AI in a mobile app earns its place where it removes steps: search by meaning instead of scrolling a catalogue, a form filled in from a photo of a document, an answer drawn from your own materials instead of a help section. The decision is not only which feature to add, but where the model will run — in the cloud, on your server or on the phone itself.
Key takeaways
They earn their place where the user is currently doing extra work: hunting through a long catalogue, retyping data from a document or waiting for support. A feature should replace several steps with one, not add yet another screen to the app.
| Feature | What changes for the user | What data it needs |
|---|---|---|
| An assistant inside the app | Asks in plain words and gets an answer from your materials instead of reading the help pages | A knowledge base: instructions, service terms, common questions |
| Search by meaning | Finds what is needed from a description of the task rather than an exact word match | A catalogue with descriptions, attributes and synonyms |
| Personal recommendations | Sees a selection built on their own history rather than a general list | History of purchases, views and actions in the app |
| Reading a document or photo | Photographs an invoice or a form and the fields fill themselves in | Sample documents and the rules for checking fields |
| Voice input | Dictates instead of typing: while driving, on the shop floor, out on a job | Speech recordings and a dictionary of your terms |
| Hints inside a form | Gets a hint before the mistake rather than an error after submitting | Completion rules and the history of past enquiries |
Recommendations are a large topic of their own: how they are built, what they are scored on and why they get things wrong in the first weeks is covered in our article on smart matching and recommendations. What matters here is different — where the model runs and what the app stores will ask for in return.
There are three options, and for a mobile app the third is the one most often overlooked: the model can run on the phone itself, with no call to a server.
Apple has opened such a model to developers: the Foundation Models framework gives access to the same on-device model at the core of Apple Intelligence, works offline, and its inference is free of cost; the framework ships with iOS 26 on devices where Apple Intelligence is enabled. On Android the same role is played by Gemini Nano through the ML Kit GenAI APIs: prompts run locally, data never goes to the cloud, and the ready-made operations cover summarising, rewriting, image description and speech recognition.
| Criterion | Cloud over an API | Your own model on a server | Model on the device |
|---|---|---|---|
| Works without internet | No | No | Yes |
| Where the data goes | To the model provider | Into your own perimeter | Nowhere, it stays on the phone |
| Hard tasks | The strongest models are available this way only | Open models trail the cloud ones | Short texts, parsing and rewording |
| What the price is made of | A charge per request | A server with a graphics card plus upkeep | Inference is free, but each platform is supported separately |
| Where it falls short | Network latency and rate limits | Load at peak hours | Missing on some devices and OS versions |
For most products the working setup is mixed: short summaries, rewording and speech recognition run on the device, harder requests go to the cloud, and the app knows in advance what to do when no on-device model is available. The product then stays useful without a connection, and the bill for requests does not grow in step with the audience.
Two requirements apply to any app with AI, and both are checked at review. The first is about consent to share data, the second about reporting generated content.
Apple has singled out AI services in rule 5.1.2(i): a developer must clearly disclose where personal data will be shared with third parties, including with third-party AI, and obtain explicit permission before doing so. A line in the privacy policy saying "we may use third-party services" does not meet this: a clear consent screen is needed before the first transfer.
Google Play asks for something else: apps that generate content using AI must let users report or flag offensive content without leaving the app, and the developer is expected to feed those reports into content filtering. A button next to the assistant's answer covers the requirement.
Important The consent screen and the report button are designed alongside the rest of the screens, not bolted on after a rejection at review. What else the stores check is covered in our article on publishing to the App Store and Google Play.
Send the minimum: only what the feature cannot work without. Everything else stays in your database and is put back on your own side.
A photo of a passport, a medical record or a fingerprint is not merely "sensitive data" but a separate category in law. Article 27-1 of the law on personal data, as amended by law No. ZRU-1125, requires biometric and genetic data of individuals, along with the data of users of telecommunications operators, to be held in databases on the territory of Uzbekistan. Other personal data may be processed abroad if the conditions in the same article are met.
The second limit is about decisions. Law No. ZRU-1115 of 21 January 2026 prohibits relying exclusively on the conclusions of artificial intelligence systems when taking legally significant decisions that affect a person's rights and freedoms. For an app that means a refusal, an account block or a health conclusion cannot rest on the model's answer alone.
The risks that arrive together with AI — an instruction hidden in a document, excess data in an answer, over-broad rights — are covered in our article on AI security in a company.
The model key is never kept in the app: it is pulled out of the installation file in minutes. Requests go through your own server, and it decides who is allowed what.
How the server side of a mobile app is built, and why it is designed before the screens, is covered in our article on the backend and API for a mobile app.
The first feature is the one with a measurable result and data you already hold. Anything that needs data gathered from scratch waits for the next round.
Tip Do not start with an assistant on the home screen. One feature inside a clear scenario — search, form filling, a hint — produces a result you can see in the numbers. An assistant that covers everything collects more questions than value.
The list covers the store requirements and the usual reasons for rejection. Go through it before the build is sent for review.
The price is made of three parts: working through the scenario, building the feature and paying for requests to the model each month. Building the app itself is counted separately.
Starting prices are listed on our AI and automation page. The charge for requests depends on how much text passes through the model; features moved onto the device do not add to that line at all.
Syntra Systems starts not with the choice of model but with the screen: what the person does today, how many steps they take and which of them can be removed. Then we look at the data — whether there are descriptions, a history and sample documents, without which the feature guesses instead of answering.
Our own product is built the same way. In the Syntra Health app, an AI assistant with voice and several personas explains test results, keeps a food diary and flags deviations, while marketplace recommendations are matched to a person's own readings; the interface runs in Russian, Uzbek and English. We follow the same order in client projects: scenario, data, place of processing, pilot, and only then the screen.
Let’s discuss your project
Tell us what you need, and we will estimate the timeline and cost and suggest a solution.
Yes. An AI feature is connected through your own server: the app sends a request to your API, which calls the model and returns the answer. Rewriting the app is usually unnecessary, but the screens gain new states — waiting for an answer, a refusal from the model, and working with no connection.
No. In most cases it is enough to connect a knowledge base and write answering rules: the model answers from your documents and cites the source. Fine-tuning changes the style and format of answers rather than the knowledge, needs labelled examples and has to be repeated whenever the base model is updated.
An on-device model works offline, data never leaves the phone, and requests are free — on iPhone that's Foundation Models with iOS 26 and Apple Intelligence enabled, on Android it's Gemini Nano. In return it's weaker than a cloud model, so the phone handles summaries, rephrasing and speech recognition, while harder requests go to the cloud.
Apps are rejected for breaking the rules, not for using AI. A common cause is sharing personal data with a third-party AI service without the explicit permission that rule 5.1.2(i) requires. The second usual finding is a mismatch between the data collection details in App Store Connect and what the app actually does.
If the app shows the user text generated by a model, the Google Play requirement applies to it. The report button goes next to the answer, and the reports that come in are reviewed and used to tune filters and answering rules.
An exact monthly figure can't be named in advance — the main line item is the charge for requests to the model, and it grows with the volume of text passing through it. Features moved onto the phone itself don't add to that bill. Development is costed separately: an AI assistant built on your materials starts from $3,000.
Plan for it in advance: mark the answer as generated, give a button that calls a human and keep a log of enquiries. In sensitive areas the model's answer stays a suggestion and the decision belongs to a specialist — which the law also requires in legally significant cases.
Sources
