Case study
App Store rejection over AI consent: what our app had to change
Our own iPhone app was turned down over two App Store rules on personal data. The consent question was too general: it named no AI provider, covered one feature out of three, and came later than the reviewer looked.
Key takeaways
- A consent question that says only that an AI will help does not tell people which company receives their words.
- Chat, translation and document reading each send something different, so the screen should list them one by one.
- Ask before the first message goes, and ask again when the wording changes.
- A no should leave the app usable, with a person answering instead of the assistant.
- Test the change of mind: someone who says no, writes, then turns the AI on.
If your app sends what customers type to an AI service, the App Store reviewer reads that part closely, and a general "an AI will help you" question is not enough. We learned it on our own iPhone app: the first version was turned down in mid-October 2026 over two App Store rules about personal data. The app has an AI assistant that answers our clients, translates messages and reads a company's trade licence at sign-up.
What a rejected AI app means for the business
Until Apple approves a version, nobody can download the app, and every feature in it waits with it, including the parts the reviewer did not question. The same rules come back with each later update, so an AI feature explained in words the reviewer does not accept can hold back any release after it. In our case the app had already done what owners are usually told to do, and it still was not enough: it asked people whether they agreed to an AI helping them, and the reviewer said no to that.
Why a general "an AI will help you" question fails review
It failed on three counts, all of them about being too general. The question named no provider: it said an AI would help, not which company receives the words. It spoke only of the assistant's replies, while two other features sent text to the AI as well, namely translating messages and reading a trade licence at sign-up. And it came before the first message, while the reviewer opened the assistant's screen first, so the asking was not where the reviewer looked.
How we traced the rejection to its cause
We read the two rules against the question our app asked, as a reviewer would, and found those three gaps. Then we listed every place where a client's words can reach an AI, including the ones a client never sees, and made sure each one asks before it sends anything.
Answering Apple raised a question of our own: are we comfortable naming the provider on that screen? For personal data we settled on a provider whose written data-protection terms are in force and which does not train its models on what we send, in every country we serve.
A last walk through the app the way a reviewer does, trying each answer in turn, came later, on 26 October 2026, and found one more path. A person who first chose only a person, then changed their mind and turned the AI on, was answered by nobody: the conversation was still waiting for our team, and the assistant stayed silent because it had been told to.
What we changed
The question is now a screen titled "Before you start: AI and your data". It names the AI provider for the account, taken from our platform so that it always matches what really happens, and it lists what each feature sends:
- The assistant's replies: what the person writes, up to 10 earlier messages from their company's chat, their name and role, their company's name and country, and the account details they ask about.
- Translation: the text alone.
- The trade licence: its text.
Under that list it says what is never sent: passwords, sign-in codes, card details, files, photos or voice. Nothing goes until the person taps "Agree and continue".
"Only a person, no AI" keeps the app usable: our team writes every reply, nothing of theirs goes to the AI, and their licence is checked by a person instead. A no stands until the person changes it in Settings. The question now comes at four moments: the first look at the assistant, the first message, a licence about to be read and a translation. People who agreed to the old wording are asked again, because a yes counts only for the words it was given to.
The same screen went into our Android app and our web portal, and the privacy policy was rewritten in all five languages to name every AI provider and list what each feature receives. We closed the last gap too: when someone who said no, or never answered, turns the AI on, every conversation of theirs still waiting for our team goes back to the assistant, and what they wrote meanwhile is answered at once. A conversation the team has already picked up stays with the team, and the team sees in the conversation why it went back.
Where it stands now
Within a day of the rejection, the question, the record of who agreed to which wording, the Android app, the web portal and the privacy policy were changed, and the corrected iPhone version was back with Apple on 14 October 2026. Our reply to the review walks through each point of the two rules and shows where the app answers it.
In our web portal the new question is live, and so is the second pass that brings the assistant back into a waiting conversation, closed on 26 October 2026. The iPhone version is with Apple's review, and the decision is awaited.
What you can check yourself
- Open your app as a new user and see whether it names the AI company before your first message goes.
- List every feature that sends text, photos or documents to an AI, not only the chat, and check that the question covers each one.
- Read the AI part of your privacy policy: does it name the provider and say what each feature sends?
- Say no to the AI in your own app and check that the feature still works, with a person answering instead.
- Say no, write a message, then turn the AI on: does someone answer you, and how quickly?
What to do if the review has already turned you down
If your app is already stuck in review, the useful move is not to send the same version again: send it back with a reply that answers each rule the reviewer cited, point by point, and shows where in the app it is answered. Take the rejection message as a list. If it names two rules, write two short paragraphs, one for each, and under each say which screen now does what the rule asks. Point to the exact places a reviewer will open: the consent screen, the first message, the privacy policy. Ask whoever looks after your app to show you that reply before it goes, and to explain any rule you do not follow in plain words. Our own reply did that for the two rules our reviewer named, and the corrected version went back with it. Keep the reply factual and short; the reviewer needs to find the answer, not read an argument. If a rule still does not seem to fit your app, ask for it to be explained before you promise a change.
Questions people ask
Does Apple require me to name which AI company receives my customers' data?
Our reviewer was not satisfied with a question that said only that an AI would help. The rule in question covers what a privacy policy must say and asking people before their data goes to another company, and naming the provider is how the app shows who receives it. If you are unsure how the rule applies to your app, read the review notes with whoever looks after it.
My app has an AI chat and a translation feature. Does one consent screen cover both?
One screen can cover both, as long as it lists what each feature sends instead of speaking in general terms. Our first question mentioned only the assistant's replies, while translation and reading a trade licence also sent text to the AI. The version we sent back lists all three separately.
What happens to people who agreed to my old, vaguer wording?
We asked them again, because a yes counts only for the words it was given to. Someone who agreed to "an AI will help" has not agreed to a named provider receiving a described list of what is sent. Keep a record of which wording each person agreed to, so a change of wording can be handled cleanly.
If a customer says no to the AI, can they still use the app?
Ours can: a person on our team writes every reply, nothing of theirs goes to the AI, and their licence is checked by a person instead. A no stays until the person changes it in Settings. A no that leaves the app unusable is not much of a choice.
How long does a corrected version take to go back through App Store review?
The changes are the part you control; the review time is not, and ours is still awaited. Our changes were made within a day and the corrected version was back with Apple on 14 October 2026. Until the reviewer accepts the new wording, plan for the app not being available.