top of page

A Receipt for Every Decision: Stuart Wakefield on Building Trust into Digital Government

Charlie Hamer
4 days ago
6 min read

When New Zealand's Ministry of Education built Te Rito, the information exchange that follows a learner as they move between schools, one of the hardest-working design artefacts was a privacy impact assessment. Documents like that are usually written in legalese, for lawyers. Stuart Wakefield's team rewrote theirs in plain language and tested it with parents until they could understand exactly what the system did, and didn't do, with their child's data. "There are little things like that," he says, "but important ways of communicating what a system does or doesn't do."


Stuart has led digital and data at the Ministry for more than a decade, first as Chief Information Officer and since 2018 as Chief Digital Officer. His career runs through the centre of New Zealand's public sector, from directing government ICT strategy in the Office of the Government CIO to leading ICT for a major district health board. He also chairs TechLeaders, the country's community of senior technology leaders, and is a considered public voice on AI and automated decision-making in government.


What stayed with me from our conversation is how much of his thinking starts from a single obligation. Public services are for everyone. A bank can segment its customers; a ministry of education cannot. Everything else Stuart has built follows from taking that seriously, from active consent in Te Rito to his case for decision receipts on automated decisions, and it adds up to an answer to the question every government is asking: how fast can we adopt AI without losing the public's trust?


The obligation you can't segment away

Stuart spent part of his career in the private sector, and the contrast still frames his thinking. A for-profit business "can make a reasonable commercial choice to segment your customer base and focus just on a segment that you find profitable," he says. Mandatory and legislated public services carry no such option. Universal provision means careful attention to digital equity and accessibility, and it means keeping channels open for the people who depend on them.


A long tenure inside one agency has taught him a second lesson: change takes longer than the business case says. "There's often quite a long tail of user adoption," he says. "I do see quite a lot of over-optimistic assumptions about how quickly change will occur, particularly in large ecosystems." The honest business case builds that tail into its assumptions from day one.


Levers for a system you don't control

Education is devolved. The Ministry works with thousands of largely self-governing schools and kura, and its users start at age two. "You don't have a lot of direct control levers in the same way you might in a single large enterprise," Stuart says. What works is a toolkit, "standards and accreditation schemes, funding incentives, or sometimes direct intervention in the market", combined in whatever mix drives the outcome.


Children raise the bar further. Stuart's starting position "is always to maximise the beneficial effects while minimising or eliminating the harmful ones". Screen time is the familiar example; the effect of AI on a child's cognitive development is the emerging one, and the evidence is unsettled. "Where we don't have clear evidence, we do need to proceed with caution."


Consent a parent can read

Te Rito was co-designed with sector advisory groups around a principle: protect the mana of the child and their whānau. In practice, that meant active consent. "Students, their parents and guardians are the ones who ultimately decide whether their data is made available," Stuart says. The plain-language privacy impact assessment flowed from the same principle, because consent means little if the document explaining it needs a law degree.


The hard part is patience. Landing Te Rito took years of work with the sector before it fitted the way schools and kura operate. Stuart's view is that the time buys durability. Involving the people a system serves "can add time and complexity to an initiative, but in my view it gets you to a better outcome and a much more enduring solution."


A receipt for every machine decision

Automated decision-making is well established in government, and so are its failures; Stuart points to the UK Post Office scandal and Australia's robodebt. Historically, agencies automated narrow cases with tightly defined rules and deterministic outcomes. What he sees emerging now is agentic AI positioned to augment or replace human judgment, even where discretionary decisions are vested in statutory powers. "That is a prize worth pursuing," he says, "but we do need to think about being able to explain and potentially defend the decision an agent has made or recommended."


His answer is the decision receipt: capture the metadata at the moment a consequential decision is made. That means the rules and evidence behind it plus the exact version of the model that made the call. The version matters because models change quickly and behave in non-deterministic ways. "A decision made by an agent a year from now on the same data set could theoretically be different than one made today." In many jurisdictions citizens hold a legal right to challenge decisions, so an agency has to be able to reconstruct what its machine did.


He is sceptical of "human in the loop" as a permanent safeguard. It is, he says, "a nice comfort blanket statement, but I do wonder how effective that will be over time when we become more reliant on machines making decisions". Once a machine is right 99 times in a hundred, the human checking it grows complacent and the oversight hollows out. The durable fix is structural: design receipts into the next generation of digital public infrastructure at the start, because retrofitting them later is close to impossible. DPI is gathering international momentum and useful patterns are starting to form. Stuart sees the receipt as a candidate to become one of them.


Assurance the market wants to buy

Safer Technology for Schools takes a different route to trust: one trans-Tasman scheme assesses the security and privacy of edtech products so that thousands of individual schools don't have to. Joining forces with Australia, where standards were already closely aligned, made the scheme workable "at scale and cost effectively".


What keeps it off the shelf is that it is sold, to both sides of the market. For vendors, accreditation is a differentiator rather than a compliance cost: a New Zealand edtech accredited at home can enter any Australian state education market under the same scheme. For schools, the demand side, the message is to buy accredited products. "In short, it is selling the concept to both the supply and demand sides."


For a jurisdiction starting from zero, his version one is modest: "I would start with the most sensitive systems first, those that contain sensitive personal information, and adopt an existing security standard if one doesn't exist, then start assessing systems against that."


Re-pitch the value proposition

Multi-year programmes outlast their budgets and their ministers. People change out in delivery teams and in the governance and ministerial layers above them. Stuart's discipline is to hold the purpose steady while everything around it moves. "It is essential to be able to continually pitch the value proposition to people who may not have been involved at the earlier stages." Don't assume new arrivals will buy in by default.


When pressure to cut corners arrives, the test is knowing which scope drives the value and which is optional. "Defend the core but be prepared to compromise and trade off anything else."


Keep the metadata with the data

Ask Stuart how to engineer trust and he starts somewhere unglamorous. "Understanding the classification of the data, the basis on which data was collected and what for must be preserved over time so that data is not able to be misused or taken out of context," he wrote to me before our call. Copies of data drift between systems and lose their provenance, and AI sharpens the problem because it puts old data to new uses.


The other half is giving people a working interface to their own permissions. Stuart calls it a citizen account view: a place to see who you have granted access or agency to and revoke it, much as you can already view and change your approved tax agent. "Without this trust-based architecture I think we will struggle with social licence for the use of agentic AI for more sophisticated use cases."


He is watching the other side of the agentic shift too. Governments are busy working out how to onboard their own agents; citizens and businesses are already arriving with theirs. The performance differentiator he expects over the next decade is "whether someone else's AI agent can both discover your services and then interact with them".



The pattern across the conversation is consistent. Take the universal service obligation seriously, then build the infrastructure that honours it, whether that is consent a parent can read or a receipt that lets a machine's decision be challenged. It maps to what we call Government 3.0: connected, citizen-centred government where trust is engineered in rather than asserted.


Stuart's closing advice for leaders in the middle of the AI wave came without hesitation: "Don't ever lose sight of the people you are here to serve." What the hype never prices in is the stakeholder who isn't ready yet, which is why he argues for maintaining "legacy options for as long as it is necessary for them to continue to be served".


AI is already sitting in classrooms while teachers and students work out its place. Stuart's bet, and mine, is that the governments who move fastest will be the ones who made themselves trustworthy first.

 
 
 

Comments


bottom of page