AI in your app: what you can add and what it costs

Almost every app request we get today comes with the same line: “and we want AI.” Fair enough. The problem is it doesn't say what. AI can be a three-line summary, a search that understands questions, a camera that reads receipts or an assistant that talks to customers. Each of these has a different cost, timeline and risk.

The biggest price difference isn't in the feature. It's in where the AI runs: on the user's own phone or on a cloud server. In this guide I split the two paths using the official prices published by Apple and Google, walk through a monthly cost estimate and explain the App Store rule that is already rejecting apps with poorly explained AI.

What you can add to an app with AI

Before talking price, it helps to know what's feasible today without lab research. These are the features that show up most in business apps:

• Summaries: of chats, documents, customer history, daily reports.
• Natural language search: the user types “March order with late delivery” and the app finds it.
• Photo and document reading: invoices, prescriptions, payment slips, IDs, menus, labels.
• Ready-made text: suggested replies, product descriptions, reminders in the right tone.
• In-app assistant: answers questions, books appointments, checks orders, always tied to your business data.
• Personalized suggestions: workouts, itineraries, products, next steps, based on what the user has already done.

None of this requires training your own model. It requires knowing where each feature should run.

On-device AI: no per-use cost, with limits

The first path is AI that runs on the device. Both platforms already offer it.

On iPhone, the Foundation Models framework gives apps access to the roughly 3-billion-parameter model behind Apple Intelligence. According to Apple, inference is “free of cost,” meaning no per-call charge, it works offline and keeps data on the device. It arrived with iOS 26.

On Android, the ML Kit GenAI APIs use Gemini Nano on the phone itself to summarize, proofread, rewrite, describe images and respond to prompts. Google says there is no “additional server cost” per call and the feature works the same without internet.

Sounds perfect. There are three limits that belong in your scope:

1. Not every device can run it. Apple requires a device that supports Apple Intelligence, with the feature turned on. On Android, it's a list of selected devices, mostly Pixel, Galaxy and other flagships.
2. The model is smaller. It handles short, well-defined tasks well. It doesn't replace a large model for long reasoning.
3. Android has quotas and usage rules. The system limits calls per app and only lets the AI run while the app is open on screen.

In practice, on-device AI is great for features that improve the experience, but it can't be the only way into an essential feature. If your customer uses an entry-level phone, they need another path.

Cloud AI: paid per token, and prices change

The second path is calling a model in the cloud. Here you pay per token, the chunk of text that goes in and out of the model. The app owner pays the bill, every month, as long as there's usage.

Google's official prices, checked in October 2026, show the gap between models. In the Gemini API pricing table, Gemini 2.5 Flash-Lite costs $0.10 per million input tokens and $0.40 per million output tokens. Gemini 3.8 Flash costs $0.75 for input and $3.75 for output until December 31, 2026. Starting January 1, 2027, it goes to $1.50 and $7.50. The price doubles in two months.

This is the catch almost nobody puts in the spreadsheet: cloud AI cost isn't fixed. It depends on the vendor's price table, and the table changes.

A word of caution about the free tier. On the same page, Google states that content sent on the free tier may be used to improve its products. For an app with customer data, that rules out the free tier in production.

Monthly cost: a sample calculation

A quick calculation shows the order of magnitude. Picture an app with 10,000 active users, each using the AI 20 times a month. That's 200,000 requests. Say each request sends 1,000 tokens and gets 500 back. Over the month, that's 200 million input tokens and 100 million output tokens.

• Gemini 2.5 Flash-Lite: $20 for input plus $40 for output. About $60 a month.
• Gemini 3.8 Flash, 2026 price: $150 plus $375. About $525 a month.
• Gemini 3.8 Flash, 2027 price: about $1,050 a month.
• Same feature running on the phone: zero inference cost, for users with a supported device.

The same feature can cost anywhere from zero to over a thousand dollars a month. Architecture decides, and it gets defined at the start of the project.

This cost adds to the rest of running the app: servers, app stores, monitoring and updates. And the cost of building the app itself depends on other factors, covered in how much it costs to build an app in 2026.

The Apple rule that rejects poorly explained AI

There's one point that catches a lot of companies at publishing time. On November 13, 2025, Apple updated rule 5.1.2(i) of the App Store guidelines. It now says apps must clearly disclose when personal data goes to third parties, “including with third-party AI,” and get explicit permission before sending it.

According to TechCrunch, it was the first time Apple mentioned third-party AI in this rule. A line buried in the terms of use won't cut it. The app needs a screen that says what is sent, to whom, and asks for consent.

If you're hiring, this becomes three scope items:

1. Map which features send personal data to the cloud.
2. Design the permission screen and the path for users who say “no.”
3. Prefer on-device AI for features that handle sensitive data, such as health, finance and documents.

If these items aren't in the proposal, someone will find out at store review, with the app finished and the launch date set.

How we do it at the Studio

We start with the question “what task does the user want to get done faster?”, not “where do we put AI?” Then we sort each feature into one of three boxes: runs on the phone, runs in the cloud or runs on both, with the cloud as plan B for unsupported devices.

That way, you get a fixed price after discovery and a deadline written into the contract. The code, the files and the app store and AI provider accounts are in your company's name. Our in-house dev team does the work, and we use Claude Code and Figma for design and build.

That's how we launched AI Vou Eu, an AI-powered tourism SaaS, in 4 weeks. And every system we deliver ships with an AI agent that runs daily, flags issues, suggests improvements and checks that the data makes sense. It's the same thinking we use to add AI to your company's system without creating risk: a clear function, protected data and measured cost.

What this means for you

Before asking for a quote on an AI app, bring three answers:

• Which AI features you want and which one is essential for the app to work.
• How many users you expect in the first year and how often each will use the feature.
• What kind of data the AI will receive: names, photos, documents, health, money.

With that, you can compare proposals on what matters: where each feature runs, what it costs per month at your volume and how the app gets through Apple and Google review. If you want an AI app live, with predictable cost and within store rules, see how we work on app development.