What's New in Ema
New to the platform? Start with the Introduction and Getting Started. This page is written for builders and partner teams who already know Ema and want to know what changed.
October 1, 2026: evaluate orchestrators at scale, plus custom metrics and a smarter Autopilot
The headline of this release is AI Employee Evaluation: for Orchestrator AI Employees, Ema now automatically builds test datasets from a goal you describe and runs bulk evaluations against your workflows, grading every simulated conversation on both the answer it gave and the route it took. You can also build custom metrics charts for any AI Employee or across your whole Workspace, Voice AI Employees launch with multi-language support, and Autopilot gets voice dictation and memory that carries across sessions. Recruiters also gain a full multi-sequence outreach platform.
Below are the headline changes:
| Capability | What it means for you |
|---|---|
| AI Employee Evaluation | Orchestrator AI Employees now have a Training tab where you can test them against dozens of realistic requests at once, instead of chatting through a few by hand. Describe what the tests should cover and optionally add SOPs, policies, or past tickets, and Ema writes a test dataset of up to 100 test cases for you. An evaluation run then plays every test case as a simulated conversation with the published orchestrator, with connections to your other apps simulated so nothing touches your live systems, and grades each one on Response accuracy (did it give the right answer) and Plan accuracy (did it take the right steps, Tools, and connected AI Employees). Filter to the failures to see what went wrong, fix and publish, then Re-evaluate on the same dataset to confirm the fix. See Evaluating Orchestrator AI Employees for details. |
| Custom metrics | Build your own charts to track whatever you're interested in, for a single AI Employee or across your whole Workspace. An AI Employee's Metrics tab starts with a default chart set for its type, and you can add, edit, rearrange, and remove charts: pick a dataset such as Workflow or Tools, choose a chart type (KPI, line, bar, pie, heatmap, table, and more), then choose what to measure and how to group it. The Workspace Metrics page uses the same chart builder across everything your Workspace runs, including roll-ups across child Workspaces, and a view can be saved as a template. Your edits stay in a draft until you publish, and you can revert to an earlier version at any time. See Building a Custom Metrics View and Building a Workspace Metrics View for details. |
| Voice AI Employees with multi-language support | Voice AI Employees handle phone calls, running your workflows as Tools mid-conversation, and they launch with multi-language support. In the setup wizard's Basic settings, choose the languages the AI Employee answers in and mark one as the default. It answers in the default language and switches only among the languages you selected. Supported languages include English, Spanish, French, German, Portuguese (Brazil), Italian, Dutch, Japanese, Mandarin Chinese, Korean, Hindi, and more. See Building a Voice AI Employee for details. |
| Dictation in Autopilot | No need to type: select the microphone button in the Autopilot composer and speak, and Autopilot transcribes what you say into your message. See Autopilot for details. |
| Recruiter multi-sequence outreach platform | Recruiter now has a Sequences page with funnels, sequence creation, duplication, and test-send, plus mid-campaign step edits that reach in-flight candidates, shortlist visibility into which sequence/step a candidate is in, do-not-contact suppression enforcement, and draft recovery for the sequence builder. See Outreach for details. |
| Cross-session memory in Autopilot | Autopilot now stores memory across sessions, so what it learned in earlier conversations carries into new ones and you don't have to start from scratch each time you open it. See Autopilot for details. |