Home / Writing

AI product

How to Add AI to Your Mobile App Without Losing User Trust (or Your Budget)

By William Souef · 2026-10-13 · 9 min read

Every founder is being asked the same question: "Where's the AI?" Adding a chatbot to tick the box rarely helps. The AI features that work solve one job users already struggle with, do it reliably, and fail gracefully. Here's the process I use, based on shipping AI in my own apps.

Start from a job, not a model

Good AI use cases in mobile apps share three traits: the user already does the task (or gives up on it), the output is easy to check, and a slightly imperfect answer is still useful. Examples:

• Breaking an overwhelming task into small steps, as in Rockie.
• Turning a photo or description into structured information, like a food scan.
• Summarising or classifying what the user already wrote.
• Generating content you review before it ships, like the stories in Histia.

Weak use cases: open-ended chat with no clear job, anything where a wrong answer is dangerous, and features users never asked for.

On-device or cloud?

	On-device model	Cloud API
Cost per request	Free	Pay per use
Privacy	Data stays on the phone	Data leaves the device
Capability	Smaller model, narrow tasks	Larger models, harder tasks
Availability	Only on supported devices	Any device with a connection
Offline	Yes	No

For narrow, frequent tasks, on-device is often the better product: private, instant and free. But it isn't available on every iPhone, so you need a plan for users without it. In Rockie, a deterministic engine built from hand-written templates covers those users, and the interface never shows which engine ran. I explain that architecture in Designing an LLM feature users actually trust.

Design for reliability

• Narrow the job. "Return 4–6 steps of 2–5 minutes" beats "help the user".
• Use structured output. Ask for data, not prose, so the app can validate it.
• Add a quality gate. Check every output against rules before showing it, and cover those rules with tests.
• Plan the failure path. One repair attempt, then a fallback. Never an endless spinner.
• Keep the user in control. Let them edit, regenerate or split the result.

Control the cost

Cloud AI costs scale with usage, which is a problem in a subscription app where your best users are your most expensive ones. Ways to keep it sane: cache results that don't change, use the smallest model that passes your quality gate, move frequent narrow tasks on-device, put heavy AI behind the paywall, and set per-user limits before you need them.

Privacy and App Review

Tell users what's processed and where, update your privacy details in App Store Connect, and don't send personal data to a third-party model without saying so. If the feature touches health, money or children, go slower and be stricter.

Measure what matters

"The model answered" is not success. Track whether users act on the output: did they start the first step, save the result, come back the next day, convert after using the feature? Those numbers tell you whether the AI is a feature or a demo.