We've pulled the key exchanges from the session covering how autonomous AI makes decisions, why bolting intelligence onto legacy systems creates ceilings, and what genuinely separates native AI from the noise.

Start with the Q&A below to get straight to the insights, then catch the full session on demand to see autonomous AI in action and watch the live platform demo.

Evaluating AI Solutions and Their Competitive Advantages

What's the most common misunderstanding people have about AI?

There are probably a few, but one of the big ones is that AI is a singular thing. The long-term goal of AI would be a single AI system that could do anything humans are capable of. But for all the power we have from large language models today, they're still quite different from how our brains work.

When we look at the different applications of artificial intelligence, there are many algorithms that can be used for pattern detection, decision-making, and controlling or optimizing systems. We're still in a world where you have to pick and choose the algorithms you want to use, and pick the ones that are appropriate for the problems you want to solve. And appropriate doesn't just mean does it give a good answer. It's does it give that answer quickly enough? 

In programmatic advertising, we're making billions of inferences every second. Large language models aren't fast enough for that, and they're expensive. Sometimes picking a machine learning algorithm that is much more narrow in its general utility but can respond much more quickly and is much more efficient actually makes it accessible from a cost perspective.

Does Quantcast use LLMs for autonomous advertising?

We do use LLMs. Large language models, things like Gemini, ChatGPT, Claude, and so on. We use them for a lot of the interface capabilities: assessing and determining how suitable creatives are for driving outcomes, interpreting data, and translating requests into charts.

But in terms of the autonomous advertising itself, the decisions around the appropriate audience, context, pricing, and optimization, large language models aren't really well-suited for that. A lot of that analysis is based on numerical data. It's mathematical in nature. 

The most important thing is the number of decisions that need to be made and the speed at which they need to be made. We have to make decisions about millions of new opportunities on behalf of our clients every second, and we can only wait milliseconds. 

What we've utilized large language models for is putting into words and helping to verbalize and visualize some of that complexity, in order to help get a little bit closer to understanding it, while knowing that the underlying system is way better at making those decisions at that speed.

If we use the AI system for research, can data be kept confidential and safe?

Absolutely. When marketers are using our system and setting up their predictive audience models using Audiences by Q, that data, your first-party data as a marketer, is used only to support your advertising models and your campaign execution. It's not available to anyone else.

How is AI different from machine learning, and what does that mean for how platforms operate?

The way I think about this is in terms of the configuration, execution, and optimization of an advertising campaign. Marketers have clear goals. They have their own data in terms of high-value engagements, conversions, and the returns that advertising campaigns are delivering. That's really clear instruction for AI systems in terms of what they're optimizing to.

One of the things that's really powerful in using large language models on top of this is the ability to interpret and help understand what's taking place and get a glimpse into how these systems are making their decisions. 

AI systems are powerful at certain decision-making because the way they make decisions is different from us. Our intelligence evolved over millions of years. A lot of what we think of as human intelligence evolved to make us more effective in our three-dimensional physical world. We tend to find patterns in a lower number of dimensions than machines are able to.

When you use large language models, your inputs are converted into hundreds or thousands of dimensions. That's led to some of the black box concerns about AI. But we've found really effective uses of the interpretation capabilities of large language models to navigate these immense data spaces and extract explanation and insight, and provide them in a way that we can understand. 

These techniques can do things we're not capable of, but applications of AI can also help us gain more confidence and trust in the decisions that have been made, and produce insights that can inform other aspects of decision-making, whether that's marketing strategy and messaging or even customer understanding that leads to better product development.

With so many AI tools flooding the market, how can you tell what truly sets one apart?

You have to look at it in the context of what you're trying to achieve. It doesn't matter what buzzwords get used. The proof is in the pudding. If you're a marketer with a clear performance goal and you've got online conversion events, if you try Quantcast relative to your incumbent DSP, we will deliver substantially better performance. The system is so easy to test. 

Getting into why you'd get differences in results, it really comes back to something we've talked about today. Large language models are amazing and will be applied to all sorts of software, but they are fundamentally constrained by the architecture of that software. If that software is set up to enable you to pull levers, even a large number of levers, then all a large language model can do is pull those levers.

But if you've reimagined your architecture to operate across immense data sets, trillions of records, and algorithmic training that can deal with tens of petabytes of data processed every day, you have the opportunity to do something far beyond human imagination. It's fundamentally rethinking the way you solve problems from the bottom up, from the data and the outcomes, rather than from the top-down hypothesis.

Think about the difference between getting into an Uber versus a Waymo. In an Uber, the experience is down to the driver. You might have a good ride or a bad ride. You provide feedback, that feedback gets captured, and maybe the company tells the driver someone liked or didn't like their drive. It's a very inconsistent way of trying to make improvements.

When you're in a Waymo, that Waymo is making millions of decisions a second. Every one of those decisions is instrumented. Every single ride can get better from the next, because autonomous systems aren't just making decisions; they're recording the decisions made and the results they deliver. That provides incredibly rich context for asking questions, understanding what's going on, finding new insights, and ultimately improving.

A/B Testing and Campaign Optimization: How Autonomous AI Changes the Approach

Two of the most practical questions from the session focused on how Quantcast approaches testing and campaign structure. If you're evaluating how autonomous AI fits into your existing workflow, these are good places to start.

How do you factor in A/B testing in campaign planning?

Being able to run A/B tests is such a great way to prove out different approaches. It's a very important strategy we've used with our customers. Even when talking about a traditional approach versus [Quantcast’s] Audiences by Q approach, having a rigorous assessment of the differences in impact and performance will be the real proof in the results.

In almost any endeavor, your rate of progress and advancement is proportional to your rate of experimentation. One of the challenges with the legacy DSP approach is that it is so time-consuming to set up and sustain campaigns, and to constantly monitor and change them, that whilst you might have a great idea for an alternative media strategy, executing and testing it means doubling the work. If you can make it cheap enough to do those experiments, you can do many, many more of them and find the new approaches that work really well.

A huge advantage of having a solution that can autonomously configure and optimize campaigns is that you're free to use your imagination and run additional tests. Built into the Quantcast platform is a robust A/B testing capability that lets you set up tests, manage different splits, and compare results so you can take advantage of that autonomous execution.

What buying models does Quantcast use?

It's really about the optimization objectives. Campaigns using the Quantcast platform are ultimately bought on a CPM, or typically a dynamic CPM, because having that flexibility means every time we put in constraints to a machine learning or AI system, we run the risk of overly constraining it.

When you start with a traditional DSP and an expert system, they don't do anything out of the box. You have to create rules. As the operator, you know that if you refine and add more rules, your performance gets better. With our system, the goal setting is typically around an outcome. 

A classic example would be a conversion. You're optimizing towards a conversion, you want to get the lowest CPA, or you're optimizing towards ROAS. AI systems are goal-based systems. They're driven by incentives, just like we are. The more specific we can be about those incentives, the more powerful the solution can be for you.

You don't necessarily have to average all of your customers into a single CPA. If you know you have certain segments that are more valuable, you can instruct the AI system through either passing that information through our live tag or using a conversion API. You can set up distinct predictive models and campaign optimization for different audience subsets that are most valuable to you.

You can watch the full webinar on demand here.