AI can generate answers, automate tasks and help employees work faster. But making it work reliably across an organisation is proving more complicated than adopting the technology itself.
Across 156 questions submitted through Slido at CDAIO New Zealand 2026, data and AI professionals raised concerns that went well beyond model selection. They wanted to know how to prove business value, maintain trust in AI-generated outputs, manage costs, improve data foundations and encourage adoption without losing control.
Taken together, the questions point to five tensions that leaders must resolve as they move from AI experimentation to sustained business impact.
Organisations are under pressure to demonstrate AI value quickly. Yet many are still dealing with fragmented data ownership, inconsistent source systems and questions about the quality of the information feeding their models.
One attendee asked how much AI investment was actually going towards fixing the underlying data architecture. Another questioned whether AI had made meaningful progress in working with messy data, after earlier expectations that the problem would soon become less significant.
These are not separate challenges. AI initiatives depend on the quality, accessibility and context of enterprise information, while exposing weaknesses that may have gone unresolved for years.
The answer is not necessarily to perfect every data asset before starting with AI. That risks turning preparation into a permanent holding pattern. Instead, leaders need to identify which data problems materially affect a use case, address those first and establish clear ownership when issues cross departmental boundaries.
The question for leaders: Are we fixing the data problems that prevent business outcomes, or investing in foundational improvements without a clear connection to business priorities?
Human oversight is often presented as a safeguard against AI errors. But having someone review an output does not guarantee that the review is effective.
Questions at the conference explored how organisations manage inconsistent answers, identify partially correct outputs and prevent employees from approving AI-generated work without properly checking it.
One particularly important concern was automation bias: the tendency for people to accept a convincing AI response without applying sufficient independent judgement. As AI outputs become more familiar and frequently correct, reviewers may become less vigilant.
This creates a difficult problem. Organisations want to automate more work, but the ability to identify errors remains essential, especially when decisions carry financial, legal or reputational consequences.
Effective oversight requires more than a human approval step. It needs clear standards for acceptable outputs, appropriate testing, escalation procedures and a way to determine when a person must independently verify a recommendation. The level of scrutiny should reflect the consequences of getting a decision wrong.
The question for leaders: Can our employees recognise when AI is wrong, and do our review processes provide evidence that they are checking rather than simply approving its work?
Employees want AI tools that help them do their jobs. Organisations, meanwhile, need to manage security, privacy, spending and the risks associated with unapproved tools.
Several questions explored how to encourage experimentation without creating a fragmented environment of competing platforms, uncontrolled costs and inconsistent governance. Others asked how to respond when employees want access to tools beyond those officially provided.
Simply restricting access may not resolve the underlying demand. If approved tools cannot meet practical needs, employees may seek alternatives, potentially creating risks that are harder for the organisation to see.
The challenge is to make responsible AI use easier than working around the rules. That means providing useful approved tools, establishing clear boundaries for experimentation and making it straightforward for employees to request new capabilities.
Cost management also needs to extend beyond subscription fees. Leaders must understand usage across AI workflows, including multiple model calls and downstream services, and balance lower consumption against the quality of results.
The question for leaders: Are our controls helping employees use AI responsibly, or making it easier for them to work around the organisation?
AI investment increasingly comes with expectations of demonstrable business outcomes. Yet measuring those outcomes is not always straightforward.
Attendees asked how to calculate ROI, validate benefits claimed when use cases are registered and demonstrate value to boards. Others questioned how organisations should weigh cost optimisation against growth, or justify investment when promised benefits remain uncertain.
The difficulty is distinguishing activity from impact. The number of pilots launched, employees trained or tasks automated may indicate progress, but none necessarily proves that the organisation is better off.
A credible business case needs a baseline, an identified business problem and measures that reflect the intended outcome. Depending on the use case, that might mean reduced processing time, fewer errors, improved customer outcomes or additional revenue. Costs should include implementation, ongoing operation and the human effort needed to review and maintain the system.
Not every benefit will translate neatly into immediate financial returns. Better decision-making or improved employee capability may matter, but leaders should explain how those improvements will be assessed rather than treating them as assumed benefits.
The question for leaders: Who is accountable for verifying that the promised benefits have actually materialised, and what evidence would convince the business that the investment was worthwhile?
AI is changing what employees can produce and how quickly they can produce it. That raises a less immediate, but equally important, question: what happens to the expertise people need to do their jobs well?
Questions from the audience touched on AI literacy, employee fatigue, changing data team roles and the risk of losing the ability to judge whether an answer is correct. Others asked how organisations can retain knowledge as employees become increasingly reliant on AI.
The tension is that productivity gains today could come at the expense of capability tomorrow if people stop developing or practising the skills required to evaluate AI-generated work.
This does not mean organisations should avoid automation. It means they need to be deliberate about which tasks AI should take over, which decisions require human judgement and which capabilities employees must continue to develop.
Training people to operate AI tools is only part of the answer. Employees also need enough subject knowledge to challenge outputs, recognise uncertainty and take responsibility for the decisions they make.
The question for leaders: As AI takes on more work, how will we ensure our people retain the expertise needed to supervise it, challenge it and operate without it when necessary?
The real challenge is resolving the trade-offs
The questions raised at CDAIO New Zealand do not suggest that organisations lack ideas for using AI. They reveal the harder decisions that emerge when those ideas meet the realities of enterprise operations.
Leaders must balance speed with sound data foundations, automation with meaningful oversight, experimentation with control, investment with measurable returns, and productivity with long-term capability.
There is no universal formula for resolving these tensions. The right balance will depend on the business objective, the consequences of failure and the organisation's ability to manage the risks involved.
But one principle runs through all five: AI success depends on the decisions organisations make around the technology, not simply on the technology they adopt.
For data and AI leaders, the next stage is not just about finding more opportunities to use AI. It is about building the foundations, accountability and organisational capability to make those opportunities deliver lasting value.
The next CDAIO New Zealand happens October 2027, check out the full CDAIO ANZ series CDAIO Sydney, CDAIO Brisbane, CDAIO Melbourne, and CDAIO Perth to hear how data, analytics and AI leaders are solving real-world challenges.
Reach out to Vanessa Jalleh for more information.