Building is really addictive.

That may not sound like a conventional product-development principle, but for Ben Haywood, chief product officer at New Zealand’s Stuff Group, it captures one of the biggest challenges of building products in the AI era.

At the INMA Asia-Pacific News Media Summit, Haywood shared how Stuff has been experimenting with AI to rethink product development — from how teams collaborate and test ideas to how journalists, designers, engineers, and product managers work together.

Start with patience, not AI

Haywood began with a quality that may seem counterintuitive in an industry increasingly focused on speed: patience.

“I believe that one of the most valuable traits of product managers and product thinkers is patience,” he said.

By patience, he meant taking the time to understand a problem, trust the product process, explore possible solutions, and “hold your nerve” while building and delivering the right one.

That can be particularly difficult in news organisations, where deadlines, competition, and the need to get information to audiences quickly are built into the culture.

The danger, Haywood said, is that product teams get pulled into that same rush. 

“It’s easy to skip the thinking,” he said, with teams potentially becoming focused on building a list of things requested by others rather than stepping back to consider what will actually set the organisation up for the future.

Stuff had already responded to this problem by moving away from rapid delivery sprints toward a “mission” model: Teams would spend a defined period focused on one problem and one outcome, with business or audience metrics guiding the work. The aim was to ship something every week and continuously build on it.

AI initially made that existing process faster, Haywood said, but he realised that something more fundamental had changed.

AI changes the economics of product development

Haywood described the emergence of agentic coding tools such as Claude Code and Codex as “a really big step change” that “completely transformed the risk economics of product development.”

As AI dramatically lowered the cost and time of building and testing ideas, Stuff began exploring whether it could rethink how its teams worked. It created small, standalone groups of experienced staff, freed from their usual responsibilities and given permission to embrace AI and build from scratch.

Two major opportunities became the testing ground: a new editorial workflow and a new community product. Together, they became what Haywood called an “accidental A/B test” — two teams tackling different problems with very different approaches.

Two problems, two approaches

The editorial workflow project was deliberately exploratory. Stuff wanted to help journalists produce journalism in the many formats audiences increasingly demand without taking time away from reporting. The team included a principal engineer, a principal designer, an editor-in-chief, a national and in-depth editor, and Haywood as the product manager.

“We started with one engineer working with three devices and scores of agents, no tickets, no backlogs,” Haywood said.

The idea was to explore as many possibilities as possible, then identify which ones were worth pursuing. But the sheer abundance of possibilities created a new challenge.

The team had to determine which features could genuinely make a difference in the newsroom, polish those ideas, and bring them together into a workflow that journalists could actually use.

The community product followed a very different model.

Here, the team already had a clear understanding of the opportunity and could quickly define a minimum viable product (MVP). As a result, development was structured from the outset. The team used AI to help establish the product purpose, experience principles, technical architecture, design system, and launch scope.

The resulting process became a partnership between humans and AI agents. An idea could be turned into a specification by an agent, reviewed and revised by people, built by an agent, and reviewed again before a person ultimately approved it for deployment.

The principle was deliberately clear: “a human in the loop at every stage.”

AI is blurring traditional product roles

One of Stuff’s most important discoveries was that AI blurred the lines between traditional product roles.

“When you’re all using the same tooling,” Haywood said, what each person brings is “your insight and your expertise, your experience, your judgment, your taste.”

Journalists, designers, engineers, and product managers could work directly alongside one another rather than operating through transactional handoffs. This was particularly valuable when domain experts were involved from the beginning.

Haywood also found that experienced people were critical to getting good results from AI. The tools could produce enormous amounts of work, but determining whether that work was actually good required expertise and judgement.

AI also expanded creative possibilities. “When you’re not worried about how long or how complex something is to achieve,” Haywood said, teams gained greater freedom to be “more creative and more ambitious about what’s possible.”

Speed comes with new risks

The experiments also exposed the limits of AI-first development.

Moving quickly sometimes created instability, with overlapping work and new features breaking existing functionality. AI-generated design proved difficult to control, too, as tools sometimes imposed their own stylistic preferences despite Stuff’s established design system.

Perhaps the most important risk, however, was psychological. Haywood described building with AI as addictive: Because ideas could become products so quickly, it was tempting to skip the thinking and keep building.

“If we can take the time to really think things through and go back to those original principles of thinking about the problem,” he said, “we get to a better answer more quickly.”

There is also a human limit to the pace of AI-assisted work. “Just because the agents can work 24/7, it doesn’t mean that our people should,” Haywood warned.

The next step: bringing the lessons into the newsroom

Stuff now plans to take what it has learned from these isolated experiments and bring the approach into its mainstream product teams.

Haywood highlighted five priorities: 

1. Giving teams permission to experiment and break things.

2. Establishing strong foundations around data and design systems.

3. Automating testing and deployment.

4. Keeping domain experts directly involved.

5. Maintaining a healthy scepticism about AI.

For Haywood, that final principle is essential.

“It’s really easy to get kind of bedazzled by it,” he said. AI can produce results that appear astonishingly sophisticated, but “it’s not smarter than us.”

That is particularly important for news organisations, where editorial judgement and institutional knowledge remain fundamental.

“Nothing beats our taste and judgment,” Haywood said.

Stuff’s experience suggests that AI-first product development is not about removing humans from the process. It is about changing what humans spend their time doing.

AI can make building dramatically faster. It can make experimentation cheaper, enable teams to explore more possibilities, and blur the boundaries between traditional roles. But those advantages make judgement more important, not less.

As Haywood put it: “You run the AI, not the other way around.”

Banner art: Adobe stock image by Shweta.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .