The core-loop test
If users come to the product for what the model does — it reads, drafts, decides, recommends — it is AI-native and should be designed that way from the first screen. If the model improves a job the app already does — better search, smarter intake, summaries — add features to what you have.
| AI features added | AI-native app | |
|---|---|---|
| Design | New states in existing flows | Uncertainty, refusal and escalation designed into every screen |
| Architecture | Model calls inside existing services | Retrieval, memory, tool use, cost control as core services |
| Testing | Eval set per feature | Eval harness across the product |
| Typical cost | from A$25k per feature | A$60k – A$180k |
| Time | 4–8 weeks | 10–20 weeks |
Privacy and governance
Both need consent flows, data minimisation and Australian residency; AI-native products also need documented decision logic and a model-change policy, because the behaviour the user relies on can shift when the model does. The Privacy Act reforms landing in December 2026 make this a board-level question, not an engineering one.
Cross-platform or native
React Native or Flutter cover most AI-native products well; on-device models, camera-heavy or hardware-dependent features still favour Swift and Kotlin. Choose on the roadmap, not the first release.
Run a one- to two-week discovery that produces a prototype of every AI state and an evaluation set. It fixes the price and, more importantly, tells you whether the model earns its place.
Questions people ask
Can we start with features and go native later?
Yes, if the architecture isolates the model layer from day one. We build it that way so the upgrade path exists.
Are AI-native apps eligible for the R&D Tax Incentive?
Frequently, where there is genuine technical uncertainty. We document it as we go.
Related: AI-native & mobile apps · Custom software · Grant-funded builds