Guardrails for AI in delivery: the propose, approve, record pattern

There are two kinds of AI guardrail, and only one of them works.

The first kind is a promise. It lives in a policy doc or an onboarding slide: always review the agent's changes before accepting them. It sounds responsible. It survives right up until the sprint is behind, the change looks fine at a glance, and someone clicks accept without really reading. A guardrail that depends on human discipline under pressure is not a guardrail. It is a good intention with a paper trail.

The second kind is a workflow. The oversight is not something you are asked to do, it is something the system does to every change, whether anyone remembers to or not. That is the pattern worth building, and it has three moves.

Propose

The agent never applies a change directly. It prepares one. Point it at an initiative and ask it to pull the next two items into the cycle, and it does not move them. It drafts the move and hands it back:

> Pull the next two Checkout items into this cycle.
# drafting proposal · nothing applied yet
Proposed: move 2 items (I6825, I6746) into the current cycle,
update affected dates. Waiting on your go-ahead.

The important word is proposed. Nothing has changed on the board. The agent has done the legwork, the reasoning, and the mechanical preparation, and then it has stopped at the exact point where judgment begins.

Approve

A person makes the call. Not a rubber stamp, an actual decision, because the proposal gives them what they need to make one: what will change, what it affects, and what it costs downstream. Approve it and the change is applied through the real tools. Decline it and nothing happened.

Approval is the step most "AI in your PM tool" pitches quietly skip. They demo the agent doing something impressive and imply a human was watching. Watching is not approving. The gate only counts if the change genuinely cannot proceed without it.

Record

Every action, the agent's and the human's, lands on one shared record with a name on it. The proposal is attributed to the agent that made it. The approval is attributed to the person who gave it. The applied change shows both, in order. Six weeks later, when someone asks why a date moved, the answer is not a shrug. It is a record.

Recording is what turns oversight into accountability. A change you approved but cannot trace is only half-governed. The record closes the loop.

Why the pattern has to be structural

You could try to run propose, approve, record by convention: agree as a team that no one applies an AI change without review, and hope it holds. It will hold on a calm week and break on a hard one, which is precisely when you needed it.

Building it into the workflow removes the discipline requirement. The agent cannot apply a change because the system does not let it. The human cannot skip the record because the system writes it. The guardrail does not rely on anyone being careful, which means it is still there on the day everyone is rushing.

The same logic extends to code. In the DevStride delivery loop, work an agent produces lands in staging and reaches production only when a person signs off. The gate is a state the work has to pass through, not a reminder someone has to honor.

The honest edge

Two things worth saying plainly, because a careful reader will test both.

First, propose, approve, record governs how changes are applied, not whether the underlying system is a sandbox. When you approve a write, it is real and immediate. There is no undo waiting behind it. Read first, approve deliberately.

Second, the pattern is only as strong as its enforcement. A "propose then approve" flow that a user can toggle off is back to being a promise. The value is entirely in the change not being able to happen any other way.

Get that right and you have the thing every AI-in-delivery vendor gestures at and few actually build: speed you can trust, because the trust is in the workflow and not in the pitch.

One of four principles in the AI-era delivery playbook. To see the gate running on a real plan, get a guided walkthrough.