Usage

The Usage block (usage.svc) is the telemetry substrate under Flow Intelligence. It captures a PII-free event stream from your applications and answers a question most analytics tools cannot:

For any business process, how much wall-clock time is spent by a human, how much by AI, and how much is pure waiting — and is that ratio improving or degrading?

That framing matters if you are deploying AI into real processes. “The model is fast” and “the process got faster” are different claims, and only the second is worth anything to the business paying for it.

What it does

Capture. An always-on, PII-free base event stream — name, time, actor class, correlation id, tenancy — declared in your application’s telemetry.yaml. Events carry no payload, which is what makes always-on capture safe to leave enabled.

Decompose time. Every elapsed interval splits into human, AI and idle, with thinking versus working and reasoning versus processing broken out.

Span engine. Business users define a span — “starts at event A, ends at event B” — after the fact. The block pairs start and end events per case and materialises the full duration vector retrospectively, over events already captured.

Contrast. Trend over time, plus the headline comparison: AI-assisted versus manual cohorts. Did the loop save time?

Why after-the-fact spans matter

Conventional instrumentation makes you decide what to measure before you measure it. You add a timer around the thing you suspect is slow, deploy, wait, and discover you bracketed the wrong interval.

Here, capture is broad and cheap, and spans are defined later against events already recorded. A question that occurs to someone in a review — “how long between a customer submitting and an agent first responding?” — gets answered from existing data, with history, rather than starting a two-week instrumentation cycle.

Interval union, not addition

The subtle correctness property. When a human and an AI are working on the same case at the same time, naive accounting counts that period twice and produces total time exceeding wall-clock time.

The engine resolves overlap by interval union: concurrent human and AI work is counted once. Time decomposition sums to real elapsed time, which is what makes the human/AI/idle ratio a number you can put in front of a business owner without caveats.

Next

• Core Concepts — actor classes, interval union, after-the-fact spans
• Operations — declaring telemetry, recompute, storage and monitoring