The Agentic Loop AI Agents Reshaping Marketing Workflows and Decisions

Introduction to AI Marketers Guild APAC and Praveer Kochhar (0:05)

Welcome to AI Marketers Guild APAC. I'm Nicole Quayle, co-founder of the Asia Pacific AI Marketers Guild, which started three years ago in North America. We partnered to provide a regional lens, showcasing local AI pioneers, tools, marketing best practices, and world-class AI thought leaders through monthly webinars. Today, we introduce Praveer Kochhar, co-founder and CPO of Kogo Tech Labs, who has over two decades of experience across technology, community, and entrepreneurship. Kogo Tech Labs is building a sovereign AI platform enabling enterprises to deploy self-teaching agents with governance and control. This evolved from India's only AI travel expert app into something much bigger. We're exploring what this means for businesses, teams, and the future of work.

The Pendulum of AI and its Value (3:08)

I'm Praveer Kochhar. I'll split today's discussion into two segments. I've realized that AI is like a pendulum, constantly shifting perception; one morning it's good, the next it's bad. Yet, we all depend on it. Try switching off ChatGPT for a day to write a five-page proposal, and you'll see the value it brings. It's crunching something, building some benefit. We want to discuss what that benefit is and how to extend it. The topic of autonomous work can be scary, but it's a question I'm trying to answer as we build Kogo AI and Kogo Workspace.

Introducing the Agentic Loop Framework (4:06)

I'm going to introduce a framework I've developed and use when talking to C-suite customers about the value of agentic systems. This framework helps me push my understanding and development, especially for presentations like this. The core idea of Kogo was built around a manifesto: "Human potential meets AI." The concept is to expand human potential with agentic superpowers or AI, rather than the other way around. The agentic loop focuses on this: it's not about removing humans from the loop, but about putting AI to work on everything around the decision, with the human remaining at the decision point.

Traditional Decision-Making vs. Agentic Loop (6:14)

Let's talk about decisions. Pre-generative AI, Herbert Simon introduced "intelligence design and choice," coining "satisfice" – taking the first available option rather than the best. We also had the OODA loop (observe, orient, decide, act), heavily used by armed forces, where orientation to the environment is key. Then came pattern recognition, and frameworks like the DIKW pyramid. The truth is, the domain expert, the person taking the decision, was the only processor. The traditional decision loop involves collecting data, analyzing it, gaining insights, taking decisions, and acting, followed by a feedback loop. This process takes weeks, with multiple iterations, handoffs, meetings, and collaborations. By the time you act, the world has moved on. Today, with 3-second attention spans on social media, trends move much faster. While you could scale data and reporting, you couldn't scale interpretation and judgment. This is where AI brings great benefit.

The Agentic Loop: A New Operating Model (8:55)

This is the new way of doing things, the future of how we operate, and many of us already are. An agent handles data collection, analysis, and option generation. Web search was the first large-scale agentic system. Execution is also emerging for agents. The one thing only you can do is make decisions. The human stays at the decision point, which is the centrifugal force moving everything around it. The agentic loop aims to make decisions faster and enable more decisions per person per day, compressing the decision-making timeline.

Human-Agent Partnership Model (10:13)

This human-agent partnership model is where augmentation needs to happen. We are already doing this, but I'm outlining what it means. What we are good at, and what won't be replaced soon, is judgment, values, and wisdom. This will become more important as we deal with more agents, learning from them like co-workers, exponentially increasing our wisdom. The agent provides speed, scale, and consistency, acting as the fiber optic cable to data and decisions, while you provide the direction. For agentic systems to scale, three things must be enforced:

1. Bounded Autonomy: The agent works within the lines you draw.
2. Decision Explainability: The agent's decisions must be explainable to build trust.
3. Override by Design: You must be able to undo, redo, or stop actions. I use a sales development agent that generates personalized emails, but I have it put them in drafts for my review. I add my judgment based on business wisdom, which is more important than mass-scale personalization. The agent provides bandwidth, but I provide direction and decide where to step in. Our roles are not limited; we are promoted in terms of power. My 35-person tech team now architects pipelines, conducting code rather than just writing logic. Their role has shifted to governance. Organizational structures will become smaller, more efficient, with small teams building large companies, providing personalized, deep service, as marketing expertise and wisdom become paramount. Agents will not take these critical decisions anytime soon.

