Skip to content

Controlled AI Development: make authority and verification explicit

Controlled AI Development means giving AI-assisted work a clear purpose, a bounded scope, and reviewable evidence. Human-in-the-loop development is useful when the human has an actual decision to make and enough context to make it.

Educational resource · PM Harness Concept & Knowledge Hub

Control is an agreement about what may change

A prompt such as “be careful” is not a useful control. Specify the accepted requirement, what changes are permitted, which assumptions need review, and what the agent must report. The review boundary should be understood before execution begins.

This does not mean a person approves every keystroke. Reversible work within a clear scope can proceed with delegated authority. Changes to a user promise, a consequential policy, or an external commitment may require a separate decision.

PM Harness connects these boundaries to intent and product decisions. This page describes a method, not an execution or approval platform provided by yourproduct.chat.

Scale the review to the consequences

ChangeExample review boundaryEvidence to request
Small, reversible presentation changeDelegate the implementation within the agreed text and layout.Diff and relevant visual checks.
Behavior change with limited scopeAgree on acceptance cases and review the plan.Tests, failure handling, and a scope summary.
Permission, payment, or sensitive-data changeRequire qualified owner review of the behavior before implementation and release.Boundary analysis, targeted checks, and a recovery plan.
Irreversible or externally visible actionObtain specific authorization for the action and target.Confirmed scope, timing, and consequences.

These categories are illustrative. The same change can have very different consequences in two products. Review should address the real exposure, not a fixed ceremony attached to every task.

Write a small execution contract

Accepted objective:
Requirement and decision versions:
Permitted scope:
Explicit exclusions:
Assumptions the agent may make:
Questions that require escalation:
Checks required before review:
Actions that require separate approval:
Evidence expected in the completion report:
Recovery owner and stop conditions:

A completion report should distinguish completed work, checks actually run, checks not run, and known limitations. “All good” is not a useful substitute for evidence a reviewer can inspect.

Approval must apply to a known version and scope. A later material change needs a new review rather than inheriting authorization from an earlier proposal.

Example: add a preview without sending real notifications

Fictional task: let an administrator preview a notification before sending it. The accepted scope includes displaying recipient information and a sample message; sending notifications is outside this change.

An agent may implement the preview and its empty or error states within the approved plan. If it discovers that the existing interface couples preview and send behavior, it should stop and raise that dependency rather than triggering a real send or adding a new delivery workflow.

The reviewer checks that opening the preview has no sending side effect, that access restrictions still apply, and that missing data produces an understandable state. A passing screenshot check alone would not establish those behavioral facts.

When a new send flow is wanted, create a separate product decision and specification. Do not let a convenient code path decide the product scope.

Verification has two different jobs

Check the implementation

Verify required behavior, important boundaries, regressions, and relevant failure paths. Inspect the change and the evidence rather than relying only on the agent's description of its own work.

Check the product outcome

After release, assess whether the intended user task improved and whether an unwanted effect appeared. Completion of the implementation checks does not establish the product outcome.

Define recovery conditions before release. If a check fails, preserve what is known, pause the affected action, and ask the responsible owner whether to revise, revert, or continue with a new plan.

Connect the review to a decision record and follow the AI coding workflow for the complete sequence.