Configuring what gets extracted

Capture gives you a transcript. What turns a transcript into something usable is the extraction configuration, and most of it comes down to signals.

Signals

A signal is something worth noticing: a pricing enquiry, a competitor named, a commitment made, a risk raised. Auron only finds what you have defined.

A definition has a name, a description, a category, example phrases and labels. Example phrases carry most of the weight. Two or three phrases in the words your customers actually use will outperform a long abstract description of the concept.

Definitions are written in the console under Signals, and can be imported in bulk from a CSV if you are bringing an existing taxonomy across.

Categories

Categories group definitions so dashboards and reports can roll them up. Design them the way you would want the report to read, because that is what they become.

Attaching signals to agents

This is the step people miss. A definition does nothing until it is attached to an agent, on that agent's Signals tab. An agent with none runs a perfectly good conversation and detects nothing.

If a session comes back with an empty Signals tab, check this before anything else.

Reading what came out

Each detection carries a confidence percentage, the insight in a sentence, and the quote it was drawn from. Read the quote first: it is the evidence, and it is what tells you whether the interpretation is fair or whether the definition needs tightening.

Detections appear inside the session, on the organization-wide signals screen, in Analytics, and on the phone.

The constitution

The AI constitution holds the behavioural rules that apply to every agent, in three parts: how agents conduct conversations, how signals are extracted, and how they are reviewed. Each rule can be edited, switched off, or restored to its default.

Change it for anything that should be true of every agent, such as a standard of evidence a signal must meet. Anything specific to one agent belongs in that agent's canvas.

Enrichment

The signals screen can trigger an enrichment pass over what has been detected, producing the composite views that the dashboards and the knowledge graph read from.

Tuning over time

Treat the first few weeks as calibration. Read a sample of detections against their quotes, drop definitions that never fire, and rewrite the ones firing on the wrong thing. A catalog of twelve well-tuned signals is worth more than sixty that nobody trusts.