Autonomy as a Spectrum (15:15)

Autonomy is not an all-or-nothing spectrum; it's a range from level zero to four. You choose the level of supervised autonomy you want for specific tasks, whether it's human-on-the-loop, human-in-the-loop, or manual. It's often a combination. For example, if a budget shifts 5%, an agent can decide automatically with a set threshold. If it shifts 20%, the agent can be instructed to ask you. For creative refreshes, you can auto-generate or stay in the loop for every word. You define the rules as you build with these agents. The agentic loop compresses the insight-to-action loop across multiple dimensions by gathering and analyzing data. Your role changes to architect, conductor, and governor. Every loop makes the next one faster, smaller, and more efficient. And you control the spectrum of autonomy.

Applying the Agentic Loop to Marketing (17:24)

How does the agentic loop apply to marketing? We partnered with a company that built an entire marketing operating system on Kogo OS within Google Workspace. On Google.ai, you'll find a marketplace with 367 agency agents, including specialized ones built by our partners, working as an end-to-end marketing operating system. We want real experts to build real value. This agent's first rule is human-led strategy governing the entire loop. Its foundation is a brand OS and data, which the agent can gather online or be fed. It connects to data, has 17 pre-built skills (intel creation, commerce, brand positioning, etc.), validates output with a quality gate, and connects to external databases. The strategy flows clockwise, and data flows counter-clockwise, providing feedback. The purpose was to build a marketing agency as a system, not disjointed islands. Because the core agent has memory and institutional knowledge, it gets smarter with every campaign, learning and remembering detailed information. It's one human-governed system with 17 tracks, one foundation, validating every output and learning from every campaign. What stays human is your strategy, brand, budget, the decision to ship to a customer, final judgment, taste, and approval.

Case Study 1: Hyundai Ad Creative Generation (20:41)

Using this agent, I'll show the first demo. Our partner, Langur (part of the LS Group in India), a marketing strategy and brand consulting company, received a brief from Hyundai India. Hyundai has hundreds of dealers running ads across regions, facing challenges in scaling across dealerships, markets, formats, and messaging due to India's diverse languages and cultural contexts. The options were to give it to an agency for three weeks or feed it to the 367 marketing agency agent. The agent starts by gathering brand intelligence, then product intelligence, consumer intelligence, builds a creative framework, handles localization, writes the copy, and provides a final ad output in two hours. This efficiency is because the agent performs deep research.

Demo 1: Hyundai Ad Creative Process (23:26)

I'm inside Google Workspace, using an agent called Notse, which functions as a chief of staff, remembering everything and building interfaces. All agents in Google Workspace run inside virtual machines, having their own files, memory, and learning capabilities. Notse has everything slotted: research, learnings, observations. I can share it and provide files for it to learn on boot-up. Agents have preloaded skills, and you can build your own. This platform allows multiple workspaces with different agents. There's an agent marketplace where I'm highlighting the 367 marketing agency agent, a full-stack performance marketing agent with 17 pre-built tracks.

For the Hyundai example, we used a version of this agent. It creates folders based on its skills, like performance ad analysis, Google search term analysis, etc. Since we lacked internal Hyundai data, we told the agent to learn about the brand. Using its online research skill, it generated a detailed brand understanding document, covering personality, mission, leadership, origin, funding, target audience, emotional insights, product portfolio (India-specific pricing), brand pillars, tone of voice, key differentiators, vulnerabilities, distribution channels, customer voice, social proof, app ratings, direct quotes, and even identified gaps in information for deeper research.

Next, it generated detailed product understanding documents for each product, like the Ionic 5, including brand overview, product identity, spec sheet, positioning, competitive differentiation, core technologies, features-benefit matrix, usage modes, packaging, key selling points, and post-purchase experience.

For Ionic, it then moved to consumer intelligence, analyzing forums and reviews (like Team-BHP), creating a push-pull map, strategy pack, language bank, competitive map, buyer archetypes, and category landscape. The push-pull map showed what buyers avoid, what pushed them to purchase, and creative angles. The strategy pack detailed brand and product strategy, competitive landscape, recommended positioning, and creative strategy territories, along with seeds for social content. It built clear buyer archetypes: silent evangelist, range anxious realist, burnt early adopter, value skeptic, practical family driver, and tech forward charger. For each, it defined core anxiety, decision filters, and proof needed, forming the basis for a creative strategy. This strategy included brand context, consumer cohorts, competitive white space, tone of voice architecture, funnel messaging, proof architecture, creative production metrics, diversity and QA scorecard, and a 3-month retention pipeline.

