Article
June 9, 2026
Google, Product Manager, Pixel Phones
Member
For the last decade, software product economics followed a simple rule: once you built the product, the marginal cost of serving the next user was almost zero.
But as the industry pivots from conversational chatbots to autonomous, “agentic” AI, that fundamental economic law is being rewritten.
Product leaders are racing to integrate autonomous agents and orchestration frameworks into their roadmaps, drawn by their ability to plan, reason, and execute complex workflows without human intervention.
However, there is a strategic trap hidden within this shift. While AI models are rapidly commoditizing, the compute required to run autonomous agents is growing exponentially. Historically, software costs were anchored by infrastructure that scaled gradually. You added capacity as usage grew, but the marginal cost per interaction remained relatively low and predictable. In the agentic era, AI systems have shifted costs toward a hyper-incremental, pay-per-action reality, making those fixed infrastructure costs far less visible. By failing to fully account for this “compute tax,” product managers risk engineering products where usage scales faster than margins, turning growth into a liability.
Drawing on my experience in AI-native consumer hardware, this shift feels massive yet familiar. In hardware, scaling is always constrained by how efficiently the supply chain is managed. In the agentic era, compute is starting to behave the same way.
To lead in the agentic era, product managers must shift from traditional software economics to a compute supply-chain mindset. The challenge extends beyond infrastructure cost. In agentic systems, the optimization surface expands across several variables, including model selection, context management, orchestration design, and workload placement (cloud vs edge). These variables shape whether a product scales profitably. The key is to control cost by optimizing these variables and constraints, much like a classic supply chain optimization problem.
What follows is a framework for navigating agentic AI economics and building products that scale economically.
To understand the challenge, we must look at how agentic AI behavior impacts the bottom line.
The Chatbot Era (Linear Costs): In the first wave of generative AI, unit economics were relatively predictable. A single user prompt typically triggered a single primary model call, with bounded compute per interaction. While production systems could include additional steps, such as retrieving information, refining inputs, or formatting outputs, the overall cost per interaction remained relatively contained and easier to model. This made costs largely linear and allowed PMs to model variable expenses into SaaS subscriptions or hardware LTV.
The Agentic Era (The Multiplier): We are now entering the era of the agent, where orchestration frameworks allow AI to act autonomously. Instead of just answering a single question, the agent plans, searches, evaluates, and corrects. This multiplies the compute needed for each user prompt.
Note: This math is intentionally simplified to illustrate the direction of the cost curve. In my experience, strong teams reduce this cost by using techniques like routing simpler sub-tasks to smaller models, summarizing context, and caching responses. However, those levers do not eliminate the underlying economic sensitivity of agentic systems.
At first glance, the solution seems simple: pass the cost to the user. But in practice, this usually breaks down. Consumers and enterprise buyers strongly prefer predictable pricing, and are wary of software bills that scale unpredictably with usage. As a result, most AI products are forced to mask variable compute costs behind subscriptions, usage caps, or credit systems.
Agentic AI fundamentally disrupts a core assumption of Product Led Growth (PLG) by increasing cost as usage scales. In traditional software economics, high engagement is the ultimate goal; engagement and margin move in the same direction.
In the agentic era, unchecked engagement becomes a margin risk. A purely LLM-based architecture does not scale economically. This is where many teams start to feel real tension between growth and margins. Success now depends on ensuring value scales faster than cost.
The Agentic Cost Curve:
User Engagement ↑ → Agent Loops ↑ → Compute Cost ↑↑↑
As product leaders, we no longer fully control our cost structure, as it increasingly depends on external model pricing and third-party APIs. In this environment, we must intentionally shape user engagement rather than blindly scaling it.
To survive this shift, product managers need to stop treating compute as an infinite software utility and start managing it as a physical supply chain. Our scope extends well beyond designing features or upgrades. We now orchestrate a distributed system of on-device compute, cloud-based models, and external services.
In the agentic era, compute economics behave like a waterbed. Push down on one cost, and it surfaces elsewhere. For example, smaller models increase error rates, and compressed context can trigger retry loops. We cannot eliminate the underlying costs. We can only strategically reduce them by optimizing the architecture.
