Separate the goal from the AI model.
First agree on the result, the limits and the checks it must pass. Then choose a model and tools that fit the work. Changing a model should not change the task or the approved permissions.
This gives you a steady question at each step: “What information, tool or decision does this task need next?”
Keep a shared task record.
Keep a record of the goal, plan, approved sources, drafts, open questions and approvals outside any single chat. A person joining the task later should be able to see what is agreed and what remains open.
The record should explain what changed, why it changed and which information informed each decision. Keep the original evidence as well as any AI summary.
WORKFLOW CONTROLS
Records / Permissions / Tool choice / Review
Shared records. Clear limits. Results people can check.
Set clear permission limits.
Write down the allowed tools, approved inputs, data boundaries, and actions that require approval. A planner should not be able to expand those permissions merely by describing a new plan.
For example, a research task might allow an AI tool to organize approved files. Sending an email, opening other systems or publishing the report would need separate permission.
Choose tools for the next step.
Choose the tool or model that fits the next step. Consider the task, the information available, the cost limit and the checks the result needs to pass.
Choosing a tool does not prove that its result is correct. Check the result and keep a record of how the tool was chosen.
Decide when the work is ready.
Separate “a draft exists,” “checks have run,” and “an authorized person approved it.” These are different states, not interchangeable labels for completion.
A fictional report could be drafted and checked while still waiting for permission to publish. Keep that waiting state visible.