AI in Mobile Apps 2026: What Works, What Is Hype

I build mobile products with LLMs and computer vision in production, not demos. The real decisions: API or on-device, serverless or a standing backend.

Everyone Wants an "AI App" Now

In 2026 the app stores are full of products that wired a language model into a single screen and called themselves "AI-powered". Most are gone within a quarter, not because AI doesn't work, but because it didn't solve anything in them.

I build mobile apps where LLMs and computer vision are the core of the product, not a garnish. Below are the decisions I make in production, along with their consequences.

API or On-Device

A model behind an API (OpenAI, Anthropic, Google) gives you the best quality and zero model-maintenance work. You pay per token, so cost scales linearly with usage: great at the start, dangerous at scale. No network, no product.

On-device (Core ML, Apple Intelligence, Gemini Nano) flips the equation: zero cost per request, full privacy, works on a plane. But the models are smaller, and the differences between devices become your problem.

Good products mix both: fast, cheap operations locally, heavy reasoning in the cloud. That split has to be designed before the architecture exists. Bolting it on later hurts.

Serverless: a Lesson I Paid For

An AI backend on serverless (Modal in my case; Lambda or Cloud Run in yours) scales to zero: dev and staging cost nothing while nobody uses them.

But serverless that scales to zero has a cold start. The container boots, the model loads, and the user stares at a spinner for a dozen-plus seconds on the first request after a quiet period.

Guess who always lands on that "first request after a quiet period"? The Apple reviewer. App Review tests your app after hours of backend inactivity, hits the cold start, sees a hanging screen, and rejects the release. I went through this on one of the projects I've worked on.

The fix was simple: production keeps one warm container at all times; dev and staging still scale to zero. The cost of one small container versus a rejection cycle in App Review: that math defends itself. You just have to know the decision exists at all.

Computer Vision Is a Product, Not a Feature

I've worked on consumer products where image recognition is the heart of the app. The biggest lesson: in CV there is no "works / doesn't work". There's a quality curve, and a decision about what you do with uncertainty.

Users will forgive the model a mistake if the UX is ready for it: show a few candidates instead of one "confident" answer and make correction effortless. CV products that pretend to be infallible lose trust at the first slip.

What Is Hype

• "Let's add AI because investors keep asking": a feature without a problem to solve is token cost with no return.
• Chat as the universal interface: sometimes the best UX for AI is a button, not a conversation.
• "The model does everything, the app is just a client": offline, sync, costs and edge cases are still product engineering.

Where to Start

Start with one question: what does the user get in exchange for waiting and for my token bill? If the answer is concrete, the rest is a set of known, solvable engineering problems. You've just read a few of them.

I build these products end-to-end: architecture, model integration, store release. Planning an AI-powered mobile product? Get in touch.