Sahit Garapati leads personalized guidance and AI-powered automation at Amazon Ads, where his team builds the systems that help 1.5 million advertisers launch and optimize campaigns. His work sits at the intersection of recommendation science, product design, and advertiser trust, a combination that demands both technical depth and a genuine understanding of how brands actually spend money.
That understanding comes from an unusual career path. Before joining Amazon, Sahit spent years on the brand side at Kimberly-Clark, where he managed P&Ls for household names like Huggies and Depend, and built machine learning tools that reshaped how budgets were allocated across retail media networks. The experience of justifying every dollar to a finance team now informs how he thinks about building products advertisers will actually trust and use.
We spoke with him about what autonomous advertising really looks like at scale, why most AI pilots never make it to production, and how his brand-side background shapes what he builds for advertisers.
You lead personalized guidance and AI-powered automation at Amazon Ads for a large and diverse advertiser base. What does autonomous advertising actually mean at that scale?
This is the central question for product leaders in this space. The term autonomous is often used interchangeably with simple automation, but it represents a much deeper strategic shift. That shift is driven by the explosion of insights and the growing complexity of digital advertising. When you are operating across a global advertiser base and billions of impressions, the number of micro-decisions involved, such as which bid to place, which creative to show, or which audience to reach, quickly becomes impossible for a human to manage directly. This leads to a common misunderstanding. Autonomous does not mean fully hands-off. Instead, it reflects a change in the role humans play. We are moving from a world where people manually manage campaigns to one where they manage intelligent systems that operate campaigns on their behalf. In this model, advertisers act as strategists. They define objectives, provide high-quality creative inputs, and set boundaries and expectations. The AI operates within those constraints, executing decisions at a speed and level of granularity that humans cannot match. I often compare this to advanced driver-assistance systems in vehicles. These systems are designed to handle complex, real-time conditions while keeping a human in control. At scale, an advertising system must do the same with data. This creates a trust-and-control paradox. Advertisers benefit most when they allow the system to handle moment-to-moment execution, but they need transparency, guardrails, and reporting to feel confident doing so. As a result, the real product challenge is not just building better models. It is designing interfaces and controls that help people understand, guide, and trust the AI.
Recommendation systems can behave unpredictably as they grow. What problems did you see as you scaled AI-powered guidance to a much larger audience?
Scaling AI systems is rarely linear. It is usually a series of breaking points that force architectural change. At smaller scales, systems can rely on precise, brute-force calculations for tasks like similarity or ranking. At enterprise scale, those same approaches break down due to latency and computational constraints. The first major shift is moving away from perfect answers toward fast, reliable answers that preserve intent. This is not about lowering quality. A good enough match means a statistically reliable, intent-preserving decision that meets defined quality thresholds while operating within real-time constraints. To achieve this, systems rely on abstraction. For example, instead of comparing every individual to everyone else, they group behavior into meaningful neighborhoods and reason within those contexts. This allows the system to move quickly without losing relevance. At that scale, the system no longer reasons about people as isolated individuals. It understands them in relation to broader behavioral groupings that reflect shared patterns and intent. The challenge is ensuring that this abstraction does not dilute what matters most in the moment. For a long time, the industry responded to this challenge by scaling what you might think of as the system’s memory, meaning the representations and signals learned during training. While those inputs grew dramatically, the reasoning capability of the models themselves plateaued. The next shift is about scaling reasoning, not just data. By building more expressive models that can learn from sequences and evolving context, systems can extract more value from the same signals. This inversion, scaling intelligence rather than just inputs, is foundational to the next generation of recommendation engines.
You’ve talked about speeding up experimentation. What slows teams down the most, and how do you remove those blockers?
In AI product development, the experimentation cycle is the innovation cycle. Velocity here is a competitive advantage. When teams move slowly, it is usually a sign of deeper structural issues rather than a lack of ideas. Organizationally, teams are often constrained by limited experimentation capacity or stuck waiting for statistical certainty in low-traffic tests. Technically, there is also a real cost to maintaining multiple experimental paths in production. But the most significant bottleneck is that many experimentation platforms were not designed for AI systems. Traditional A/B testing tools work well for deterministic changes, such as testing a new UI element. They struggle with probabilistic and generative systems. Comparing stochastic outputs, defining success for generative responses, and managing the cost of running multiple model variants all expose a mismatch between the tools and the problem. The solution is investing in AI-native experimentation platforms. These systems are as much about risk management as they are about speed. When teams can roll out changes gradually, analyze performance in near real time, and test complex model variants safely, the cost of failure drops. Lower risk encourages experimentation, and that confidence is what ultimately creates velocity.
You’ve built recommendation ranking models that blend immediate context with longer-term patterns. How does a system like that determine what to surface in the moment?
The system is constantly answering one question: who is this user right now? It does this by balancing longer-term behavioral patterns with immediate, in-session signals. Long-term data provides stability, while real-time behavior captures emerging intent. The challenge arises when those signals conflict. A sophisticated ranking system does not treat them as competing truths. Instead, real-time behavior becomes the lens through which historical patterns are interpreted. The system elevates the parts of a person’s history that are relevant in that moment and down weights what no longer applies. Real-time intelligence is expensive, so product leaders must be deliberate about where it delivers meaningful value. Knowing when immediacy truly matters versus when slower signals are sufficient is what separates systems that feel genuinely responsive from those that look impressive in theory but feel disconnected in practice.
