AI · 7 min read ·
Applied AI in mobile apps that earns its place
Most AI features shipped in 2025 were a chat box bolted to a product that did not need one. The features that stuck did something narrow and verifiable: summarising, extracting, ranking, or removing a form.
Pick tasks with a checkable answer
Extraction, classification, summarisation and search ranking all have ground truth you can evaluate. Open-ended generation inside a regulated workflow does not. Start where you can measure accuracy and show the user the source.
- Document and receipt extraction
- RAG assistants over your own knowledge base
- Smart triage, routing and prioritisation
- Predictive churn, demand and risk scoring
On-device versus cloud
Small on-device models handle vision, wake-word, and privacy-sensitive text without a round trip and work offline. Cloud models handle reasoning and long context. The best products route between them and never make the user wait on a network call for something local hardware can answer.
Design for wrongness
Every AI surface needs a confidence signal, an obvious correction path, and logging of what the model saw and returned. Users forgive a wrong suggestion they can fix in one tap; they abandon a product that is confidently wrong with no recourse.
The takeaway
Ship AI where the output is checkable and the fix is one tap. Everything else is a demo.