Based on this, it generated ad copies for different buyer types, customized to specific markets like Mumbai, Delhi NCR, and Bangalore, changing narrative, content, and cultural context. The final output included visual directions for a vision model. The agent integrates image generation, so you don't need to switch tools. For example, a creative for Mumbai featured the Worli Sea Link, and because of range anxiety, the CTA was "owner reports" instead of "book a test drive." This entire end-to-end process, showing what AI can do with human governance, was generated in two hours. This is being replicated across multiple brands to reduce the marketing loop.

Case Study 2: Dove Social Intelligence Engine (35:39)

Langur developed a different strategy. They took Agent Zero, our general-purpose agent (like a college graduate with basic skills), and taught it new skills, transforming it into a social intelligence engine. They built strategy dashboards with it. The brief was to identify and track social media trends before they become big, monitor conversations on a single dashboard, turn them into opportunities or threats, and gather insights based on brand DNA, detecting and categorizing signals.

Demo 2: Dove Social Intelligence Dashboard (36:02)

The result was a real-time signal architecture for brand teams, moving from raw social noise to crisp, actionable brand decisions. We gave the agent capabilities to use scrapers and integrate with external platforms. It performed deep research, analyzed, and built a dashboard in about eight hours. This dashboard tracks category trends, brand and competitor signals, consumer and macro signals, and influencer/creator signals. Its sources include social platforms (ShareChat, Reddit), Google Trends, beauty/lifestyle blogs, and news. It generates real-time alerts, weekly digests, brand intelligence briefs, monthly reports, competitor activity radar, and ongoing monitoring, providing brand-specific signals and priorities. You can go deeper into specific trends with volumes, seeing what's growing (e.g., self-care Sunday, skin barrier repair, everything showers). It categorizes what to act on, prepare for, and monitor. It also shows platform distribution and sentiment for trends, top hashtags, and product mentions, based on available data. We architected the desired output, and the agent built a system that tracks and gathers this information in real time, including cultural, macro, micro trends, and seasonal moments. This intelligence can then feed into another agent with the brand DNA to generate creative briefs, meta-copy, or creatives. This demonstrates how the entire pipeline, from monitoring real-world events to creating effective creatives, has shrunk.

The Compounding Value of AI Systems (40:19)

The agentic loop is shrinking. Currently, we are the limitation because we struggle to zoom out beyond individual tasks, often just prompting AI hoping it remembers. We are missing the compounding value of AI systems that store all my data, memory, and past work in one ecosystem. This means not pulling/pushing files between tools, but working in an environment where everything learns and compounds, making life easier. The loop becomes shorter and more efficient, freeing up time for travel and exciting decision-making work, rather than laborious tasks like staring at spreadsheets.

Kogo's Competitive Strategy and AI Sovereignty (42:33)

Regarding competition with models like Claude, everything I showed was built on open-source models; we do not use closed-source models on Kogo AI. Kogo is a proprietary sovereign platform. It's 30x cheaper than other models, with no limits on sessions, and we aim to offer unlimited tokens, which is the holy grail of AI. We believe AI should be private, sovereign, and open-source. For example, the latest open-source Chinese models like DeepSeek focus on efficiency through caching, where 95-99% of generated tokens are cached, drastically reducing GPU consumption and cost. This breakthrough in caching means only a small percentage of tokens are regenerated, making models much cheaper. AI will become cheaper, commoditized, private, and open-source will eventually win. We are flag-bearers, showing that we can achieve the same capabilities with open-source models, even at 92% benchmarking versus Claude 4.7. Privacy is paramount; I control my data and conversations with my AI, which I see as a great freedom. It's not rented intelligence but an "intelligence tradeoff" where you feed agents your intelligence for subsidized tokens.

Advice for Getting Started with AI (48:21)

Many people experience data overload and analysis paralysis with too many AI tools, diving in, freaking out, and stopping. I wouldn't stop the "freak out" because chaos often leads to clarity. Everyone should go through that realization. Use everything available, but be conscious of AI's compounding effect. Ask yourself: Is what you're doing compounding value and making the loop smaller and more efficient daily? Jensen Huang's philosophy of moving the needle a little bit every day, leading to a compounding effect, applies here. If you analyze your AI strategy with this parameter, it will yield dividends. A simple trick when you sign in: add Agent Zero as your first agent. Then, tell it to research "best prompt generation techniques" and create a skill for you to generate prompts. You'll never have to write a prompt again. This is exciting, building a pathway for people who might not know how to build the right prompt or get the best use out of an agent.