“If you’re going to save money on anything, don’t save it on cybersecurity," says Datacom CEO Greg Davidson
Datacom’s 2025 AI Impact in New Zealand report shows local organisations are embracing artificial intelligence. Eighty-seven per cent of businesses use some form of AI and close to a third are working with advanced AI.
Yet, Datacom CEO Greg Davidson says that while AI tools like ChatGPT are everywhere: “Most of what’s going on is still proof of concept, experiments and pilot projects. Many businesses, particularly in New Zealand, have yet to move on from that.
“Ninety per cent of what businesses are doing at the moment is chat; they are not fundamentally trying to solve an engineering problem.”
Workers who use an AI chatbot might ask questions and accept what comes back. Davidson says this is useful for small productivity boosts, but it is not transformative.
He says to get the best from enterprise AI, you need to deliberately design, test and build systems so that the AI delivers reliable, repeatable, business-critical outcomes. That requires engineering discipline.
Datacom’s work with Datapay’s AI-powered Payroll Assistant illustrates this point.
Davidson says: “We had to figure out a way of testing it. We had to carefully curate what information it referred to and we had to fundamentally alter how the RAG pipeline worked in order to get those accurate answers. In essence, instead of chat, it’s an engineering problem.”
RAG, or Retrieval-Augmented Generation, is the technical term for when an AI searches for a trusted, verified external information source, maybe the text of the original legislation or company manuals, instead of simply pulling information from a large language model or its training data.
This is important because Davidson says AIs are designed to please their users. That can mean they guess or make up plausible-sounding answers if they don’t have the correct information to hand. This is sometimes called hallucination.
He says when Datacom built Payroll Assistant, it restricted its information to documents from the IRD, ACC and MBIE. Then it engineered the AI to stop it guessing.
Real productivity benefits come when AI is engineered into processes rather than used as a novelty but that does not come cheaply.
For Davidson, a key is to invest only in the projects that can shift the dial. “There’s no point spending, you know, 1000 engineering hours on a problem that’s not going to give you the equivalent benefit on the other side of it.”
In a tight economy, boards hesitate to fund multi-stage projects without guaranteed returns.
“During a downturn, people want to know that they can get the benefit from the work that they’re commissioning.” Getting a return can involve a number of iterations.
Davidson says Datacom had to take a leap of faith when developing its Payroll Assistant AI to get the results it wanted. “The problem is that most organisations don’t want that degree of uncertainty. If they commission work, they want to know what they will get to justify the cost of the engineering effort.”
What makes this hard is understanding when to invest in engineering. Davidson says it can be deceptive because AI chat delivers ‘pretty good’ results. In a few places, the results can be amazing, but quality control is difficult.
“Most people are so impressed with the outcome, they don’t second-guess it. They take the lazy answer. When something produces the right answer, 99 times out of 100, spotting the one becomes a pain.”
New Zealand organisations’ focus on AI chat suits the main AI providers’ business models. Open AI with ChatGPT, Anthropic with Claude and Google with Gemini all earn revenue selling per-seat chatbot subscriptions.
Davidson says AI is most valuable when businesses treat it like any other critical system built — with testing, curated data, security guardrails and ongoing quality assurance. Without that engineering approach, he says AI stays stuck at the modest productivity gains from the “chatting with a bot” stage.
While local companies weigh up how far and how fast to push into AI, Davidson says the one area that cannot be deferred is cybersecurity. He says: “If you’re going to save money on anything, don’t save it on cybersecurity.”
The threat landscape is worsening, driven both by organised criminal groups and by state-sponsored actors. Too many New Zealand organisations are still behind on the basic controls that regulators regard as minimum good practice.
Added to that, worldwide AI adoption itself introduces new risks: models must be tested, secured and monitored to prevent bias, misuse or breaches.