AI App Development: Add AI to Your Existing App

The ones that save your users time on something they already do in the app. Summarising long content, search that understands plain English, filling in forms from a photo, and suggested replies all fit that description.

A lot of people are tired of AI features they never asked for, so a chat window added because competitors have one rarely gets used. We'd start with the task in your app that takes users longest or where they drop off, and ask whether a model makes it faster. Put the feature where that task happens, make it easy to ignore, and measure whether people use it. If a feature needs a cloud model and runs constantly, check the token cost per user against what that user pays you. The App Gameplan helps you decide what to build before you commit.

Apple can reject it if the app sends personal data to an AI provider without asking first. Guideline 5.1.2(i) says you must clearly disclose where personal data will be shared with third parties, including third-party AI, and get explicit permission before doing so.

In practice, an app that sends user content to OpenAI, Anthropic or Google needs a consent screen that names the provider and the data, and no request can go out before the user agrees. That includes the first time someone opens the AI feature. Getting the order of events wrong in code is enough for a rejection even when the consent screen exists. On-device models such as Apple's own Foundation Models don't send data to a third party. Google Play has a separate AI-Generated Content policy for apps that generate content with AI, including chatbots, image generators and apps that create voice or video of real people. Google says it isn't intended, for now, to cover productivity apps that use AI to improve an existing feature.

Yes, but it takes more work than in a native app. Neither Apple nor Google publishes an official Flutter plugin for on-device generative AI, and Firebase AI Logic documents its hybrid on-device-with-cloud-fallback mode for Android, iOS and the web, with no Flutter guide.

So there are two routes: a community package, or platform channels written directly against Apple's Foundation Models framework on iOS and ML Kit's GenAI APIs on Android. We write the channels, which gives you control over the fallback and over API changes. Budget for them from the start. Calling cloud models from OpenAI, Anthropic or Google from Flutter is simpler, since it's an ordinary network request, usually through your own server. We build a lot in Flutter, and we maintain flutter_markdown_plus, a package for displaying Markdown, the format models answer in.

It will if it runs on the phone. Features built on Apple's Foundation Models framework or Gemini Nano run on the device, so they keep working with no signal. Features that call OpenAI, Claude or Gemini in the cloud need a connection.

If offline use matters, we design the feature around what a small on-device model can do: summarising, extracting details, classifying and rewriting short text. Anything that needs a larger model waits until the phone is back online, and the app tells the user that's what is happening. Remember that on-device models only run on recent phones, so users on older handsets get the cloud version or no AI feature. We decide that fallback with you before building.

OpenAI doesn't train on data sent to its API unless you opt in, Anthropic doesn't train on API data by default, and Google doesn't use prompts or responses from its paid Gemini API to improve its products. None of that makes your app GDPR compliant by itself.

You still need a lawful basis for sending personal data, a data processing agreement with the provider, and a retention position: OpenAI keeps API data in abuse-monitoring logs for up to 30 days unless it approves you for reduced retention, its Responses API stores conversation state for 30 days by default, and it offers European data residency to approved customers. Apps serving users in the UK or EEA must use Google's paid Gemini API, where Google doesn't use prompts or responses to improve its products but does log them for a limited period. Apple separately requires explicit user permission before personal data goes to a third-party AI. We're not lawyers, so have your data protection officer sign off anything regulated. For health or financial data, an on-device model keeps it off third-party servers altogether.

Yes. ChatGPT itself is a consumer product, so apps use the model behind it through the OpenAI API, and the same applies to Anthropic's Claude API and Google's Gemini API. We connect your iOS, Android or Flutter app to whichever fits the feature, usually through your own server, so API keys never ship inside the app.

Apple's guideline 5.1.2(i) requires you to tell users their personal data is going to a third-party AI and to get explicit permission before sending it, so the integration includes a consent step. Because you pay per token, we estimate the monthly cost first, and we plan for the day the provider retires the model you built on. On iPhone, Apple's new LanguageModel protocol lets the same code call Claude or Gemini, which makes switching between providers easier.

By moving the frequent requests onto the phone. Cloud models from OpenAI, Anthropic and Google charge per token, so one user action that triggers several model calls can cost far more than it looks, and that cost grows with every user while your subscription price stays flat.

We look for requests that run often and are simple, such as summarising, classifying or rewriting short text, and run those on Gemini Nano or Apple's on-device model, which have no per-token charge. The cloud model handles only what needs it. On client projects that made this move, we've seen AI running costs fall by 30 to 50%, though your figure depends on how much a small on-device model can do for your feature. Once live, we monitor token spend so you see a rise before the invoice does. On iPhone, Apple's Private Cloud Compute model is also free to developers in the App Store Small Business Program with fewer than two million first-time downloads and Apple's entitlement.

Yes, in most cases. We add AI features to iOS, Android and Flutter apps that are already live, working in your existing codebase. The feature ships in a normal app update, and you keep your app, your users' data and your store listings.

What we need to decide first is where each feature runs. A cloud model from OpenAI, Anthropic or Google handles harder reasoning and works on any phone, but you pay per token and need a connection. An on-device model, Apple's Foundation Models framework or Gemini Nano, costs nothing per request and works offline, but only on recent phones. A rebuild only comes into it when the app itself is in poor shape, and in that case we'd tell you before quoting. The App Gameplan answers this for a fixed £3,500, credited against development.

