AI Assistant

Human review queue

Prompt versions, source references, safety flags and accepted versions are controlled before AI output can affect a workflow.

Review queue

review

3

Drafts waiting for a human decision

Registered prompts

ready

4

Versioned prompts with source rules

Safety flags

warning

3

1 critical flags

Model runs

review

4

75% average evaluation

Review decision

0/100

Support note wording

gpt-5-review - 91% confidence - Support note NOTE-441

needs human review

Draft wording only. Worker should verify that it preserves the observed support and participant response.

Human decision
review

Decision is still required.

Source references
ready

NOTE-441 excerpt, Participant goal G-12

Critical safety
ready

No protected workflow action detected.

Safety flags

0

No safety flags on this draft

Protected actions stay blocked

  • AI output can draft explanations, but cannot submit claims, close incidents, approve invoices or alter budgets without a human workflow decision.

Model telemetry

38/100

Evaluation

75%

1 low scoring

Latency

1638 ms

Average provider response

Cost

52.8c

4,390 tokens

Guardrails

2

Failed checks

Provider run capture
ready

4 provider runs recorded against draft outputs.

Prompt evaluation
warning

1 runs need prompt review.

Guardrail evidence
blocked

2 failed guardrail checks must remain blocked.

StatusEvaluationLatencyTokensCostGuardrails
OpenAIgpt-5-reviewcompleted91%1280 ms146019.8c3/3
OpenAIgpt-5-reviewdegraded74%2420 ms119015.7c3/3
OpenAIgpt-5-reviewfailed52%1860 ms91011.1c1/3
Azure OpenAIgpt-5-mini-reviewcompleted84%990 ms8306.2c3/3

Review queue

DraftConfidenceFlagsStatus

Support note wording

Support note NOTE-441

91%Noneneeds human review

Invoice OCR extraction

Provider invoice PINV-203

58%1needs human review

Claim readiness explanation

Claim CLAIM-144

62%2needs human review

Prompt registry

PromptControlsStatus

Support note refinement

v4 - support_note

support_note_refinement

Draft only, No new facts, Cite note excerpt

Active

Invoice OCR explanation

v3 - provider_invoice

invoice_extraction

Confidence threshold, Source required, Human correction

Active

Claim readiness explainer

v2 - claim

claim_admin

No submission, Validation evidence, Reviewer approval

Active

Incident learning draft

v1 - incident

incident_learning

No closure, Manager review, Practice-standard mapping

Draft