← AI App Building (No-Code & Low-Code)
Day 13 of 14
Building an AI Product Strategy
The Commoditisation Trap
As AI models become more capable and cheaper, features built purely on 'we use AI' will be commoditised. Any competitor can call the same Claude API and replicate the surface behaviour of your AI feature. The defensible layer isn't the AI itself — it's the data, the domain expertise, and the workflow integration.
For BathCheck: a generic 'AI assessment report generator' is replicable. But 'AI trained on 50,000 actual OT bathroom assessments, with outputs fine-tuned to Australian NDIS reporting standards, integrated into the OT's existing workflow, and connected to their Xero for automatic invoicing' is not easily replicated. The moat is: (a) proprietary training data, (b) domain-specific accuracy, (c) deep workflow integration.
Data network effects are the strongest AI moat for specialised apps. As more OTs use BathCheck, you accumulate more assessment data. This data can be used (with proper consent and privacy controls) to improve your AI models, making BathCheck more accurate for your specific domain than any general-purpose AI tool. This is the basis of a genuinely defensible AI product.
Fine-Tuning vs Prompting vs RAG
Three approaches to improving AI output for your domain, in order of increasing investment and defensibility:
Prompt engineering (hours to implement): improve output quality by providing better context, examples, and instructions. The fastest iteration cycle. Appropriate for early stage and for tasks where prompt-tunable improvements are sufficient. Not defensible — competitors can see and copy prompting strategies.
RAG (days to implement): add domain-specific knowledge to each query via retrieval. Provides access to information not in the model's training data. More defensible if your document corpus is proprietary (your accumulated assessment templates, your NDIS compliance documentation). Pairs well with Supabase pgvector.
Fine-tuning (weeks to months, significant data required): train the model on your domain-specific examples. Requires thousands of high-quality input-output pairs. The most defensible approach — the fine-tuned model embodies your domain knowledge in a way that's hard to reverse-engineer. For BathCheck, fine-tuning on 10,000+ actual OT assessment reports would produce report quality that generic models can't match at any prompt.
⚡ Today's Action
Identify your highest-value AI feature and ask: what would make it 3x better than any generic AI could produce? Is the answer better prompting, better knowledge (RAG), or better training data (fine-tuning)? Map out the data collection strategy for the approach that fits your current stage.
💡 Pro Tip
Start building your training dataset now, even before you plan to fine-tune. Every time your users edit an AI-generated output, log both the original and the edited version. After 1,000 edits, you have a fine-tuning dataset. After 10,000, you have a genuine competitive advantage.