AMA: TikTok Head of Product, AI-powered Automated Services, Adrianne Wang Martinson on AI Product Management

One of the biggest challenges in shipping agentic AI is deciding how much autonomy to grant the agent—and doing so in a way that is safe, predictable, and enterprise-ready. Unlike traditional AI features that primarily focus on prediction or generation, agentic systems are responsible for understanding intent, planning actions, and executing decisions. This makes autonomy a first-order product decision. Too little autonomy and the agent becomes a brittle assistant; too much autonomy and it intro ...Read More

A common approach building agentic workflows is to focus on model capability; however, the true differentiator is the operational infrastructure that allows agents to act safely and predictably in the real world. The most critical pillars are: 1. Observability: From Black Box to White Box Agentic workflows involve non-linear, multi-step reasoning. Without visibility, failures feel random and unfixable. Infrastructure Requirement: We move beyond simple request/response logs to Traceable Reasoning ...Read More

Since ChatGPT went viral in late 2022, raw model intelligence has rapidly become a utility. Access to powerful language models is no longer a differentiator—it is a prerequisite. To build a durable AI product, teams must pivot from “what the model knows” to “how the system behaves.” The Table Stakes: These capabilities are now the electricity of modern software. If you have them, you’re in the game—but not necessarily winning it. Frontier Model AccessEvery competitor has access to the same under ...Read More

A dataset is “good enough” when it supports learning, evaluation, and improvement, not when it’s large or perfectly clean. Three early signals matter most: Coverage – The data reflects real-world behavior, including important edge cases, not just ideal scenarios. Clarity – Humans can consistently agree on what “good” looks like, or disagreements lead to clearer rules. Feedback – There is a measurable outcome that allows the system to learn and improve over time. If the data enables these three t ...Read More

This is not a binary build vs. buy decision. In the agentic AI world, the right approach is build and buy. A useful way to think about it is: Buy the intelligence (foundation models) Buy the plumbing (integration and orchestration infrastructure) Build the brain (how the agent reasons, retrieves context, and executes workflows) Most Orgs/teams should avoid building commoditized layers that evolve quickly and offer little differentiation. Buying these allows you to move faster and focus resources ...Read More

AI systems are probabilistic, so applying traditional product playbooks often leads to fragile products and misplaced confidence. The 4 Most Common Mistakes I have seen are: 1. Launching fully autonomous agents too early PMs sometimes ship AI systems that make end-to-end decisions and take actions without human review from day one. This fails because agent failure modes are impossible to fully predict upfront. What works instead: Start with human-in-the-loop designs and increase autonomy gradual ...Read More