How our AI works
Atriverse uses AI models to draft briefs, run agent steps, check facts and answer questions. Every AI result is labelled AI-generated, and you can open Why this? to see the reasons or sources behind it.
Human oversight
- Agents can only use tools on an approved list. Anything else is blocked and recorded.
- Any action that reaches the outside world — sending a message, writing data, spending money — waits for a person to approve it.
- Each agent has daily limits on actions and cost, and is paused automatically if it hits them.
- Administrators can switch any model or agent off instantly, or stop all agents at once.
- Models are monitored daily for drift, accuracy, safety flags and cost.
Giving feedback
Under any AI answer, use the thumbs up or down next to Useful?. Your feedback is stored with that answer and used to measure each model's accuracy. To report a problem, email security@atriverse.com.
Models we use
Agent action gate
high riskAllow-list, approval and cap checks on every agent tool call.
- For
- Deciding whether an agent may run a tool now.
- Not for
- —
- Known limits
- Only as good as the allow-list and caps.
A person reviews its results before they take effect.
Agent step executor
high riskRuns one workflow agent step and writes its output (may call tools).
- For
- Research and drafting inside an approved workflow.
- Not for
- Acting on external systems without approval; final decisions without review.
- Known limits
- Can hallucinate facts; outputs are checked by the Signal B grader.
A person reviews its results before they take effect.
Jarvis chat
high riskConversational operator assistant with tools over the org graph and vault.
- For
- Answering questions about the company and navigating the app.
- Not for
- Taking irreversible actions without confirmation.
- Known limits
- Can be wrong about live state; sources shown when tools are used.
Results are shown with an AI-generated label.
Research step executor
high riskResearch-heavy agent steps routed to search-grounded models.
- For
- Finding and citing sources for factual steps.
- Not for
- Unsourced claims; private data lookup.
- Known limits
- Search results may be stale or wrong.
A person reviews its results before they take effect.
Run risk classifier
high riskRules that assign a run its risk tier and autonomy level.
- For
- Gating autonomy and MCP access per run.
- Not for
- Replacing human judgement on novel risks.
- Known limits
- Keyword/policy rules; can miss novel risks.
Results are shown with an AI-generated label.
Signal B grader
high riskGrades step outputs and verifies claims against sources.
- For
- Pass/fail gate on agent outputs before humans see them.
- Not for
- Sole judge of truth for high-stakes decisions.
- Known limits
- Grader can agree with confident but wrong outputs.
Results are shown with an AI-generated label.
Deck imagery
low riskGenerates abstract cover/closing images for exports.
- For
- Decorative abstract artwork.
- Not for
- People, logos, text, factual imagery.
- Known limits
- Images are decorative only.
Results are shown with an AI-generated label.
Deck narrative writer
low riskWrites the slide narrative for run exports (PDF/PPTX).
- For
- Summarising a finished run into slides.
- Not for
- New claims not in the run.
- Known limits
- Falls back to a deterministic outline if JSON is malformed.
Results are shown with an AI-generated label.
General routed calls
low riskAny other routed model call (summaries, misc helpers).
- For
- Internal helper text.
- Not for
- User-facing decisions.
- Known limits
- General LLM limits.
Results are shown with an AI-generated label.
Memory embeddings
low riskEmbeds briefs, claims and vault notes for semantic recall.
- For
- Similarity search over the operator's own memory.
- Not for
- Identity or sensitive inference.
- Known limits
- Semantic matches can be loose.
Results are shown with an AI-generated label.
Voice synthesis
low riskReads Jarvis replies aloud.
- For
- Spoken replies.
- Not for
- Impersonating real people.
- Known limits
- Voice is synthetic.
Results are shown with an AI-generated label.
Voice transcription
low riskTranscribes spoken requests to Jarvis.
- For
- Dictating to Jarvis.
- Not for
- Recording third parties without consent.
- Known limits
- Accents/noise reduce accuracy.
Results are shown with an AI-generated label.
Brief clarifier
medium riskAsks clarifying questions and drafts the locked brief from a task request.
- For
- Turning an operator request into a clear, scoped brief.
- Not for
- Legal, medical or financial advice; deciding scope without operator confirmation.
- Known limits
- May miss implicit constraints; can over-ask on simple tasks.
Results are shown with an AI-generated label.
Cascade / replanner
medium riskDecides whether to retry, amend or replan after failures.
- For
- Recovering runs from tool and model failures.
- Not for
- Changing the approved goal.
- Known limits
- May loop on unrecoverable failures (bounded by attempt caps).
Results are shown with an AI-generated label.
Doctor decision brief
medium riskWrites a diagnosis and decision brief for unhealthy org tasks.
- For
- Suggesting fixes for operator review.
- Not for
- Applying fixes automatically.
- Known limits
- Based only on recorded KPIs and events.
A person reviews its results before they take effect.
Workflow designer
medium riskDesigns the multi-agent workflow (agents, tools, order) for a brief; also drafts personas.
- For
- Proposing a workflow the operator approves before it runs.
- Not for
- Running without approval; choosing tools outside the allow-list.
- Known limits
- Can propose more agents than needed; tool choice is advisory only.
A person reviews its results before they take effect.