It's been three years since ChatGPT made enterprise AI a boardroom conversation. In that time, we've gone from "what is this?" to "how do we make it work?" The hype cycle has run its course. What remains is a clearer, more honest picture of what enterprise AI actually delivers, and what it doesn't.
What You Need to Know
- Enterprise AI works. But not the way most organisations expected. The value is in augmenting workflows, not replacing them. In compounding capability over time, not in one-off projects.
- The biggest lesson: foundation beats features. Organisations that invested in shared AI infrastructure are pulling ahead. Those that built isolated features are stuck.
- Pilot fatigue is real. Many enterprises have run 5-10 AI pilots. Few have scaled beyond them. The gap is operational, not technical.
- The NZ/AU market has matured significantly. Two years ago, most organisations were asking "should we use AI?" Now they're asking "how do we use AI well?" That's a meaningful shift.
3
years since ChatGPT launched, transforming enterprise AI from research topic to boardroom priority
Source: OpenAI, ChatGPT launch, November 2022
What We Got Right
Starting with business problems, not technology. Every engagement we've run starts with the question: "What's the actual problem?" Not "What can AI do?" The projects that delivered the most value were the ones scoped around specific, measurable business outcomes. Claims processing time. Compliance checking accuracy. Document turnaround.
Building for compound value. Our core thesis since 2023 has been that enterprise AI's value compounds: each capability you build makes the next one cheaper and faster. Three years in, this has proven true. Clients who built shared infrastructure (data pipelines, model orchestration, governance frameworks) are deploying new capabilities in weeks. Clients who built standalone features are starting from scratch each time.
Investing in discovery. The discovery sprint has been our most important process innovation. Two to four weeks of structured investigation before committing to a build. It's caught bad data, wrong assumptions, and misaligned expectations before they became expensive mistakes.
What We Got Wrong
Underestimating change management. Early on, we believed that if the technology worked and the ROI was clear, adoption would follow. It didn't. Every successful deployment has required deliberate, sustained effort to get people using the tools. The technology was 40% of the work. Adoption was 60%.
Overestimating model capability for edge cases. AI models are excellent at the 80% of tasks that follow common patterns. They struggle with the 20% that don't. We've learned to design for the edge cases from day one, with human fallbacks, confidence thresholds, and escalation paths.
Moving too fast on some pilots. A few early engagements went from discovery to production in 8 weeks. Too fast. The systems worked technically but hadn't been tested against enough real-world variation. We've since built more structured testing phases into every delivery.
The State of Play
Three years in, here's where enterprise AI actually stands:
What Works Reliably
- Document processing and extraction. AI reading documents, extracting structured data, comparing against requirements. This is the most consistently valuable enterprise AI pattern.
- Knowledge retrieval (RAG). Connecting AI to organisational knowledge bases for question answering, research assistance, and advisory tools.
- Classification and routing. Automatically categorising inputs (emails, claims, support tickets) and routing them to the right handler.
- Draft generation. AI producing first drafts of reports, responses, and analyses for human review and refinement.
What Works With Caveats
- Decision support. AI recommending actions for human approval. Works well when confidence thresholds are set correctly. Fails when the human in the loop stops actually reviewing.
- Customer-facing AI. Chatbots and virtual assistants with domain-specific knowledge. Works for narrow, well-defined domains. Fails for broad, open-ended queries.
- Process automation. AI handling end-to-end workflows. Works for standardised, high-volume processes. Struggles with processes that have many exceptions.
What Doesn't Work Yet
- Autonomous decision-making for consequential decisions. The technology can do it. The governance, liability, and trust frameworks can't support it yet.
- General-purpose enterprise AI. The dream of one AI system that handles everything remains a dream. Domain-specific, task-specific approaches win.
- AI without data quality investment. Still the most common failure pattern. Organisations that skip data work still get burned.
23%
of NZ/AU enterprise AI pilots have scaled to production, up from 8% two years ago
Source: RIVER, compiled from NZTech and direct engagement data, 2025
What's Next
The next phase of enterprise AI isn't about new models. The models are good enough for most enterprise tasks. The next phase is about:
Operational maturity. Building the organisational capability to run AI systems at scale: monitoring, maintenance, governance, continuous improvement.
Compound deployment. Moving from individual AI capabilities to platforms where each capability builds on shared infrastructure. The economics of AI improve dramatically when you're not starting from scratch each time.
Measured outcomes. Moving past "we have AI" to "our AI delivers measurable business value." This requires better instrumentation, clearer success criteria, and the discipline to shut down AI initiatives that don't deliver.
The hype phase is over. The hard work phase has been under way for a while. And the organisations that committed to it early are starting to see returns that justify the investment. That's not a story the hype cycle tells well, but it's the one that matters.
It's not the overnight revolution the hype promised - it's a slow compounding of capability that becomes visible over 12-18 months. The organisations that understood this and invested accordingly are the ones we're proudest to work with.
Isaac Rolfe
Managing Director