In January, dentsu announced the launch of Generative Audiences, a high-definition audience and personas simulation tool for media planning and buying. Dentsu hopes the development of the tool will help marketers move away from siloed audience segmentation towards AI-generated audiences, that can think and act like real consumer groups – shortening the time frame from planning to activation.
To find out more, FutureWeek caught up with Ben Gott, dentsu’s President of Data and Tech.
What makes Generative Audiences special?
It’s an insight and activation tool. For a long time, we struggled to bring insights and activation together as an industry.
Over the last five years, we’ve been seeing that the quality of media campaigns relies on the ability to make bespoke copy and ads.
It’s interesting because it sits at the intersection of media and creative. Because it sits at the intersection of these two workflows, you can build an immediate plan, audience and creative in a single workflow.
What is the AI functionality within the tool?
We’re not about enriching an off-the-shelf LLM. This has been a bespoke build from the ground-up. We have built a large language model – what we call a large knowledge model.
The AI interventions come in three places. This includes modelling, so being able to create huge models quickly which requires a large amount of compute and use of AI. The natural language interface, which allows teams to interrogate massive amounts of data in day-to-day language. Then there’s the agentic workflow, which allows us to string together that media value chain so that the manual steps of carrying an audience over to the team that builds or activates can be carried out with an agent.
Is this an evolution of targeting as it’s always been? Or is it a complete revolution?
We think this is new. Shortening insight to action is really powerful. The ability to move at pace while maintaining proper analytics and transparency is genuinely new. In the past, we’ve had a trade-off between pace and rigour – so you could move fast by activating an audience directly in Meta or Google, or by pressing a black-box button.
With this, you couldn’t ensure everything was privacy safe, or ensure that you’ll actually get good results. So this provides a new model.
I think the industry has needed a new model because consumers are very aware of how their data gets used.
How does it fit into the wider dentsu tech stack?
It’s part of our dentsu.connect operating system, which is an end-to-end media and creative workflow tool. It helps our over a thousand media activation and planning people in the UK. It helps them to bring insights and data into their day-to-day workflow – whether that’s a brief landing from a client or activating something live like planning growth, audiences or media.
What data are you using to power the technology?
We are able to match a client’s first party data if they want, and that’s obviously in a way that is privacy safe so it doesn’t leave their environment. We overlay that with a series of very high quality signals. We tend to group these high quality signals into categories, such as cultural signals and what’s happening in the world, like the Winter Olympics, for example.
We overlay commerce signals, which is what people in the market want to buy. We match to context signals, which is what people are doing on the internet at that time. Then we also build in behavioural intent, demographic overlays, and a GEO overlay.
The key thing is that we can configure at an audience level which of these signals we want to use.
How are you approaching data quality and modelling when building these AI-generated audiences?
Deterministic data sets can be a bit shaky when you start to extrapolate continuously, while third-party data has been under some strain for a while. We’re trying to blend both of those techniques so that a user can get the best of deterministic when precision is needed, but you can get the scale and reach of something more signal-based.
What makes this really powerful is the modelling power you can put behind this with AI. Our heritage with data and analytics means that we’ve been running and optimising these models for decades.
We’ve pulled in really expansive data sets and a strong blend of signals – not just IDs. We then overlay that with the ability to have transparency and to calibrate the results. So users aren’t just pressing ‘Go’ and hoping.
How are you mitigating things like stereotype and bias reinforcement?
This is an issue that we’re very aware of and we have built from the ground-up to guardrail us from this as much as possible. As an industry we have been working on pen portraits for years, which can be horribly cringy and stereotyping at times.
We as an agency needed to have control of the models we’ve built. Only building on top of stock LLMs can be horrible for having Western bias built into it, for example.
We have anti bias training for all of our developers, and we build and test against that.
When we have been building these models, we’ve been able to calibrate the data against real-world data and outcomes, so we can understand whether any hallucinations or bias are creeping in.
Are there any new skills planners and strategists need to have to use tools like this?
This might be obvious now, but planners and strategists need to be AI literate. We’re very focused on making sure our teams have the right tools and know how to use them effectively. We know there’s a style of natural language that works best with AI technologies – so things like knowing how to prompt well is important.
This tech allows for planners and strategists to elongate their roles. This has happened already in Martech but the Media industry is catching up. We now have planners and strategists acting as analysts because they’ve got these incredible models at their fingertips and they’re able to interrogate those models in their own language.
Planners and strategists have always been data-driven and understood the need to grasp insights, but this is an absolute step-change in their ability to do that. They don’t need to understand SQL or Python code, and they don’t need to rely on Data Scientists who sit on the other side of the room.
But it’s not just about a skillset as much as it is about a mindset and with great power comes great responsibility. We make sure our team knows what *model overfitting means and what the downfall of this could be etc.
How does it fit into a marketer’s traditional way of working?
If we split insights and activation, a senior marketer will be less focused on the activation end, they’re going to be on the insight end. This kind of tool brings insights into the organisation in a way that everyone can understand. It spreads insights across the organisation. We think about this as a co-pilot and an extension of them and who they are.
For example, a head of digital can use the tool to explain to a CFO what the audience they’re targeting is and why, because they can have a live conversation with the audience about their products and services.
It’s also being used in Sales teams at motor dealerships and in insurance businesses, for example, so that those selling the product can get insights despite being removed from the HQ strategy.
*when a model learns training data too closely causing it to perform well on training data but badly on new, unseen data.