By never letting the app act on a model's raw output. Facts that have to be right, such as prices, dates, account details or anything medical, should come from your own data, with the model only finding or phrasing them.

For misuse, the main risk is prompt injection: a user or a document tells the model to ignore its instructions. Where a model decides what appears on screen, it sends a structured description, and the app builds the screen from a fixed registry of components you've approved. Anything outside the registry is ignored, and the model can't inject code because it never produces code. We also keep the model's permissions narrow, so a successful injection can't reach anything sensitive, and we test the feature with the kinds of prompts users actually try.

Language models answer in Markdown, and flutter_markdown_plus is a Flutter package that turns Markdown into a readable screen. Foresight Mobile publishes and maintains flutter_markdown_plus, which gets more than 550,000 downloads a month and scores 160 out of 160 pub points.

Model answers contain tables, code blocks, lists and sometimes equations, and a plain text widget shows them as raw symbols. Our package renders them properly, and its LaTeX companion handles equations. If your AI feature is in a Flutter app and the output displays wrongly, we can fix it in the package itself as well as in your app.

Some did. The Digital Omnibus on AI, Regulation (EU) 2026/1744, came into force on 27 July 2026. It moved high-risk obligations under Annex III to 2 December 2027, and Annex I systems embedded in regulated products to 2 August 2028.

The Article 50 transparency duties kept their date and have applied since 2 August 2026. The Omnibus added one transition: providers of generative AI systems that were already on the market before 2 August 2026 have until 2 December 2026 to meet the Article 50(2) duty to mark AI-generated content. The same date brings two new prohibitions: AI systems that generate sexually explicit or intimate images, video or audio of an identifiable person without their consent, and AI systems that generate child sexual abuse material.

For an existing app, that means checking now that users are told when they're dealing with an AI, and that generated content will be marked by 2 December. We can review your app against both.

It depends on four things: whether the feature must work offline, how often it runs, how sensitive the data is, and how hard the reasoning is. Offline or sensitive usually means the phone. Long, multi-step reasoning means the cloud.

On-device models such as Gemini Nano and Apple's Foundation Models model are small. They're good at summarising, classifying, extracting and rewriting, cost nothing per request and keep data on the handset. Apple's on-device model has a 4,096-token context window per session, so long documents won't fit. Cloud models from OpenAI, Anthropic and Google handle harder reasoning and broader knowledge, and bill per token. Most apps use both, with frequent simple tasks on the phone and the rest sent to the cloud, plus a cloud fallback for phones that can't run the on-device model.

After a four-week App Gameplan, adding AI features to an existing app typically takes 10 to 14 weeks. Moving existing cloud AI work onto the phone takes 12 to 16 weeks, and a new app built around AI takes 4 to 7 months.

The discovery phase matters more for AI than for most app work. Whether a feature runs on the phone or in the cloud, what the fallback is for phones that can't run the model, what it will cost to run, and which consent rules apply all shape the build, and changing them halfway through is expensive. Timelines move with the number of AI features, the state of your existing codebase, and how much testing on real devices the on-device work needs.

Adding AI features to an existing app usually costs £20,000 to £40,000. A new app built around AI, with on-device models, cloud fallback and compliance work, usually costs £40,000 to £80,000.

The build cost depends on how many AI features you want, how the work splits between the phone and the cloud, and how many of your systems the app connects to. Running costs sit on top: cloud models bill per token, so the monthly bill grows with usage, while on-device models have no per-token charge. Apple's Private Cloud Compute model is free for developers in the App Store Small Business Program with fewer than two million first-time downloads and Apple's entitlement. The App Gameplan gives you a fixed-cost development quote after four weeks, for £3,500 credited against development.

Yes, if the app is available in the EU or its AI output is used there. Article 2 covers providers placing AI systems on the EU market wherever they're established, and providers and deployers outside the EU whose AI output is used in the Union. A UK company's app in the EU App Store or Play Store is in scope.

What applies depends on what the AI does. Prohibited practices, such as manipulative techniques and emotion recognition in the workplace, have been banned since February 2025. The Article 50 transparency duties have applied since 2 August 2026: people must be told when they're interacting with an AI system, and providers of systems that generate text, images, audio or video must mark that output as AI-generated in a machine-readable way. Most apps that add AI features sit at this level. The high-risk rules cover areas such as employment, education and law enforcement, and usually don't apply to them.

Recent flagships. On iPhone, Apple's on-device model needs a device that supports Apple Intelligence: iPhone 15 Pro, iPhone 15 Pro Max, iPhone Air, and the iPhone 16 range or later, set to a supported language. On Android, ML Kit's GenAI APIs running Gemini Nano support phones including the Google Pixel 9, 10 and 11, the Samsung Galaxy S25 and S26, and the OnePlus 13 to 15, plus models from several other manufacturers.

That means many of your users, often most of them, can't run it yet. Before building, we check your analytics for the share of users on supported devices and design a fallback for everyone else: a cloud model, a simpler feature, or a clear message. On Android the model lives in AICore, an operating system service, and on iPhone it ships with iOS, so neither adds to your app's download size.