Building AI Features Into Mobile Apps

By May 2023, the question I hear most from mobile teams is: "How do we add AI to our app?" It's usually asked with some combination of excitement and desperation. The assumption is that AI is a feature you just bolt on, like notifications or push notifications. It's not that simple.

I've spent the last few months experimenting with AI integration in mobile contexts. Some of it works beautifully. Most of it is hype. Here's what I've learned.

What Actually Works on Mobile

Let me start with the wins. Real use cases where AI on mobile creates value:

Camera-based recognition. Your app sees what the camera is looking at, identifies objects, and acts on that. In an ERP context: a warehouse worker points their phone at a barcode-less item, the app recognizes the product by appearance and returns SKU and pricing. That's powerful. Computer vision works reasonably well on-device now with models like Vision Transformers. It's fast enough for real-time preview.

Voice input. Typing on mobile is slow. Voice is faster. But it's only useful if the system understands what you're saying well enough to act on it. In an ERP context: "Receive 50 units of product SKU-1234 into location A2." With GPT-4 processing the transcription, you can actually make that work. The system understands intent reliably enough to confirm the action without requiring you to type.

Smart notifications. Most mobile apps send dumb notifications—"You have a new order." AI-powered notifications could be smarter: "You have a high-priority order from your top customer; they've ordered this exact product 12 times before, so it's probably time-sensitive." This requires integrating AI into your backend to understand context, not running AI on the device.

Contextual search. "Show me orders from customers in Dubai who ordered more than $10K in the last quarter." Instead of building a query builder, the app listens to natural language and translates it to actual filters. On-device models for NLU are getting decent. This one actually works.

What Doesn't Work (And Why People Keep Trying)

Long-form text generation on mobile is a trap. The headline is always "AI chatbot in your app!" The reality is that chatbots are useful when they're solving a specific problem. A general-purpose AI assistant in your ops app? Nobody uses it. People use the app to do their job, not to chat.

Real use case I saw: a large logistics company added a "ChatGPT for logistics" feature to their mobile app. Drivers could ask questions. Sounds good. Usage was 0.1% of drivers. Why? Because when a driver needs information, they need it embedded in the workflow. "Show me which route is fastest given current traffic" is useful. "Open a chat interface and ask the AI about routes" is not.

Predictive analytics on mobile is another one that sounds good but rarely works. The fantasy is that your mobile app predicts something useful to the user. The reality is that predictions require data, context, and thinking that happens server-side. Mobile users want answers, not information. Surfacing a complex prediction requires UI design that most apps don't have. It becomes noise.

Another trap: using AI for feature recommendations or personalization in an enterprise app. Most enterprise users don't care. They have jobs to do. If your AP team is entering invoices, they want a fast interface to enter invoices. They don't want the app to recommend a different task. That's distracting.

The Real Pattern

Here's what I've noticed works: AI on mobile is most useful when it:

1. Reduces friction in a specific task. Not "general purpose intelligence," but "make this particular thing faster/easier."
2. Doesn't require waiting for the user. If it takes 3 seconds to get an AI response, that's annoying. Friction redirection, not friction reduction.
3. Surfaces answers, not conversations. Users want outcomes, not chats.
4. Works offline or with graceful degradation. Mobile networks are unreliable. If your AI feature requires constant connectivity, it will fail.

The winning pattern looks like this: user takes action → data is understood → the right next step is suggested or pre-filled → user confirms. Minimal AI, maximum usefulness.

Practical Implementation Approach

Here's how I think about adding AI to mobile apps:

Start with on-device models for fast feedback. Use smaller models locally. Vision, NLU, entity recognition. These can happen in milliseconds. Don't hit the server for every interaction.

Use server-side AI for complex reasoning. Your cloud inference can handle heavier models. Use it for decisions that require context you can't fit on a phone.

Design workflows, not features. Don't say "let's add a chat interface." Say "let's make data entry 30% faster using voice + context." Then figure out the AI that enables it.

Measure actual usage. If a feature gets sub-1% adoption, it's not working. The fact that it's powered by AI doesn't change that. Pull it and move on.

Test with actual users in actual contexts. Simulation doesn't matter. Real field workers, real networks, real pressure. That's where you find out if it actually helps.

For ERP Mobile Apps Specifically

The AI features that matter in ERP mobile apps:

• Barcode/QR recognition for data capture instead of typing
• Voice commands for common workflows
• Smart filters on large lists (show me what matters right now)
• Anomaly flagging (this order looks weird, review it)
• Predictive search in dropdowns

These aren't sexy AI applications. None of them are conversations with AI. But they actually reduce time on task and improve accuracy. And that's what matters in an operations app.

The Uncomfortable Truth

AI on mobile is useful, but it's not a feature in itself. It's a tool that makes specific workflows better. If you're adding AI to your app because the market expects it, or because your competitors are talking about it, you're wasting time and money.

If you're adding AI to solve a specific problem—data entry is too slow, classification accuracy is too low, users are making mistakes—then you might have something. Start there.

The apps that'll win in 2023 and beyond aren't the ones that shout about AI. They're the ones that use AI invisibly to make work faster and easier. That's the bar we should be setting.