Executive summary

This internal policy Q&A assistant scores well on use-case fit, data handling, and security basics, but accountability for AI outputs is still shared across teams. Hard gate HG-09 caps the result at Production with Guards until a named owner is assigned. Treat this as an advisory checklist for the next engineering and governance conversation — not certification.

Hard gates applied

• HG-09 → ceiling Production with Guards

  No clear accountable owner for production AI outputs

Dimension scores

Dimension notes

• D7

  Ownership is the binding constraint: without a single accountable owner for production answers, escalation paths and decision rights stay ambiguous even when tooling looks mature.

• D3

  Evaluation coverage is good enough for a guarded rollout, but groundedness and regression gates should be tightened before any claim of full production readiness.

• D5

  Reliability patterns exist, yet fallback and rollback drills should be rehearsed so a model or prompt regression does not strand employees mid-policy lookup.

Risks

• [high] Unclear ownership of AI answers

  When ownership is shared, incorrect or outdated policy answers may lack a clear fixer, reviewer, and escalation path.

• [medium] Eval drift under content change

  Policy corpora change often; without stronger golden-set gates, groundedness can slip quietly after prompt or retrieval updates.

• [medium] Weak failover for model outages

  Employees may be left without a documented fallback when the model or retrieval path fails during peak policy questions.

Remediation plan

1. 1. Name a single accountable owner (S)

   Assign one role accountable for production AI outputs, with a named backup and an escalation path for disputed answers.

2. 2. Add human review for high-impact answers (M)

   Route answers that change entitlements, leave, or safety-critical policies through a human review queue before they are treated as authoritative.

3. 3. Harden golden-set regression gates (M)

   Expand fixtures for groundedness and schema validity; block prompt deploys that regress the suite.

4. 4. Document and drill fallbacks (S)

   Publish a fallback path (search index or human helpdesk) and rehearse prompt/model rollback.

Citations

• CORP-GEN-01 — Advisory assessments are not certifications
• CORP-HUM-01 — Name an accountable owner for AI outputs
• CORP-HUM-02 — Human review for high-impact actions
• CORP-EVAL-01 — Golden sets and regression gates
• CORP-EVAL-02 — Measure groundedness and schema validity
• CORP-OPS-01 — Plan fallbacks when the model fails
• CORP-OPS-02 — Make prompt and model rollback easy