How agencies use AI for feedback rounds on designs

The short answer

In a design feedback round, AI is good at the text and bad at the pictures. Give it every comment from the client side and it will group them per screen or asset, merge the duplicates, flag where two people want opposite things, and draft the questions you need answered before the designer starts. It should not decide what the client meant by "make it feel warmer", which stakeholder wins, or whether a round counts against the contract.

The time saved is real because consolidation is pure typing. On a job with a marketing lead, a founder and a product manager all commenting, it is often an hour per round.

How a feedback round goes wrong

Most rounds that run long fail the same way. None of it is about design.

• Feedback arrives in four places: an email, a PDF with comments, a call, a message to the designer directly.
• Two people on the client side disagree, and nobody notices until the designer has done both.
• Vague notes get guessed at. "Can the hero be stronger?" becomes a bigger headline, and the client meant a different photo.
• Nobody counts the rounds, so round three of three quietly becomes round six.
• The freelancer works from a forwarded email and misses the second half of the thread.

AI fixes the second, third and fifth of these if the comments are in one place first. It does nothing for the first and fourth; those need a process.

Step 1Collect the feedback where the design is

Ask the client to comment on the design itself, not in an email about it. A note pinned on the image says which element it means. A comment at second 14 of a video says which moment. Email feedback loses that, and it is the information the AI needs most, because it cannot look at the design and work it out.

If the client insists on email, paste each email into the same place as soon as it arrives, with who sent it and the date. Feedback scattered over a week is the hardest kind to consolidate, for people and for AI.

Step 2Consolidate with AI

Once everything is in one place, give the AI the comments and a strict job. The prompt below works with any assistant.

Below is client feedback on round 2 of a website design, from three
people. Each comment has the author and the screen or element.

1. Group the comments by screen, then by element.
2. Merge comments that ask for the same change. Keep every author.
3. List any two comments that conflict, side by side.
4. List comments too vague to act on, and write one question for each.
5. Do not suggest design solutions. Do not drop any comment.

Feedback:
[paste here]

The last line matters. Left alone, an assistant will helpfully propose solutions ("consider increasing contrast"), and those end up in front of the designer as if the client said them. You want the client's words, organised, not new ones.

Read the result against the original. Check that nothing was dropped and that merged comments really asked for the same thing. Two people saying "the button" may mean two different buttons.

Step 3Ask the client only the questions

Do not send the designer anything yet. Send the client the conflicts and the vague comments, as short questions with options where you can:

Two things to settle before we start round 3:

1. Home page header: Anna asked for the product photo, Ben for the
   team photo. Which one?
2. "The pricing page feels busy": is that the number of plans, or the
   amount of text under each one?

Everything else is clear and goes into round 3 today.

The AI can draft this from step 2. Edit it so it sounds like you, and keep it to the questions. Clients answer a two-question email the same day; they leave a full recap for the weekend.

Step 4Turn the list into tasks

With the answers in, the consolidated list becomes the designer's to-do list for the round: one task per change, grouped by screen, with the original comment attached so nobody has to interpret twice.

If a freelancer does the work, they should get the list and the pinned design together, not a forwarded email. The fewer hand-offs between the client's words and the person changing the file, the fewer guesses.

Step 5Close the round

A round ends when the new version is uploaded and the client approves it or comments on it. Upload each round as a new version rather than a new file, so the client can compare it with the last one and you keep a record of what was approved.

Count the round when it closes, against what the contract says. If you do not have a revision clause yet, the revision clause guide has wording you can adapt, and the contract templates include one. How to run the rounds themselves for image-heavy work is in managing client feedback rounds.

What not to hand to the AI

• Interpreting taste. "Warmer", "more premium" and "pop" are questions for the client, not for the model.
• Picking a winner between stakeholders. That is a client decision, and sometimes a political one.
• Judging whether a request is in scope. The AI does not know what you quoted. You do, and the free quote template is a good place to keep the lines a round is measured against.
• Anything that goes to the client unread.

How this works in OIO

OIO runs feedback rounds on the ticket. The client clicks an image to drop a numbered pin with a comment, or comments on a video at the second, optionally on a spot of the frame. Each round goes up as a new version, with a compare slider on images, and the client clicks Approve this version when it is right. A ticket waiting on the client shows a Needs feedback marker, and the client gets one reminder email a day, which you can switch off.

Client contacts can pin from a ticket link without an account, and freelancers work from their own ticket link: they read the brief and the feedback, upload the next version and mark it done.

The AI tools work on the text around it. When feedback arrives by email, paste it into the ticket's note, and Clean up tidies it without changing the facts, Extract tasks turns it into to-dos, and Translate puts a note from a client in another country into your language. Paste the consolidated list into the assistant and it drafts the questions back to the client. It does not look at the pinned image itself, so the pin, which says where, and the comment, which says what, stay the source.

For the wider picture of where AI fits on a client job, see AI project management for agencies.

Questions