How It Works

EfficientAI evaluates voice agents the way your customers experience them: by placing real calls against the agent and driving those calls with simulated callers.

Everything on the platform is built from six objects. Each one is explained in full on its own page in this section — this page shows how they fit together.

The object model

An evaluation suite is the join table: it pairs one agent with every combination of the personas and scenarios you attach. Each combination becomes one or more runs, and each run is scored by the metrics you enabled.

Combination math (nCr + pairing)

Evaluation suite runs are based on persona-scenario pairing:

If personas and scenarios are already selected in the suite, total combinations are P x S (Cartesian product).
If subsets are selected from larger pools first, use nCr for each selection stage:
- Personas selected: nCk
- Scenarios selected: mCr
- Then suite combinations expand to (nCk) x (mCr) before applying run count.

The evaluation lifecycle

Because the persona, scenario, and voice mapping are all pinned configuration, re-running a suite after a prompt or model change produces a comparison you can trust rather than a one-off anecdote.

Where each concept is explained

Concept	What it is	Explained in
Agent	The system under test, plus its call medium and prompts.	Agent
Persona	The simulated caller: behavior profile plus a pinned voice.	Persona
Scenario	The goal, context, and success criteria for a conversation.	Scenario
Metrics	The scoring rules applied to a finished run.	Evaluator
Evaluator	The engine that runs combinations and applies metrics.	Evaluator
Evaluation suite	The grouping that turns personas and scenarios into run combinations.	Evaluation Suite
STT / LLM / TTS	The three model layers inside a voice bundle.	Setup
Telephony vs WebRTC	The two call mediums an agent can be tested over.	Agent

Community & contact

Found a bug or have a feature request? Open a GitHub issue.
Join our Discord for faster replies!