AI product teams cannot plan around stable token economics because model providers change prices, token usage, and deprecation schedules.
2
Teams can protect their business by routing work across models, using open-weight models and CPUs where they fit, and keeping the option to switch providers.
3
The useful software factory combines model choice with data, orchestration, security controls, and workflows where agents can inspect tasks, involve teammates, and open pull requests.
Summary
Sarah Sachs explains how Notion manages the cost and dependency risks of building AI products. Model upgrades can keep the same token price while using far more output tokens, or cost substantially more while the older model is being deprecated. Sachs argues that companies should assume model suppliers are competitors and avoid tying their product to one lab. Notion's answer is a model-agnostic system that routes about 75% of traffic through its auto model, evaluates complete task trajectories, and chooses models by cost, capability, and latency. She also recommends open-weight models for moderate tasks, CPUs for deterministic work, and governance that makes data use and persistence visible. The final section connects these choices to software factories. Secure agent systems need controls around private data, untrusted content, external communication, sandboxes, and multi-agent state. A Notion demo shows agents scoping a task, asking teammates for input, consulting other agents, and opening a pull request.
AI products need a durable system of record before they can become systems
Sachs describes Notion's progression from AI as a thought partner, to an assistant, to teammates that execute repetitive processes. The harder stage is AI as a system, where workflows interface with each other and entire systems run. She says 88% of people cannot get past the assistant stage because their data is siloed and they lack a durable system of record for collaboration. In Notion's view, that record needs to support collaboration between humans and agents, humans and humans, and agents and agents.
Model pricing creates business risk that most companies cannot absorb
Sachs calls cost a structural barrier to building and serving AI products at scale. She gives two recurring examples. A reasoning model can keep the same price per token after an upgrade while consuming three times as many output tokens. Another model can cost 40% more while its predecessor is scheduled for deprecation within four months. She asks whether customers are growing 40% or making three times more revenue during that period. If a company automatically upgrades, someone pays the difference, either customers or investors.
Sachs says AI companies should treat their model supplier as a competitor. A lab that sells a first-party product can charge a markup on the tokens it sells to another product, leaving little defensible value for the buyer. Tying the product to one provider also removes the ability to exit. She argues that companies without the negotiating power of large enterprises need to preserve their own leverage rather than hope for a durable discount from one vendor.
Product knowledge matters more than trying to win token economics
Sachs says most applied AI companies should stop trying to compete with labs on model training. They can build an advantage through data flywheels, knowledge of customer needs, and orchestration. A product needs to know when it requires high capability, low price, or lower latency. Not every request belongs on the slowest and most capable model. The product experience and routing logic have to justify the cost of the models used underneath.
Traffic should be routed by the whole task, not by token price alone
Sachs recommends thinking about cost per capability per second. She says sending every request to the latest Opus model is a mistake. Large-scale data analysis may justify it in Notion, while inbox triage on Opus would rip off both the customer and Notion. Notion evaluates complete web-search trajectories rather than only the latency or price of one call. That broader evaluation can reveal which provider is best for the actual workflow.
Sachs warns that the best frontier model and provider can change every month. A large discount from one lab may therefore leave a product behind for much of the time. She calls the ability to walk away the most valuable form of leverage, even when maintaining model interoperability requires engineering work. Notion's auto model routes about 75% of its traffic and can switch among available models. Customers can also access models without being locked to one vendor.
Open weights and CPUs cover work that does not need frontier models
Sachs says open-weight models are strong enough for many moderate tasks and can provide a credible alternative during negotiations. She expects capability gaps to close over time, so teams should prepare for that option early and test models against their own tool errors and latency needs. She also says many jobs should run on CPUs. Converting a CSV to a PDF, handling Notion tool calls through a CLI, and running deterministic SQL queries do not require an LLM.
Secure software factories need controls around data and agent actions
Sachs describes the lethal trifecta as access to private data, exposure to untrusted content, and the ability to communicate externally. Autonomous systems increase the risk because fewer actions receive supervision. She recommends sandboxes, computers, and multi-agent orchestration that makes it clear what agents see, what they do, and what persists. Her Notion example shows an agent scoping a task in an active document, asking teammates for input, consulting another agent, and opening a pull request.
"If you have access to private data, exposure to untrusted content, and the ability to communicate externally, the second you have that system, you're exposing risk."18:18
Who should watch
You are building an AI product whose margins depend on third-party model prices and need a way to avoid provider lock-in.
Your team sends most requests to one frontier model and needs a practical routing strategy for moderate or deterministic work.
You are building agent workflows that can access private data, read external content, or take actions outside the system.
