Developments in Machine Learning and AI, supported by Large Language Models (LLMs), have made programmatic advertising faster and smarter in recent years – most notably when it comes to analyzing content so that brands know where to place their ads.
But the adtech ‘holy grail’ – real-time responsive ad targeting, which moves across platforms and scales to find relevant new audiences without using personal data – remains a huge technical challenge. In fact very few companies have full end-to-end systems that meet this vision in either performance or scale. Could agentic AI be the answer?
Autonomous operations
Agentic refers to AI systems which can operate and pursue goals autonomously. These systems, or agents, can develop pathways and make decisions towards an end goal, learning from closed-loop feedback and adapting accordingly along the way, with limited human oversight. Unlike traditional models that follow fixed instructions, agentic systems proactively determine how to achieve their goals based on live inputs, such as the contextual environment or performance signals.
AI companies have been building these capabilities for some time, but agentic options are only now being trialled in live business scenarios ranging from customer service to medical research and cybersecurity. Businesses can now expect to see ‘agent’ options everywhere from Apple, Google and Amazon to Microsoft and Salesforce.
An agentic vision for programmatic advertising
In adtech, agentic AI is shaping up to be a strategic layer to the automated buying and selling process, steered by guardrails and KPIs set by the advertiser and by data inputs such as known customer interests, performance metrics and contextual patterns.
The ideal AI agent or system would then learn from these data inputs in a closed-loop system, making real-time decisions about buying and selling according to what’s actually working during live campaigns. The objective is to move towards autonomous decision-making that enables smarter, more responsive advertising, with smooth customer personalization and enhanced commercial outcomes.
Adtech has undergone some fundamental operating shifts over the last five years, which on the one hand have limited the use of third-party cookies and customer data, but on the other hand, opened up new opportunities in nascent areas such as Connected TV, Digital Audio, DOOH, in-app advertising and social media.
This type of fluid, intelligent agentic-style AI is now being discussed as a viable way to build privacy-friendly ad-targeting systems which also meet modern-day consumer trends by working omnichannel across platforms.
However, data processing speed remains challenging. Understanding multimodality and categorizing text, audio, video and other data points quickly enough to allow real-time targeting decisions requires significant technical architecture. Building systems that can respond with sub-second latency, while maintaining accuracy and brand suitability, is one of the most significant hurdles in this field.
A live, working model
While the ad industry has just started talking more widely about agentic, AI company Illuma has been building the foundational elements of this for over a decade. An agentic approach – combining reinforcement learning, unsupervised models, and LLMs – is already being used to power a system that learns which contexts are proving successful for a campaign, adapts in real time and autonomously scales to proactively identify and target similar opportunities.
While many agentic AI systems are general-purpose or research-focused, Illuma applies agentic AI specifically to advertising, and offers some early real-world learnings for future developments in this space.
For example, the Illuma data team has managed to balance classification accuracy with the need for low-latency response by using transformer-based language models. These summarize performant URLs and use a computationally efficient algorithm to match them with other similar content, avoiding the need for unrealistically large data training sets.
Using this method, the Illuma system is currently processing millions of queries per second. This sort of high-speed processing is the baseline required to perform real-time learning and adaptation, which is a critical foundation step for advancing agentic AI in programmatic advertising.
Privacy concerns and cookie deprecation
Traditional audience-based targeting uses historical browsing patterns as the primary data input. For example, an individual looks at content about running shoes; so for the following weeks or months, they are targeted across the internet by sports retailers.
Agentic AI offers a privacy-friendly alternative by shifting the focus from personal data to real-time contextual interests and first-party signals. In this example, the sports retailer can ditch the historic data and instead keep up with contextual interests and trends as they evolve throughout live campaigns, with the AI agent expanding to find more people with similar interests, without ever needing to consider, store or use their personal identity.
Multimodal learning for Connected TV
LLMs have now evolved from simple models to multimodal systems that can understand more than just text. As a result, contextual technologies are now able to classify rich media including Connected TV, as they are capable of extracting audio data, separating background music, and going beyond simple data analysis to achieve a near-human understanding of the visual action taking place on-screen.
Using this information, an AI agent can then take goal-seeking decisions into Connected TV at scale; moving campaigns between platforms towards high-performant content. For example, if one context type is proving successful for a campaign on the open web, then autonomously pivoting the campaign into similar opportunities in Connected TV – all with limited human prompting.
As Connected TV continues to grow, fuelled by the shift away from linear TV and the rise of streaming platforms, the ability of agentic AI to operate across screens and media types will be a key differentiator for future-ready adtech systems.
The road ahead – accuracy, transparency and compliance
As with all AI processes, the quality and usefulness of an AI agent’s output depends heavily on the quality, granularity, and timeliness of its inputs. So as more advanced multimodal large language models (LLMs) are integrated into programmatic systems, we expect to see multiple agents working together, reviewing each other’s output, with the effect of continually driving up the accuracy of the foundational material used for their decision-making.
At the same time, as media owners move towards a more unified data framework for Connected TV and other media types, richer cross-platform signals will emerge, enhancing the effectiveness of AI-driven recommendations for advertising across platforms.
Of course, challenges remain and will continue to evolve as the technology develops. For example, bias and fairness in algorithmic decision-making must be carefully monitored and addressed, especially when agents operate autonomously at scale. Data privacy and compliance will require continual vigilance and innovation as regulations evolve, and agentic decisions will need to be understood and audited in order to ensure advertiser trust and regulatory compliance. While autonomy is a defining feature of agentic AI, human-in-the-loop frameworks may still be needed in certain high-stakes or sensitive scenarios.
As early adopters in this space, Illuma has learned that agentic AI is not a silver bullet—but it is a powerful new foundation. When embedded thoughtfully and guided by clear business objectives and ethical constraints, it’s clear that agentic systems can offer an exciting leap forward for advertising technology.
