The Context Layer
The axis AI agents are missing.
Today’s agents read what was written. Dataken sees what’s happening right now — the first-party record of what real people are doing on their computers. That’s activity context.
10% more productive time in 30 days, guaranteed.
Structured records · on-device capture · privatized by default
The missing axis
Document context alone isn’t enough.
Most context engineering today is RAG — retrieval over documents, code, and tools. That gives your agents deep knowledge of what was written down. It gives them nothing about what the user is doing right now.
Activity context is the missing second axis: the substrate a just-in-time agent needs to generate the right prompt, surface the right skill, or act without being asked. Dataken is the only company operating that layer in production.
What’s in the graph
Structured records of real work — not keylogs, not screenshots.
Every record is a structured tuple. Timestamp, application, detected context, dwell duration, recognized task. Raw signal is processed on the user’s device; only the structured record enters the graph. These four are representative, not real user data.
On-device capture · per-tenant boundaries · anonymized on ingest
The matrix
The only column that spans every row.
Every AI category below is useful — but none of them see the work itself. Dataken is the only layer that captures activity context, then acts on it.
| Capability | Generic chatbots | RAG copilots | Meeting AI | Dataken |
|---|---|---|---|---|
| Reads documents you point it at | ||||
| Summarizes meetings after the fact | ||||
| Sees what you're working on right now | ||||
| Acts without being asked | ||||
| Tenant-specific rules and skills |
Reads documents you point it at
Summarizes meetings after the fact
Sees what you're working on right now
Acts without being asked
Tenant-specific rules and skills
What it unlocks
Three capabilities no RAG agent can touch.
Agents that act on your stack, not generic.
A generic copilot can’t write into your customer’s Epic instance. OLi can — because the rules registry is per-tenant, and the activity graph makes the action specific instead of generic.
Help before the question, not after.
OLidoesn’t wait to be prompted. It sees the work, matches a recognition pattern, and acts. The human doesn’t need to know what to ask for — OLi already knows what they need.
A production dataset nobody else has.
10 billion+ activity records. 5+ years of continuous operation. 1 million records per day. Not a feature — a dataset that took years to build and can’t be bootstrapped by a new entrant.
Dataken is the only company operating the activity context layer in production. 10B records. 5+ years. Every tenant’s graph is their own. If AI agents are going to graduate from chat to action, this is the substrate they need.
10% more productive time in 30 days. Or you don’t pay.
About the context layer
Frequently asked
- What is a context layer for AI agents?
- A context layer is a persistent, structured record of what is actually happening in an organization's work, which an agent can read from instead of being told. Without one, an agent only knows what a user types into a prompt. With one, it can act on the work in flight — which is the difference between an assistant that answers and an agent that anticipates.
- Why can't we just build this with better prompts or RAG?
- Prompt engineering and retrieval both operate on what already exists in a document store. Neither observes work as it happens. The gap they cannot close is ambient context — what a person is doing right now, what they did before it, and what people in the same role usually do next. That has to be captured at the desktop, continuously, over years.
- Can we access the activity graph programmatically for our own agents?
- That is the developer track: the context layer is exposed so teams building in-house agents can use the activity graph as an input substrate rather than only consuming OLi. The developer program is currently at waitlist stage.