ViantAI: Adtech's crack at a fashion catwalk

When Viant announced their AI DSP solution last week, they would have welcomed controversy.

Not in the same way a sordid C-lister does when indulging in an unsavory online post, but to an extent that their new product would be deemed relevant enough to the AI conversation. The biggest risk for any technology platform currently, when announcing new AI solutions, is to be met with an air of indifference or a bored response along the lines of “yeah, but everyone is doing it now.” The immediate feedback received on the social platforms where they posted the launch videos proved, if nothing else, that Viant had something relevant to say.

Positivity and excitement represented the majority, although those of a more disparaging tone, despite being present, were perhaps fewer given the ambition of the product. But criticism of any kind, given how nascent GenAI still is, is an indicator of success. The video many of us have now watched was the same as the flamboyant, unconventional pageantry we see on the catwalk at fashion shows. It was an illustration of what could be done with AI—if not all right now, then certainly in the near future. It was intended to be thought-provoking and inspirational.

DSPs have a challenge relating to AI in which merely telling customers how they’ve been applying it for years, despite being true, won’t be enough. Innovations will have to be transformative to cut through the noise, and this will come with the risk of division. Embracing this and leveraging it is going to be important for any company.

So what is ViantAI?

Based on the video and PR buzz, we can regard this as a fully integrated generative AI layer that sits natively across the DSP. In other words, the ambition, I think, is to provide a natural language interface that enables the end user to remove the drudgery of platform navigation, separate system usage, and button clicking, and instead have the platform seamlessly assist you from the planning through to the execution stages (coming soon). All you need to do in return is provide a few natural language prompts to guide it along the way. We didn’t see the part in which your AI-generated media plan would be pushed through to activation in the DSP, but we know that’s coming soon, and this, of course, is what makes the product much stickier and conducive to a “one-stop shop.”

But innovations like ViantAI—and there will be many more across all corners of the ad ecosystem—raise as many new and interesting questions as they answer. Let’s consider a few.

What does this really mean for planning?

Tempting though it might be to claim this vaporises swathes of the agency planning and buying roles overnight, the reality is likely different.

The Lump of Labour fallacy is the misconception that there is a fixed amount of work to be done, and that if some work is taken by a machine, there will be less work for people. But if it becomes easier or cheaper to do a job, it doesn’t mean that job disappears. If a planner can suddenly spend a fraction of the time doing things like exporting data, aggregating data, building plans, etc., they’ll be able to plan and activate more effectively, freeing up time to work on more campaigns, better optimisations, or more strategic efforts. A virtuous circle is created.

And so the question really becomes: what can you now get your domain experts to do with this new time? From a publisher perspective, there are many other high-value initiatives that could be focused on if operations teams had more time, like partner management and optimisation, identity testing, client servicing, new product innovations—the list is extensive. The common theme here is the freedom to move from operational to more strategic pursuits that are fulfilling and fruitful for both the individuals and the organisation.

How good does the underlying model really need to be?

My guess would be, a lot of the time, not as good as you’d think.

There are trade-offs to be had in anything, but when you offer an end-to-end product as extensive as this, you will accept this coming at the cost of other things. For example, previously, if as a planner/buyer using the platform, you’d meticulously plan your audience around the client brief at the expense of many other things, but now, with a few short prompts, you can achieve all that and more, you’ll likely welcome the new approach in which a fully scoped target audience can be provisioned for you. Does it matter if the segment size, audience profiles, or site breakdowns are sometimes wrong? How wrong does it need to be and how often before it’s unusable? How do these errors cascade down into the wider ecosystem? If the output of these models becomes the source of truth for some teams or agencies, then it doesn’t matter too much whether as a publisher or SSP if you say your segment size is X or your audience breakdown is Y. If the model gives Z, then that might be the pre-filter that doesn’t get you on the plan in the first place.

This isn’t a new problem per se. As an industry, I’d hope that we’ve all become well aware of the fragile foundations that underpin most measurement and attribution methodologies. This problem is potentially worsened when AI enables this to be done on a far wider scale than before and additionally if we put our entire faith in what the model gives back, rarely with any thought to interrogate the numbers.

How transparent will these models be?

Transparency in Adtech has become a tired buzzword, but as models become increasingly pervasive and encroaching in our day-to-day activities, we’ll continue to raise valid questions around how explainable they are.

Apollo Research published ominous findings in some respects based on research conducted on the new o1 model from OpenAI, which now has breakthrough reasoning capabilities. It found what it regarded in some cases to be human-level reasoning capabilities superior to that of previous GPT models and concluded clear evidence of scheming in which the model could fake alignment with developer goals while prioritising primary goals elsewhere.

Frankly, scheming is a terrifying concept, more pressing for industries outside of adtech, but nonetheless, if we are willingly giving greater control of our current workflow to the models, we have to ask critically: who is programming the incentives in the first place and can we even know this? If I’m a buyer and the DSP tells me that the best measurement tool is X, can I really be sure that it is? I’m not referring to hallucinations here in which it gets it wrong due to an error, but rather an intentional, goal-driven output misaligned with my own that might be hard for me to pick up. It’s one thing to have shady humans in ad tech; it’s another to have shady models.

What are the 2nd order effects?

If DSPs en masse offer end-to-end generative AI tools like this, it has the power to change everyday behaviors. What new products and customer needs does this create? If, as a planner, I spend very little of my time now manually planning campaigns and audiences, does this mean I spend a greater proportion focused on more consultative pursuits, and what new platforms or products do I therefore require?

Netscape CEO Jim Barksdale famously said that the only way to make money is either bundling or unbundling. ViantAI is a bundling of what are traditionally separate parts of the digital lifecycle, but this then probably gets unbundled at some point back into more verticalised AI-specific tools. For example, generative planning tools that facilitate deeper brand research synthesised with wider consumer trend behaviors.

Another 2nd order question is how incumbents innovate to accommodate new solutions like ViantAI. If some of my USP as a data platform is in the UX of my platform, what happens when this gets stripped away by a system that aggregates, or rather bundles, that feature into their product suite? I’ll likely need to pivot and focus on offering unique value in another way.

Innovations will come, behaviors will change, but the aspect that won’t be immediately obvious is how this impacts the wider environment we work in. This is an exhilarating concept.

The best we can all do now is remain curious and proactive in the AI space. You’ll do yourself a great disservice in avoiding this.