last couple of decades and the way you interact with those is almost completely counter to how you would interact with a large language model. So we need to get people familiar with how you get what you want out of these things.

Media3.0: Meaning that search was very deductive and generative AI is much more inductive?

Wilkinson: Yes. And in terms of how it evolves, it is very important for us to go live with it in order to get people familiar with it and to do it in a way that is very much foundational. And that's pretty much how we view the roadmap for this.