Can AI usage be measured without reading employee prompts?

Yes. A telemetry system can measure numeric usage characteristics without collecting prompt text. The Upsilon draft core uses input, output, cache-write, and cache-read token counts and does not require semantic payloads to calculate its core metrics.

Does privacy-preserving telemetry prove productivity?

No. Token telemetry describes operator-layer behavior. Productivity, task quality, and business outcomes are external measures that can be joined separately when appropriate.

Data minimization

Privacy-Preserving AI Telemetry

Measure the operating pattern without requiring the semantic content of the work.

Core telemetry

Input-token count
Output-token count
Cache-write / cache-creation count
Cache-read count

Not required for core measurement

Prompt text
Response text
Source code
Repository contents

Content independence is a boundary, not a magic guarantee

Privacy still depends on implementation details such as identity handling, retention, access controls, transport, enrichment, and whether external outcome data is joined later. The important architectural point is that the core operator measurements do not require semantic inspection in the first place.

Use the right layer for the right question

Content-independent telemetry can answer questions about operating patterns. If an organization needs correctness, code quality, accepted-change data, incidents, or business outcomes, those signals should be added as explicit external layers rather than smuggled into the operator metric.

AI operator metrics →Upsilon Standard →

Loading rankings…