One Big Question is a monthly newsletter for directors and executive leaders navigating data, AI, digital transformation, and digital sustainability. Each edition is written by a governance leader or senior decision-maker and grounded in real boardroom experience and organisational accountability from New Zealand and internationally.
Issue 3 | May 2026 | by Bowen Pan
AI isn't software. It's workforce.
87% of NZ organisations now use AI, per Datacom's 2025 State of AI Index. Only 12% have scaled it. Nearly half are still piloting.
The headline says progress. The detail says we bought the tool and stopped.
This is not an AI problem. It is an organisational design problem.
There is a useful lesson from electrification. When factories first swapped steam engines for electric motors, output barely moved for nearly 30 years. The motors were better from day one, but the factories were still laid out around a central steam engine, with belts and shafts running through the building. Owners replaced the engine and called it done. The real gains came a generation later, when designers stopped retrofitting and started rebuilding around the new physics.
Most NZ organisations are in the replacement-motor phase right now. They have added the tool but not redesigned the work, and the boards governing them are following suit. One project, one model, one sign-off.
There is often a pattern in how AI projects show up in NZ companies. The AI paper comes with a budget, a vendor, a risk register, and a model approval. None of it answers who owns the output once the system is running. That question never makes it onto the page, because no one in the room is sure whose job it is to ask.
The AI systems being approved have already moved past the controls being designed for them. They take an objective, work out the steps, pull information from across the business, make judgement calls, and act. Five hours of work, done in five minutes, with no one watching each step. The work doesn't disappear, it changes shape. Execution moves into the system. People move up to judgement, strategy, and the problems that never got resourced because there were never enough hours. That's the real upside. Not just higher productivity, but a step change in the value that an organisation can deliver.
But the right reference point is not the org chart either. Org charts are built around human constraints, span of control, career paths, employment law. Agents don't have those. The unit that matters is the workflow: a sequence of handoffs that used to need multiple people, systems, and reviews, now collapsed into a single run. Govern the workflow, not the box on the chart.
Four things change when you make that shift.
1. The plumbing has to be AI-ready.
Clean data, consistent schemas, the right APIs. People need these too, but they can paper over the gaps with process and catch the mistakes because nothing moves that fast. Agents don't have that buffer. They move quickly, and they amplify whatever foundation you give them. The data and infrastructure debt you've been carrying becomes the ceiling on what AI can do for you.
2. Risk-tier the AI portfolio.
A tool that rewrites marketing copy does not need the same approval path as one that touches credit. Apply one control to all of them and you move at the speed of the slowest, most regulated use case. This is also where you decide what stays human-checked every time, especially in regulatory heavy areas. Everywhere else, the human sets the policy and watches for drift, not audits each step.
3. Adoption is cultural, not technical.
The hardest part is getting senior people to trust output they did not produce themselves. That doesn't happen with a town hall. It happens when one peer pulls visibly ahead.
4. Evaluation is something you build, not something you buy.
No vendor benchmark tells you whether the output is right for your firm. You need your own answer key, and you need to re-test against it every time the model changes. If your management can't show you their evaluation set, they don't have one.
The board doesn't need to understand how a transformer works. It does need to ask the questions it would ask of any other part of the business. Who owns the output. Who measures the work. Who decides when it gets more responsibility. Who is accountable when something goes wrong.
If your management team can't tell you who owns the output of each system, you don't have governance. You have tech procurement.
Director's Takeaway
If your board is still governing AI as a piece of software, it's time to change the narrative. Stop asking the management team for their AI strategy. Ask for the workflows AI now runs, and who is accountable for each one. Push management to table that list and for each workflow, ask how output is being evaluated.
If they can't answer, that's the gap to close before the next AI investment.
About the author:
Bowen Pan is a Kiwi business leader with global impact. He led the creation of Facebook Marketplace from the ground up and went on to hold senior product roles at Meta, Stripe, and Common Room.
Now back in New Zealand after 13 years in Silicon Valley, he advises high growth startups such as Partly and Marloo, sits on the boards of NZME, Milford and Revolut NZ and is the co-founder and co-chair of Growth NZ, an independent non-partisan group focused on the policy and structural conditions needed to lift New Zealand's long-term shared prosperity.
Bowen holds an MBA from Stanford University and 8 US patents in marketplaces design and developer applications.
