# Human Control Plane

Canonical URL: https://kevinastuhuaman.github.io/human-in-the-loop-patterns/
Repository: https://github.com/kevinastuhuaman/human-in-the-loop-patterns
Author: Kevin Astuhuaman, AI Product Manager and Berkeley Haas MBA

## What this proves

Human Control Plane is an original, interactive product study about selecting the right human-oversight pattern for consequential AI actions. It demonstrates AI product judgment, enterprise interaction design, policy modeling, approval invalidation, recovery design, and public-safe technical implementation.

It is distinct from a generic approval modal. Users can change consequence, reversibility, model uncertainty, and intent freshness. A deterministic policy engine recommends one of four patterns: auto-execute with Undo, preview and confirm, preview with fresh approval, or accountable approval.

## Scenarios

- Job application: Submit application. A browser agent prepared a selected role using profile context and ATS-specific form rules.
- Customer refund: Issue a $28 refund. The order is eligible under policy and the agent has complete transaction evidence.
- CRM enrichment: Update account ownership. The agent inferred a territory change from two public sources and an internal routing rule.
- Access change: Grant production access. A role change created a request for privileged access to a sensitive environment.

## Control patterns

- Auto-execute with Undo: Proceed immediately because harm is bounded and the action can be cleanly reversed. Owner: Operator can reverse.
- Preview and confirm: Expose the proposed change and uncertainty before a lightweight confirmation. Owner: Requesting user.
- Preview + fresh approval: Pause before the consequential action and bind approval to the exact committed state. Owner: Accountable user.
- Accountable approval: The agent may recommend, but a named owner must approve and the system records the decision. Owner: Policy owner.

## Product decisions

- Match authority to consequence: A model's capability does not grant it permission. The product assigns authority based on potential harm.
- Bind approval to state: Approval expires when inputs, evidence, navigation, or the proposed action changes.
- Design recovery with execution: Undo, reject, escalate, and fail-closed states belong in the primary workflow.
- Approval is bound to one run and one committed input state. Changing a policy input invalidates prior approval.
- Reversible, low-consequence actions can move faster when a real Undo path preserves the prior state.
- High-consequence, hard-to-reverse actions remain recommendation-only until an accountable owner approves.
- The audit record captures inputs, evidence, policy, owner, and outcome instead of exposing hidden chain-of-thought.

## IP boundary

All examples and records are synthetic. The repository contains no Trackly production source, employer IP, customer data, credentials, private prompts, or private infrastructure topology.

## Related evidence

- Portfolio: https://portfolio.kevinastuhuaman.com
- Agent Workflow Canvas: https://kevinastuhuaman.github.io/agent-workflow-canvas/
- Evals Control Room: https://kevinastuhuaman.github.io/evals-control-room/
- AI Product Builder Stack: https://kevinastuhuaman.github.io/ai-product-builder-stack/