Skip to main content
← Legal and trust

AI at Omie.

Which models Omie uses, what data reaches them, and what each role can and cannot see.

Date
9 November 2026
Whole diligence pack

A PDF is not available for this document. The full text is on this page.

In shortOmie runs on Anthropic Claude, with OpenAI for search and audio and ElevenLabs for voice. No manager, HR or Omie admin screen shows a learner’s Ada chats or reflections. Customer content does not train models.

01Which models we use

ProviderModelWhat it does
AnthropicClaude HaikuShort, frequent jobs: tagging skills, a one-line reason for a recommendation, questions inside a lesson, scoring one short answer, reading the tone of a survey comment.
AnthropicClaude SonnetWork that needs more reasoning: Ada coaching, quality checks on generated lessons, scoring role plays and capstones, scan reports and weekly digests.
OpenAItext-embedding-3-smallTurns text into vectors for search and recommendations.
OpenAIText to speechAudio lessons.
ElevenLabsVoice narrationNarrates lesson text.

We name model families here, not version numbers, because versions change more often than this page. Every provider, its region and its transfer terms are on the sub-processor list. A new provider goes through the notice the DPA promises.

02What reaches a model

FeatureWhat is sentWhat comes back
Ada, the learner coachThe question, the chat so far, and a summary of that learner: first name, role, job title, goals, challenges, learning style, skill gaps, streak, completions and lessons due.An answer for that learner.
Reflection summaryThat week of reflection prompts and answers.A short summary in that learner's own Monday email.
Role plays, capstones, short answersThe learner's answer and the scoring criteria.A score and feedback.
Scan reportsThe person's scan results and any context they gave.A written summary on their report.
Manager 1:1 prepThe learner's name, job title, learning style, target role, streak, completions and skill gaps, plus the manager's own note. Never the learner's goals, challenges or anything they told Ada.Talking points for that manager.
Manager weekly digestTeam completions, the names of the most active and quietest learners, and skill signals.A draft digest for that manager.
eNPS and pulse commentsThe comment and its score, with no name attached.Tone and themes for HR, shown only above the privacy floor.
HR people searchThe question HR typed. No member data.Search filters. Omie then finds the matches in its own database.
Lessons from your filesText from the files an author uploads.Draft lessons for that author.
Search and recommendationsSearch text and onboarding answers (role, goals).Vectors, used to match lessons.
Audio lessonsLesson scripts.Speech.

Ask Ada about my team, for managers, and Ask Ada about the org, for HR, send nothing to a model: their answers are composed from Omie’s own numbers.

Omie does not train models on customer content, and the sub-processor list records that Anthropic and OpenAI do not train on what we send through their APIs. Aggregate, anonymized usage signals improve how Omie recommends lessons.

For each call to Anthropic, Omie logs the model, the route, the token counts and the timing. The log does not hold the prompt or the reply.

03Who sees what

RoleSeesDoes not see
LearnerTheir own Ada chats, reflections and scores.Anyone else's.
ManagerProgress and skills for their own team, and what Ask Ada answers about that team.Learner Ada chats and reflections.
HR and org adminRollups across the organization. Group results are hidden below 5 people.Learner Ada chats and reflections.
Omie adminOperational screens. View as opens an account the way its owner sees it, except Ada chats and reflections, which it refuses. Each session is recorded.Learner Ada chats and reflections. A data export requested for a member: the file goes to that member, not to the admin.

How accounts and tenants are separated is on the security page.

04What a person still decides

  • No model sets hiring, promotion, pay or dismissal. Performance and potential scores in a calibration are entered by people.
  • When Ask Ada suggests a nudge, it stays a draft. Nothing is sent until the manager confirms it.
  • Lessons drafted from your files wait for the author to review and publish them.
  • A model does score role plays, capstones and short answers, and the learner sees that score and the feedback.

Model answers can be wrong. Treat Ada as a coach, not as an authority.

05How AI output is marked

Ada chat, role play and scoring responses carry an AI-Disclosure header, so software can tell they came from a model. In the AI path builder, generated items carry a visible AI label.

06Your choices

  • A learner can delete any Ada conversation.
  • A learner can export their own data, including Ada chats and reflections, from their settings.
  • A learner can add their own Anthropic key in their settings. Ada then runs on their own Anthropic account.

07Questions

Write to legal@tryomie.com and a person will answer.

Forward one link.
Every document inside.

Seven documents, versioned and dated, in one diligence pack.

  • ISO planned
  • DPA countersigned in 1 working day
  • 30 days notice on sub-processors
Contact us