EmaFusion™
EmaFusion™ is Ema's model layer. Every time an AI Employee (AIE) needs a large language model — to classify intent, extract a field, write a reply, validate a rule — the call goes through EmaFusion™ rather than to a single hard-coded model. EmaFusion™ looks at what the step needs, picks the best model for it from a broad catalog, and falls back to another model if the first one's answer isn't confident enough. You get the strengths of many models without wiring up or maintaining any of them yourself.
This means you never pin an AI Employee to "GPT" or "Claude" and hope it's the right choice for every task. You set a high-level intent — let EmaFusion™ pick, restrict it to a shortlist of models, or optimize for speed, cost, or accuracy — and EmaFusion™ handles per-call model selection underneath.
Availability. Intelligent EmaFusion™ routing is enabled per tenant. When it isn't enabled, AI Employees still run on the models from your configured providers, but model selection uses your provider's primary/fallback order rather than EmaFusion™'s per-task routing. See How EmaFusion™ works for both paths.
Why it matters
- No model lock-in. Your AI Employees aren't tied to one vendor or one model version. As new models are added to the catalog, your AI Employees can use them without any rebuild.
- Right model per task. A cheap, fast model handles a simple classification; a stronger model handles complex reasoning. EmaFusion™ makes that choice per call, so you don't overpay for simple steps or underperform on hard ones.
- Confidence-aware fallback. If the chosen model returns a low-confidence answer, EmaFusion™ cascades to the next-best candidate instead of returning a weak result.
- One place to configure. Model behavior is set on the AI Employee and inherited by every agent in it, with per-agent overrides when a specific step needs something different.
Where you configure it
EmaFusion™ has two configuration surfaces:
- The AI Employee — model selection and optimization priority for the whole AI Employee, set on the AI Employee Configuration card under the AI Employee's Configuration tab. This is the default every agent inherits. See EmaFusion™ setup.
- An individual agent — an optional Override default configuration toggle on each agent in the workflow editor, for when one step needs a different model than the AI Employee default.
Providers and API keys — the credentials EmaFusion™ uses to reach the models — are managed separately by an administrator under Admin → Workspace → LLM Providers. See EmaFusion™ setup.
In this section
| Page | What it covers |
|---|---|
| How EmaFusion™ works | Candidate selection, capability filtering, optimization priority, confidence-based cascading, and per-call token tracking. |
| EmaFusion™ setup | Setting model selection and optimization on an AI Employee, overriding per agent, using your own providers (BYOM), and configuring providers as an administrator. |
What's next
- How EmaFusion™ works — the routing pipeline, end to end.
- EmaFusion™ setup — configure it on an AI Employee and at the provider level.