You manage a large cross-functional team of engineers, scientists, and designers at Amazon Ads. How do you keep that many people moving in the same direction without slowing down?
The biggest threat to velocity is organizational silos. In AI, where data, models, and business outcomes are tightly coupled, misalignment quickly becomes costly. The most effective approach is decentralized execution anchored by a centralized vision. Leadership defines the problem and the priorities. Once those are clear, small, multidisciplinary teams own the solution. Disney offers a useful example of this model by bringing together creative, technical, and business experts to solve focused problems, such as improving how content adapts to audience context. When teams are aligned around a shared outcome, success becomes collective rather than functional. This structure changes incentives. Instead of optimizing for individual contributions, teams optimize for business impact. That shift is essential for maintaining speed and coherence at scale.
You’ve successfully secured executive sponsorship for multi-year AI initiatives. In a world focused on quarterly results, what makes executives say ‘yes’ to a long-term strategic bet on AI?
Executives are fundamentally focused on risk. Securing sponsorship is less about selling a vision and more about presenting a credible framework for managing uncertainty. That framework must clearly connect the initiative to business objectives, identify high-value early use cases, and define measurable baselines for success. The most effective proposals balance ambition with pragmatism. They show how long-term transformation is achieved through a sequence of short-term wins, each one funding and validating the next. This approach respects quarterly realities while enabling sustained investment.
Many AI products generate impressive results in a pilot but fail to drive real, scalable revenue. What separates the AI initiatives that successfully transition to production and drive business value from those that remain stuck in “pilot mode”?
This is a critical challenge in our field. We often see a divide where, despite significant investment, many organizations struggle to get measurable P&L impact from their AI pilots. This divide is often not driven by model quality. The problem is how these initiatives are built and deployed. Many pilots are run incorrectly. A common mistake is treating the pilot as pure technical validation, a science experiment to prove the model works in a sterile lab. This approach is problematic because it fails to co-create an ROI model with the customer or align with day-to-day operations. The organizations that successfully cross this divide do something fundamentally different. They do not build models; they build adaptive systems. The initiatives that stall often result in brittle tools that are misaligned with day-to-day operations and have no learning mechanism. In contrast, high performers are far more likely to have fundamentally redesigned individual workflows around the AI. They treat the initial deployment as an extension of the pilot and build dashboards to institutionalize the success criteria. This is the core difference. A pilot that stays in pilot mode is a technical artifact that proves something can work. A product that drives revenue is a fully integrated business process that proves it does work, every day, at scale. The gap is between a demo and a process.
You started in brand leadership at Kimberly-Clark before moving into product and tech at Amazon. How does understanding the brand side change what you build for advertisers?
That background provides a critical, and often missing, empathy for the customer. It is the perspective of a brand manager who has to go into a meeting with their finance team and justify every dollar of their spend. This experience fundamentally changes product priorities. For example, a product manager focused purely on the technology might prioritize a small model accuracy boost, but my consumer packaged goods background might lead me to focus on the trust deficit with the channel as a much larger business problem. From a brand’s perspective, there is a core tension around transparency. Brands often feel they are operating in a black box, unable to see the attribution algorithms or which impressions they truly bought. This means trust must become a core product feature. The channel must prioritize transparency with clear attribution dashboards and enable accountability by building measurement products that instill confidence in them. Finally, the product must be designed for the “internal sell”. The product’s job is not finished until it generates a clear, defensible report a CFO of a brand can understand.
You built machine learning tools at Kimberly-Clark that reshaped budget allocation across retail media networks, then moved to Amazon Ads to build the channel side. What do most advertisers misunderstand about how AI-powered solutions actually work?
The most significant misunderstanding I encounter is a reluctance to fully lean in. Many advertisers are dismissing the technology, viewing it as an unreliable novelty or unsuitable for their core use cases. This hesitation often stems from a belief that they must wait for a perfect or finished version of the technology. This is understandable, as these models are evolving at an unprecedented pace within a relatively short amount of time. Because the technology is still improving, advertisers believe they must wait until it is perfectly reliable. This is a critical error in my opinion. The advertisers who are breaking through growth ceilings are not waiting for perfection; they are the ones who started with their existing assets and learned through implementation. They understand the need to iterate and improve along with the models. Every month spent waiting for this perfect tool represents a lost revenue opportunity. You do not wait for the technology to be ready. You engage with it, learn from it, and iterate with it to push the boundaries of what is possible for your brand.
Where do you see AI in advertising heading over the next few years, and what should product leaders be preparing for now?
Two trends stand out. The first is the rise of agentic systems that can coordinate increasingly complex workflows while operating with clear human intent, oversight, and guardrails. These systems are designed to raise the ceiling on what advertisers and agencies can accomplish, not replace strategic decision-making. The second is the growing importance of trust and governance. As AI becomes more capable, expectations around transparency, data stewardship, and accountability rise alongside it. Product leaders must design systems where governance, human review, and explainability are built in from the start. Without trust, even the most advanced AI will fail to deliver value.
Disclaimer: Sahit is speaking on his own behalf and his opinions do not necessarily represent the view of Amazon.
