The AI Pivot

I studied Computer Science. I've been building digital products for 15 years. I've seen plenty of "the next big thing" come and go. Blockchain. VR. Whatever the last conference was excited about. Most of them were interesting. None of them changed the way I actually worked.

This was different. For the first time, I could genuinely talk to a computer and have it talk back in a way that made sense. Not keyword matching. Not scripted chatbot trees. An actual conversation, with context, with reasoning, with the ability to handle nuance. I called Rai. I called the team. Everyone needed to see this.

And then it dawned on us.

Wait. This thing is going to tell us how to do our jobs.

We're a small team of seniors. Engineers and designers who've spent their careers mastering their craft. The idea of an AI giving us suggestions on how to write code, how to structure a layout, how to architect a system? There was a moment, maybe a few days, where we just looked at each other like, "are we about to become obsolete?"

That passed quickly. But the discomfort was real, and it was useful. Because it made us take this seriously in a way we might not have otherwise.

Here's the thing about being a small team. You can't hedge. You can't spin up an "AI division" and let the rest of the company carry on as normal. When we decided to go all in on AI, that meant all in. The whole company. Every project, every conversation, every decision filtered through the question: how would AI change this?

The logic was uncomfortable but simple. If we ignored this and in two or three years found ourselves wondering "where did that come from?", that would be our own fault. We'd seen it happen to other companies. They saw the shift, assumed it didn't apply to them, and woke up one morning to a world that had moved without them.

So we sucked it up. Head down. We started with internal experiments. API trials with OpenAI as soon as they became available. Early RAG prototypes. We weren't in a rush to ship anything to clients, because we've never been comfortable releasing services until we know they're enterprise-ready. And AI in early 2023 is not enterprise-ready. But we need to understand it deeply, on our own terms, so that when the time comes we know exactly what it can (and can't) do.

The irony isn't lost on us. A team of people who specifically didn't want AI telling them what to do, betting the entire company on AI. But here we are!

Our first flagship AI project is already taking shape with UniMed.

The brief: build an AI health coach trained on New Zealand clinical data and grounded in Te Whare Tapa Whā, the Māori model of holistic wellbeing. Not a generic chatbot regurgitating WebMD articles. A coach that understands hauora (wellbeing) in the context of Aotearoa.

We've built personality controls so users can shape the experience. Sliders for how gentle or assertive you want your coach to be. Even how many emojis it uses (yes, really). It sounds playful, but the thinking is serious: health coaching only works if the person on the other end actually wants to engage with it. Giving people control over the tone is the difference between "another app I'll ignore" and something they'll actually come back to.

The plan is to fold Ora into UniMed's broader health platform over time. We're already seeing how one AI capability creates the foundation for the next one. Build the health coach, and suddenly you've got the infrastructure for clinical decision support, adjudication, triage. That compound pattern is becoming central to how we think about AI delivery.

The second flagship is coming from a very different direction.

Dr Tania Wolfgramm, our Chief Research Officer, has been doing supporting work on the Mana Wahine Kaupapa Inquiry reports with the Ministry of Justice. The work is important, but the source documents are incredibly difficult to use. Thousands of pages across hundreds of reports, scattered across formats, nearly impossible to search effectively.

After consulting with the MOJ team, Tania is leading the build of Hakamana: a RAG-powered knowledge base ingesting over 80,000 approved public documents from the MOJ corpus.

The architecture is designed for accuracy over speed. Each query first identifies the primary document bucket and secondary buckets, retrieves around 40 candidate references, cuts back to the 10-20 most relevant, then writes the answer with full citations including PDF page numbers and links to the live documents on the MOJ website.

In legal and policy research, a wrong citation isn't a minor inconvenience. It's a professional liability. Hakamana is being built for people whose work depends on being right, and the architecture reflects that. Early results are showing 95%+ retrieval accuracy, which is exactly where we need to be.

We don't know yet. Honestly. We're a few months into what we already know will be a long journey. The tools are changing weekly. The patterns are still emerging. We're learning something new on every project and every experiment.

What we do know is that we're not going to rush this. We've seen too many companies sprint AI products to market with impressive demos that fall apart under real conditions. "AI-powered" features that are barely more than a ChatGPT wrapper with a logo on top.

That's not us. We'll keep building, keep experimenting, keep learning. When we're ready to offer enterprise AI services publicly, it'll be because we've earned the right to. Not because the market is hot and we want to catch the wave.

For now, the work continues. And we're having a lot more fun than we expected.

This is the third chapter. Read Where It Started for the beginning, Why We Became RIVER for the rebrand, or continue to We Broke AI for Two Years for what happened next.