2690 field guide

What needs to be in place before AI helps a brand workflow?

A practical way to decide whether the real gap is information, steps, ownership, or an AI capability.

When a founder sees that a task happens every day, it is natural to ask whether AI can do it. Sometimes it can. Often the first problem is that the team has to find the current file, check the status in another tool, and ask someone which version is approved. AI working on top of that confusion can produce a faster answer that is still wrong.

Find the decision underneath the request

Take a real example: “Can AI tell our agency which ad creative is ready to run?” Before choosing a model, ask what “ready” means. Has legal approved the claim? Is the current file attached? Is the offer still active? Who gives the final approval? If these answers live in different chats and spreadsheets, the first useful build is a shared record of the asset and its state.

Put the information where the job happens

List the information someone needs to make the decision. Give each item an owner and a place where the current version lives. Keep the source visible. For an ad, that might be the approved file, claim, landing page, start date, owner, and the evidence of approval.

Set the order of operations

Write down what happens next and which exception changes the path. A creative asset may move from brief to draft, review, revision, approval, and launch. A rejected claim should stop the handoff; a date change should update everyone relying on the plan. Those are workflow rules, not prompts.

Add AI where it has a defined role

Once the record and steps are dependable, AI can help summarize a brief, compare feedback across reviewers, find recurring exceptions, or flag a missing approval. People still own consequential decisions. Each output should let them inspect the source and correct a mistake.

Test the result with the people who will use it, on the work they actually need to do.

Start by documenting one job and one recent failure. If the missing information or ownership is obvious, fix that first. If a bounded AI task remains, test it on real examples with a person who knows what “good” looks like.