Tier 2 Co-pilot

Tier 2 — Co-pilot (Agentic Tier Framework v2026.1)

Concise Definition

Tier 2 Co-pilot capabilities are embedded, drafting assistants that participate in a workflow by generating or pre-filling outputs which must be reviewed and explicitly accepted by a human before they change state, transact, or are filed.

The AI is triggered automatically at defined workflow steps (not ad hoc by the user), and its outputs are treated as the default proposal that humans accept, edit, or reject.

Key Characteristics of Tier 2

1. Embedded Invocation
   • The AI runs at named workflow steps (e.g., “Generate first contract redline,” “Pre-fill intake form,” “Draft research summary”).
   • Users do not decide whether to invoke the AI; the workflow does.
2. Human Review as a Gate
   • No AI output can advance the workflow without human review.
   • Review is a mandatory gate before any state change, filing, or transaction.
3. Default Acceptance Risk
   • The AI’s draft is the default starting point.
   • Reviewers tend to accept unless they notice issues, creating a risk of shallow or inattentive review.
4. No Autonomous Commit
   • The AI does not sign, submit, file, or otherwise commit actions on its own.
   • Human acceptance is required for any external or persistent effect.

Boundary with Tier 1 (Augmentation)

• Tier 1: Human chooses when to use AI; it is a tool invoked on demand.
• Tier 2: Workflow automatically invokes AI at defined steps.

Boundary with Tier 3 (Workflow Operator)

• Tier 2: Every state-changing step is human-gated.
• Tier 3: AI autonomously executes bounded workflow steps without per-step human approval.

Canonical Controls for Tier 2

To be defensible, Tier 2 deployments implement:

1. Review-Quality Telemetry
   • Override rate (how often humans change AI output).
   • Edit distance between AI draft and final version.
   • Time spent per review.
   • Falling overrides or shrinking edit distance can indicate declining supervision.
2. Materiality-Gated Review Depth
   • High-materiality outputs (e.g., major contracts, regulatory filings) require named senior reviewers.
   • Lower-materiality outputs can be reviewed by trained delegates under defined standards.
3. Default-Acceptance Design
   • UX makes thoughtless acceptance harder and engaged review easier.
   • Acceptance requires an explicit, affirmative action (e.g., “Review & Accept” with diff view), not a passive “next” click.
4. Confidentiality Boundary
   • Inputs to the AI are constrained by the use case’s confidentiality regime.
   • The boundary and rationale are documented in the Risk Register (GOV-02).
5. Verification Regime
   • Authorities, citations, and factual claims are checked before acceptance.
   • Verification steps (who verified, what was checked, how) are documented.

Tier 2 Drift Pattern

Dominant failure mode: Silent drift toward Tier 3 behavior without Tier 3 governance.

• As confidence grows, reviewers skim or skip detailed checks.
• Override rates drop; edit distances shrink; review time falls.
• In practice, the AI begins to function like a workflow operator, but controls and governance remain at Tier 2.

A defensible program must either:

1. Restore Tier 2 Discipline
   • Reinforce training, capacity, and sampling.
   • Recalibrate review expectations and thresholds.
2. Promote to Tier 3 Explicitly
   • Update classification, controls, and documentation to Tier 3.
   • Record the promotion and new control set formally.

Allowing ungoverned drift is not an acceptable path.