How EmaFusion Works
EmaFusion is Ema's proprietary mixture-of-experts model. It routes across 40+ large language models from 9+ providers -- including OpenAI GPT and O-series, Anthropic Claude, Google Gemini, Meta Llama, Mistral, Moonshot Kimi, Alibaba Qwen, and enterprise-trained custom models -- to deliver optimal results for every sub-task an AI Employee performs.
Architecture
When a task is created, the EmaFusion model recognizes the sub-task type and selects the single best model for that task's framework from its broad catalog of providers. Unlike AI tools locked to one model, EmaFusion dynamically picks the right model for each sub-task -- and, if that model's response falls below a confidence threshold, it cascades to the next-best candidate. Cross-model validation across multiple models only runs in select high-stakes scenarios; routing to a single best model is the default behavior.
Routing Logic
- Task analysis -- EmaFusion examines the incoming sub-task (e.g., conversation generation, structured query, reasoning, summarization).
- Model selection -- Based on benchmarked performance, cost, and latency data, EmaFusion ranks candidate models and picks the single best one for the sub-task.
- Execution with cascading fallback -- The selected model processes the task. If its response falls below the confidence threshold for the chosen optimization mode, EmaFusion cascades to the next-best model in the ranked list.
- Continuous improvement -- Ema regularly benchmarks and updates model performance data, so routing decisions improve over time.
Supported Model Families
EmaFusion routes across a broad set of model families. The following are the primary families available in the current release:
| Provider | Models | Typical Strengths |
|---|---|---|
| OpenAI (GPT) | GPT-5, GPT-5.2, GPT-5.4, GPT-5.4-mini, GPT-4o, GPT-4o-mini, GPT-4.1, GPT-4.1-mini, and earlier variants | General-purpose generation, structured output, tool calling |
| OpenAI (O-series) | O1, O1-mini, O3, O3-mini, O4-mini | Advanced reasoning, step-by-step problem solving, math and code |
| Anthropic | Claude Opus, Claude Sonnet, Claude Haiku (multiple 3.x and 4.x variants) | Reasoning, long-context tasks, safety-sensitive content |
| Gemini Pro, Gemini Flash (1.5, 2.x, 2.5, 3 variants including Lite) | Structured queries, multimodal tasks, translation | |
| Meta | Llama 3.1 | Cost-effective generation, open-weight flexibility |
| Mistral | Mistral Large | Multilingual tasks, code generation |
| Moonshot | Kimi K2 | Long-context understanding, agentic tasks |
| Alibaba | Qwen | Multilingual generation, code |
| DeepSeek | DeepSeek V3, DeepSeek R1 | Reasoning, code generation, cost efficiency |
| Custom (BYOM) | Your private models | Domain-specific, compliance-restricted tasks |
GPT-5.4 and GPT-5.4-mini are the latest additions to the OpenAI family available through EmaFusion. These models support temperature control and are used as the default for many agent types including search reranking, entity extraction, and SDR email generation.
Benefits
| Benefit | Description |
|---|---|
| Maximize accuracy, minimize cost | Uses the right model for each sub-task -- avoiding the cost of always defaulting to a single premium model for every request. |
| Continuous improvement | Ema benchmarks and updates model performance regularly. New models are integrated as they become available. |
| Future-proof | New models are added seamlessly with no vendor lock-in. Your AI Employees automatically benefit from advances in the LLM ecosystem. |
| Fewer hallucinations | Outputs can be cross-checked across multiple models, reducing the likelihood of hallucinated content. |
| Privacy by design | Ema automatically obfuscates sensitive data (names, emails, phone numbers) before sending to models. |
Bring Your Own Model (BYOM)
To personalize output further, you can use BYOM to integrate your custom-trained models for specialized tasks. This is useful for:
- Domain-specific models trained on proprietary data.
- Models hosted in your own infrastructure for compliance reasons.
- Private instances of commercial models (e.g., Azure OpenAI, private Gemini endpoints).
BYOM supports any model endpoint that exposes an OpenAI-compatible chat completions API. You provide the API URL and API key, and EmaFusion routes all requests for that AI Employee to your endpoint.
BYOM limits to be aware of:
- One custom model per AI Employee. You cannot configure multiple BYOM endpoints on the same AI Employee -- the one you set becomes the sole model for all of its agents.
- Not supported at the agent level. BYOM is configured at the AI Employee level only. Individual agents cannot override it with a different custom model.
- Capability mismatches can cause failures. If your custom model does not support a capability an agent requires (for example,
FUNCTION_CALLING,STRUCTURED_OUTPUT, orVISION), that agent will fail at runtime. Confirm your endpoint supports the features your workflow needs.
See Setup & Configuration for instructions on configuring BYOM.
Example Use Cases
- Customer Support AI Employee: Uses GPT-5.4 for conversation, Claude for reasoning, and Gemini for structured queries.
- Finance Analyst AI Employee: Routes to an enterprise-trained proprietary model for compliance-sensitive tasks, with fallback to GPT-4o for summaries.
- Healthcare AI Employee: Uses a private BYOM model with a HIPAA-compliant API for patient data.
- Voice AI Employee: Routes speech-to-text diarization to GPT-5.4 for high-accuracy transcription, with downstream agents using EmaFusion's balanced routing for response generation.
Configuration Hierarchy
EmaFusion configuration is resolved through a two-level hierarchy: a default set at the AI Employee level, and optional per-agent overrides that replace the default for a specific node in a workflow.
- AI Employee default. Every AI Employee has a default EmaFusion configuration -- the selected models, optimization mode, and any BYOM settings -- stored on the AI Employee itself. This is the configuration all agents use unless they override it.
- Agent-level inheritance. By default, every agent references the AI Employee's EmaFusion configuration and inherits any changes to it. This is why updating the AI Employee-level default automatically flows through to all agents that have not been overridden.
- Agent-level override. Each agent has an Override default configuration toggle in its Advanced Settings. Enabling it unbinds the agent from the AI Employee default and lets you set an agent-specific model, optimization mode, or custom model. Disabling it re-binds the agent to the AI Employee default.
- BYOM is AI Employee-scoped only. Custom models configured via BYOM are set on the AI Employee and apply to every agent underneath it. There is no agent-level BYOM override.
As a rule of thumb: configure the AI Employee-level default to match what most of your agents should use, and reach for agent-level overrides only when a specific agent has different requirements (e.g., a reasoning model for a planner, or a faster model for a latency-sensitive step).