How Do You Freeze Approved Decisions in an AI Project Canvas?

How Do You Freeze Approved Decisions in an AI Project Canvas?

Reviewed by Product Specialist at thinQit. Updated 17 November 2026.

You freeze an approved decision in an AI project canvas by marking it as locked, recording who approved it and when, and sending every later change through a written change request instead of an edit. Once a decision is frozen, AI agents and people build against it as fixed input, and the canvas shows any attempt to alter it.

What does it mean to freeze a decision?

A frozen decision is an entry in the AI project canvas that has been approved by a named person and can no longer be edited in place. It has a status, an owner, an approval date and the exact wording that was approved. The boundary is simple: a freeze protects the decision itself, not the topic around it, so open questions on the same subject stay editable.

The reason to freeze is practical. When an AI agent reads a canvas, it treats every entry as an instruction. If an approved entry can be rewritten silently, two agents can end up building against two versions of the same choice. Freezing removes that ambiguity. For the wider setup, see Setting Up an AI Project Canvas Before the Build Starts.

A product lead and an engineer signing off a printed requirements document in a meeting room

Which decisions should be frozen?

Freeze decisions that other work depends on and that are expensive to reverse. Typical examples are the page list of a website, the data entities of a web app, the navigation structure, the brand and tone rules, and the integrations the build may use. Leave exploratory notes, draft copy and unresolved questions unfrozen.

A useful test is to ask what breaks if the entry changes tomorrow. If the answer is "several screens, a data model or a launch checklist", freeze it. If the answer is "nothing yet", keep it as a draft. Over-freezing is also a risk: a canvas where everything is locked forces a change request for trivial edits and people stop reading the requests.

How do you freeze a decision step by step?

Freezing is a short procedure. Run it every time a decision is approved, not in a batch at the end of a phase.

  1. Write the decision as one unambiguous statement, for example "The site has five pages: Home, Services, Pricing, About, Contact."
  2. Record the reason in one sentence, so a later reader knows why it was chosen.
  3. Name the approver, the person with authority to accept it, and note the date.
  4. Set the entry status to approved and lock it, so it is read only for agents and for people without change rights.
  5. Link the entry to the tasks, screens or documents that depend on it.
  6. Announce the freeze in the project log, so everyone working from a copy knows to refresh it.

Step five matters most for later change handling. A dependency link tells you what a future change would touch before anyone approves it.

A developer comparing a change request ticket with an approved specification on two monitors

How do changes work after a freeze?

A frozen decision changes only through a change request that names the entry, the proposed new wording, the reason and the affected dependencies. The approver reviews it, and if accepted, the old entry is marked superseded and a new frozen entry takes its place. The old version stays visible in the history.

This keeps the audit trail intact. Nobody overwrites the original, so you can always answer the question "what did we agree, and when did it change?" The same pattern is described for documentation in How Do You Prevent Conflicting Instructions in AI Documentation?, and the record of reasons is covered in What Belongs in an AI Project Decision Log.

How should AI agents treat a frozen decision?

AI agents should treat a frozen decision as a constraint they may not alter and should flag when a task conflicts with it. If a request contradicts a frozen entry, the correct behaviour is to stop, cite the entry and ask for a change request, not to quietly choose one side. The related behaviour is explained in How Do AI Agents Handle Conflicting Instructions?.

Agents can still propose changes. The difference is that a proposal goes to a human approver, while the frozen entry stays in force until that approval happens.

A small team at a whiteboard sorting sticky notes into approved and open columns

What goes wrong when decisions are not frozen?

Three failures repeat. First, scope drift: an approved page list grows by one page in a chat message and nobody updates the canvas. Second, rework: a rebuild discards approved work because the new run read a different version. Third, disputes: when a stakeholder says "we never approved that", there is no record to settle it. The article Rebuild an AI Website Without Losing Approved Decisions shows how a frozen set survives a rebuild.

Frequently asked questions

Can a frozen decision ever be edited directly?

No. Direct edits defeat the purpose of the freeze. A change goes through a request, the approver accepts it, and the old entry is marked superseded while a new frozen entry replaces it. The history keeps both versions.

Who should be allowed to approve and unfreeze a decision?

The person who owns the outcome, such as the product owner or the client sponsor, should approve. Assign this by role before the project starts, so an agent or a developer is never the one who approves a change to their own instructions.

How many decisions should be frozen at once?

Only as many as other work depends on. Freeze structural choices such as pages, data entities and integrations, and leave drafts and open questions editable. A canvas with everything locked slows the team down without adding protection.