To operate effectively, product leaders must understand the two macro variables of the Agentic P&L:
Total Agentic Cost = Cognitive Cost + Execution Cost
This is the cost of reasoning and orchestration. Because agentic systems trigger multiple model calls per user task, this cost scales rapidly. PMs must balance this tax across three execution levers:
Note: This dynamic heavily advantages vertically integrated players, who can bypass the third-party tax by owning more of the software stack.
To see these compute supply chains in action, let’s look at two distinct architectural approaches currently emerging in the B2C market:
Real-world usage and my own experiments show that even basic, daily automated tasks rack up roughly $30 per month in API fees, while power users may exceed $200+ depending on usage intensity.
While these examples are most visible in consumer productivity tools, the same economic dynamics apply to enterprise product teams building agentic systems on managed platforms (such as AWS AgentCore or Vertex AI Agents). In these environments, the optimization challenge shifts from device vs. cloud trade-offs to shared infrastructure efficiency, including multi-tenancy, model routing, and real-time versus asynchronous execution at scale.
These architectures reveal a deeper truth: in the agentic era, product architecture is inseparable from compute costs.
At its core, building a sustainable AI product is a supply-chain optimization problem where the goal is to allocate compute (supply) efficiently and control costs without sacrificing value. In my experience, teams that manage this well tend to converge on three principles:
Start by understanding how your product actually consumes compute. Quantify your cost per task alongside your cost per user. Map each workflow to identify where expensive “flagship” reasoning is truly required versus where simpler utility tasks can be handled by cheaper, smaller models.
Categorize your model calls by their “intelligence grade,” based on the level of reasoning required and the type of model needed. This allows you to ensure each step is handled by the simplest, most efficient models and context size possible.
Apply token-saving strategies such as summarization or caching to minimize redundant compute. The goal is simple: reduce the total compute required per task without degrading user value.
Not all AI workloads belong in the same place. For consumer products, this requires explicitly deciding how compute is distributed across the device and the cloud.
Each feature should be evaluated across three dimensions: latency (does it need to be instant?), privacy (does it require local processing?), and cost (where is it most efficient to run at scale?).
In my experience building AI-native products, this leads to a tiered execution model:
This architectural split is becoming the standard. For example, Google’s newly released Gemma 4 model family spans ultra-mobile edge deployments (Effective 2B/4B) to larger server-class variants (31B Dense), reflecting a clear industry shift toward intentional, workload-tiered AI systems over monolithic ones.
While most teams focus on optimizing cognitive cost, execution cost is equally critical. If every action depends on third-party APIs, your margins remain exposed to external tolls.
Where possible, product leaders should pursue vertical integration in the execution layer by building direct integrations or native capabilities that reduce reliance on external platforms. For example, an AI agent that completes tasks within its own productivity suite captures more value than one that incurs external API costs on every action.
Owning execution is one of the most powerful levers for protecting long-term margins.
The shift from chatbots to agentic AI is a fundamental restructuring of product economics. If compute is treated as an infinite, cheap utility, the Agentic Multiplier will quietly erode margins. From my perspective, this shift is still underappreciated, and many teams are only beginning to feel its impact.
The next generation of product leaders will distinguish themselves by architecting efficient, cost-aware systems. They will treat compute as a supply chain, efficiently balancing it across device, cloud, and services.
Building the smartest AI model is no longer enough. The winners will be those who can scale intelligence economically. In the agentic era, intelligence is no longer the constraint. Economics is.
Disclaimer: The views and opinions expressed in this article are purely the author’s and do not necessarily reflect the official policy or position of Google or its affiliates.
We are proud to be a leading contributor in content for product leaders. See below for more great resources from Products That Count.
Akash Rathi is a Product Manager at Google, leading product strategy for Pixel phones at the intersection of AI and consumer hardware. He has led multi-billion dollar consumer tech products from concept through global commercialization, with his products receiving world-wide media acclaim. Akash holds an MBA from The Wharton School and a master’s degree in Industrial Engineering & Operations Research from UC Berkeley.
MemberIn this episode hosted by Sonjoy Ganguly, Emrecan Dogan will be speaking on why shipping more features doesn’t mean creating more value.
Explore the journeys of 2026’s top product leaders in our latest eBook. The Product Leader Awards spotlight leaders who are building learning organizations, making AI...
In this episode of the CPO Rising Series hosted by Jay Patel, Jonathan Shottan speaks on building the world’s largest kinetic dataset.