How agentic AI is rewiring marketing | Emir Caglayan
In Episode 1 of the NCMT podcast, hosted by the Nanyang Centre for Marketing and Technology at Nanyang Business School, Emir Caglayan, AVP of AI Strategy at WPP Media, overseeing the Google account for the APAC region, demystifies the buzz around agentic AI. Exploring its real-world marketing applications, he explains how moving from simple chat interfaces to autonomous, multi-agent workflows is redefining campaign management, creative speed, and the role of the modern marketer.
Demystifying Agentic AI: From Chatbots to Action-Takers
Since the widespread adoption of large language models (LLMs) in late 2022, AI has largely been viewed through the lens of a chat interface. We type a prompt, and the AI returns a text response. However, agentic AI represents a fundamental shift from simple generation to autonomous execution.
What separates a standard LLM from an agent is the ability to take action. As Emir Caglayan notes, referencing a definition by Google CEO Sundar Pichai, agents are “systems that combine the intelligence of advanced AI models with the access to tools so they can take actions on your behalf under your control.”
To understand this, Emir uses a factory analogy. The LLM is the brain or the core intelligence. The “harness” is the factory environment or nervous system that networks everything together. The “tools” are the specific equipment given to the AI to execute tasks. When stitched together, these elements form an assembly line where AI agents hand off tasks to one another to achieve a specific, strategic outcome.
The Six Layers of an Agent
Building an effective AI agent requires more than just prompting an LLM. It involves a systematic, multi-layered design. Emir breaks down the anatomy of an agent into six distinct layers:
- Model Selection: Choosing the right foundational model (e.g., Gemini Flash vs. Gemini Pro), assessing whether it needs to be more deterministic or more creative for the task (setting the temperature, i.e. the randomness of the response).
- Knowledge Base: Providing the grounding material. This is the specific corporate data, historical campaigns, and proprietary information the agent fetches from, ensuring it does not rely solely on its generic training data.
- Agent’s Role: Defining the agent's role through system instructions. This dictates its persona (e.g., “You are a social listening expert”), behavioural guidelines, and preferred output formats.
- Tooling and Capabilities: Equipping the agent to interact with the outside world, such as granting web search capabilities, API access to databases, or the ability to trigger other software.
- Guardrails: Setting strict operational boundaries to ensure brand safety, privacy compliance, and to prevent the agent from hallucinating or responding to irrelevant queries.
- Task Logic and Evaluation: Designing the sequence of operations and implementing evaluation frameworks to test the agent's output against predefined quality criteria.
Real-World Application: The Speed of Sports Marketing
The theoretical power of agentic AI becomes tangible when applied to complex, high-speed environments like sports marketing. Emir’s team deployed a multi-agent workflow during the ICC Women’s World Cup in India for Google.
Traditionally, capturing real-time cultural moments required a physical “newsroom” of eight to 10 cross-disciplinary experts monitoring eight-hour matches, brainstorming ideas, and pushing content. This was costly, resource-heavy, and often too slow. By the time a graphic was approved days later, the cultural moment had passed.
To solve this, Emir's team designed an autonomous, closed-loop workflow that condensed this days-long process into just 15 minutes:
- Social Listening Agent: Scans live social media conversations during the match to identify cultural triggers and trending topics.
- Sports Archivist Agent: Expands on the identified moments by pulling historical sports statistics to contextualise the event (e.g., “Is this the first time this record was broken in a decade?”).
- Synthetic Focus Group: Feeds the identified moments into simulated audience personas. These synthetic focus groups instantly signal whether the brand should engage with the moment or avoid it due to potential risks.
- Creative Ideator Agent: Harnesses the approved signals to generate three to five on-brand social media post concepts, adhering strictly to brand tone guidelines.
- Human Editorial Team: Reviews the AI-generated concepts, refines the copy, and deploys the final post with media backing.
“It’s a micro-workflow that represents what we do in marketing,” explains Emir. “Identify moments, test ideas, write based on the brand, and go live. And this was getting done agentically and really fast.”
The Evolving Role of the Human Marketer
While agentic workflows remove the cap on production volume, they introduce new challenges. Generating thousands of posts autonomously risks flooding the market with “AI slop”.
Because of this, the human marketer is not obsolete; their role is simply shifting from manual executor to strategic architect and coordinator. Emir references a crucial insight: “You can outsource your thinking, but you cannot outsource your understanding.”
Human experts are required to establish the initial strategy, maintain the agents, evaluate the data pipelines, and ensure output quality. As these closed-loop systems continuously learn from each campaign, they build a new baseline of corporate knowledge. The human's job is to ensure the AI learns the right signals and avoids drifting into poor performance.
Getting Started: Automating the Mundane
For marketing teams intimidated by the rapid pace of AI development, Emir advises against trying to boil the ocean.
“Just identify one painful, repetitive task that you have in your day and see how you can actually use AI to make it easier for you,” he suggests. Whether it is triaging emails, summarising data, or drafting initial briefs, starting small and getting hands-on is the only way to build the necessary intuition. You can read reports and listen to podcasts, but true understanding in the agentic era only comes from building and experimenting yourself.