GUIDES · AI IN YOUR PRODUCT
Most AI features look finished in a demo and fall over in production. The ones that last solve a real problem, keep specialist behaviour in the system rather than a prompt, and leave the decisions that matter with a person. This page collects what we have written about building AI that way, drawn from products we have shipped.
The short answer
AI earns its place when it takes on a defined step in real work, gets tested on real inputs, and has a person accountable for what it produces. If you cannot say who uses it, at which step and what changes when it works, you do not have an AI feature yet. Sometimes the better fix is a clearer workflow, a well-built form or a proper report.
When AI is the right call, it goes in as part of the product, not bolted onto the side. That is how we work on AI Properly Embedded engagements, and it is how the products below were built.
Is AI the answer?
Start with how to stop shipping AI demos. It covers defining "good enough" before you build, testing on real inputs, designing the failure path and deciding in advance when to switch a feature off.
If the answer is still unclear, that is a decision worth making properly. Strategy & Audit is a fixed-scope review of your product, its data and where AI would genuinely help, and the written plan is yours whether you build with us or not.
Put the craft in the system
A prompt can ask a model to behave like an expert. It cannot make it remember, measure or schedule anything. Why an AI tutor is not a chatbot with a teaching prompt walks through how we built OLi Tutor with Olympus Insights, with memory, difficulty calibration and spaced repetition engineered into the system.
The full build, from commercial strategy to the platform around the tutor, is in the OLi Tutor case study. If your product depends on a specialist craft, whether that is teaching, triage, case management or compliance, the same question applies: where does the craft live?
Keep people in the lead
Human in the loop often means a confirm button. What the AI drafts, and what a person decides shows how we designed real decision gates into Nooma, the AI practice companion we built with O-HR, and gives a four-question test for your own product. The Nooma case study covers the governance and architecture behind it.
Large organisations are working through the same questions. Our notes from HR Tech Fest Sydney cover how One NZ, Adobe, Atlassian and others are deciding what stays human.
Build it to last
AI tech debt rarely comes from the model. It comes from how the model gets wired in. Adding AI without piling on tech debt covers seven decisions that prevent it, from giving AI its own service to pinning model versions and putting a ceiling on cost. If your AI product started on a no-code stack, AI app development beyond no-code covers what breaks and what is worth keeping, and the Beyond No-Code guide covers the wider move.
Handle the data
From 10 December 2026, organisations covered by the Privacy Act must explain certain automated decisions in their privacy policies. What changes on 10 December 2026 sets out what the obligation covers, why a person making the final call may not take you outside it, and the design decisions that reduce your exposure.
Questions
At a step where people already do repetitive, well-understood work and someone can check the output. Drafting, summarising and structuring are usually safer starting points than anything that makes a decision about a person. Start with one step, measure it, then widen.
Agree what a pass looks like before you build, test it on real inputs including the awkward ones, and keep testing after launch. Watch whether people come back to it and how much they edit what it produces. Heavy editing means it is not saving the time you think.
Model costs scale with usage, so a popular feature can become an expensive one. The main levers are choosing the smallest model that passes your tests, caching repeated work, and setting a spending limit per feature before launch rather than after a spike.
Not necessarily. From 10 December 2026 the obligation covers programs that do something substantially and directly related to a decision, not only programs that make it. If AI shapes what a decision maker sees, check whether the workflow is in scope. This is general information, not legal advice.
WHO THIS ISN'T FOR
If you want a chatbot bolted onto the side by next month, or you are looking for a problem to justify an AI budget, we are the wrong team. We start from the work and add AI where it earns its place, and sometimes that means telling you not to.
Not sure AI is the answer? See Strategy & Audit
Book a free 30-minute call. We will tell you plainly where AI would earn its place, where it would not, and what it would take to build it properly.
Software that holds up under real use.
© 2026 Castle Digital. All rights reserved.