HUMAN + AI ACCOUNTABILITY FRAMEWORK (2026)

Purpose
Use this template to deploy AI copilots/agents without losing accountability. The goal: humans own outcomes; AI accelerates execution.

1) Workflow Inventory (start with 10)
For each workflow, fill in:
- Name:
- Function (Eng/Support/SalesOps/Finance/etc.):
- Volume (per week):
- Estimated labor cost (hrs/week * fully loaded $/hr):
- Downside risk if wrong (Low/Med/High):
- DRI (single name):

2) Decision Rights (choose one tier)
Tier A — READ: AI can access data sources, no outputs sent externally.
Tier B — RECOMMEND: AI proposes actions; human approves.
Tier C — WRITE: AI creates artifacts (tickets/PRs/emails drafts); human approves before execution.
Tier D — EXECUTE: AI can take actions in systems.
Rules:
- If irreversible or regulated impact exists (refunds, legal language, production deploys, IAM changes), default to Tier B/C.
- If Tier D is allowed, define thresholds (e.g., refunds ≤ $50; emails only to internal aliases; deploy only to staging).

3) Evaluation Gate (minimum viable eval)
- Dataset: 50–200 real historical cases.
- Metric: pass rate target (e.g., ≥ 92%) + 2–3 critical failures (must be 0).
- Change control: prompts/tools are versioned; eval must pass before production.
- Human QA: sample 20 outputs/week for the first month.

4) Monitoring + Error Budget
Track monthly:
- Automated actions count
- AI incidents (incorrect customer promise, policy violation, security exposure, silent regression)
- Incidents per 1,000 actions (target set per workflow)
- Business KPI impact (cycle time, deflection, CSAT, escaped defects)
Error budget rule:
- If incidents exceed budget: rollback scope, add constraints, expand eval set, re-launch.

5) Security + Data Handling
Minimum requirements:
- SSO/SAML + role-based access
- Audit logs for tool calls and outputs
- Least-privilege permissions per workflow
- Prohibited data list (secrets, keys, PII beyond necessity)
- Retention policy (what gets logged, for how long)

6) Incident Response (AI-specific)
Define:
- Kill switch owner:
- Escalation channel (Slack/Teams):
- Customer comms owner (if external impact):
- Postmortem template includes: input, tool calls, model/version, permission scope, detection method, and prevention.

7) 30/60/90-Day Plan
30 days: ship 2 workflows in Tier B/C; create eval suites.
60 days: add monitoring dashboards + QA sampling; formalize incident response.
90 days: expand to 5–8 workflows; pilot Tier D only for low-risk actions with strict thresholds.

Use this framework as your “operating contract”: AI can move fast, but humans stay accountable.