How it Works
How OLi works.
OLi watches what your people are doing on their computers, recognizes moments where it can help, and shows up with the right content or action — no prompts, no dashboards, no training.
10% more productive time in 30 days, guaranteed.
Every step runs on the activity graph
The spine
Every step runs on the activity graph— Dataken’s first-party record of real human work.
10B+ records · 5+ years of continuous operation · 1M records per day.
Capture — it starts at the desktop.
A lightweight agent on the user's machine records structured activity records — what app is active, how long they dwell, what content is on screen. On-device capture means raw data never leaves the machine unprocessed. Roughly 1M records per day in a typical deployment.
Technical detail
Each record is a structured tuple: timestamp, application, window title, dwell duration, detected content type. No keylogging. No full-screen capture. The signal is what the user is doing, not what the user is typing.
Graph — records flow into your tenant's activity graph.
Activity records stream into the graph, where they're anonymized at the architectural layer and bounded to your tenant. Your graph is yours — isolated, per-tenant, and built entirely from your own work. The graph is what makes OLi's suggestions specific rather than generic.
Technical detail
Apache Spark ingest pipeline. Per-tenant data boundaries enforced at the storage layer — not as a policy overlay, but as an architectural property. Anonymization runs inline on ingest. The rules registry that drives recognition is also per-tenant.
Recognize — OLi sees the work, not just the window.
Pattern recognition runs against the graph in real time. OLi detects friction, recognizes tasks, and identifies context — three tab-switches in 90 seconds on a compliance form, a department timesheet opening, a knowledge-base article that matches the active task. Recognition is the trigger; nothing happens until OLi sees something worth acting on.
Technical detail
Recognition rules are configured per tenant. A healthcare tenant's rules detect EHR workflows. An insurance tenant's rules detect claims processing. Same agent, different context, different actions — because the rules registry and the graph both live inside your tenant boundary.
Act — help lands at the point of work.
When OLi recognizes a moment worth acting on, it surfaces help as a desktop toast — a small notification at the point of work. Today that's usually content: a micro-learning refresh, a knowledge-base article, a break reminder. Increasingly it's skills: real actions OLi takes for the user, like building SOWs from a timesheet and routing them through Zoho Sign.
Technical detail
Skills are tenant-customized. They integrate with your stack — your EHR, your CRM, your document-signing workflow. This is what separates OLi from a generic AI assistant: the action is wired to how your organization actually works.
Learn — the loop closes back to the graph.
Every interaction feeds back into the activity graph. Recognition patterns refine against real usage — not because of a vague ML claim, but because the graph records whether the intervention actually changed behavior. Closed-loop improvement, not open-ended dashboards.
Technical detail
Privatized LLM inference by default. Any skill or Ask OLi call that invokes an LLM uses the provider's zero-retention, no-training-on-tenant-data mode. An open-source isolated-deployment option is available for security-sensitive tenants.
The full loop
One real moment, end to end.
Abstracts aside — here’s what the five-step loop looks like for a single sales moment OLi is watching right now.
Quote to SparkCo opened 4× with no reply.
Records the 4 opens of the quote PDF over 3 days.
Links the PDF opens to the original email thread and the contact.
Matches the 'warm lead, no reply after repeated opens' rule.
Drafts a nudge citing the slide the contact re-read most.
Records whether the nudge lands a reply — and refines the rule.
Privacy is architecture, not policy.
Activity-graph anonymization. Privatized LLM inference by default. Per-tenant data boundaries. On-device capture where possible. These aren’t promises — they’re how the system is built.
Read the full privacy story →