Kojo Osei

 undo All posts
Subscribe

Context Gravity

• 2 min read

I've been using the term context gravity to describe how startups can build defensibility in the AI era. It describes AI-era stickiness, just as data gravity described SaaS-era stickiness.

In the previous SaaS era, the ultimate defensibility play was to become a system of record. Stickiness came from the pain of switching. You might hate Salesforce's interface, but you couldn't leave without risking your entire sales pipeline. You could find a better project management tool, but migrating thousands of meticulously tracked issues and team conversations from Jira was a non-starter. The cost of switching was so high that you’d tolerate clunky features for years rather than attempt to rip and replace.

Two trends are weakening data gravity moats and opening the door for a new playbook:

  1. Mature integration providers make it easier to access data from systems of record (e.g. Ampersand, Merge)
  2. AI agents thrive on new types of procedural data that are not traditionally captured by systems of record. 

Data gravity is finally giving way to context gravity.

Context gravity is the depth of your domain-specific standard operating procedures (SOPs). It’s the accumulated institutional knowledge: the nuanced way your team handles edge cases, your unique stylistic preferences, and the unspoken rules that define your company's output.

This context is the new moat. It's harder to extract than raw data because the moat is in the engine that dynamically assembles the right pieces of context every time. Which SOPs are relevant? Which user preferences apply to this specific edge case? What historical corrections were made on a similar task?

Consider an AI accounting product. The product is stickiest when it has native context on a customer's operations. It knows that a customer always flags invoices from a particular vendor for manual review because they often have billing errors. It knows that a customer's CFO prefers expense reports to be summarized by project code first, then by date. It's learned that any transaction over $10k automatically triggers an alert to both the project manager and the finance lead, with a specific message template.

These are the agent interactions that will be hard for competitors to replicate.

I look for three things when I’m thinking about whether a product has high context gravity:

  1. Is it the place where work (SOPs & agent instructions) is defined? Is it where users naturally input their instructions, preferences, and operating procedures? The more precisely you can capture how work should be done, the better you can leverage the instruction following capabilities of models.
  2. Is it the place where work is executed? Where does context get dynamically assembled and orchestrated with LLMs? A defensible product must stitch together the right rules for each situation, feed them to the model, and take action on what comes back. If that orchestration happens elsewhere, you're just a fancy settings page.
  3. Is it the place where work is refined? Your product must have a built-in feedback mechanism whenever a user corrects a model's output, overrides a decision, or approves a result. This feedback helps you refine your context layer and deepen your understanding of that specific customer's operations.

I keep returning to this framework because it feels like the natural evolution from data gravity. The deepest moats in the AI era will be built on this define-execute-refine cycle.