When a Long-Term Plan consultation at Waitaki District Council drew 600 submissions—ten times the usual number—the small rural council faced a public-sector nightmare: a deluge of public input with no extra staff to process it. The solution came in the form of artificial intelligence, but with a governance framework so strict that every AI output was vetted by a person, every licence was ring-fenced, and not a single decision was left to a machine.

“We looked at it seriously in February of last year,” says Teresa McCallum, the council’s chief digital officer, who joined in early 2023. “We started with Copilot. So Microsoft Copilot is a large language model based on the same type of training as ChatGPT, but it is ring-fenced in the Microsoft environment.” That pragmatic choice—beginning inside the familiar Microsoft 365 ecosystem where staff already worked in Word, Excel, and PowerPoint—set the tone for an AI rollout that has since become a template for other councils.

Waitaki’s journey is not about bolt-on innovation. It is a story of careful tool selection, relentless human oversight, and a governance-first mindset that treats AI as an augmentation tool, not a replacement for human judgment. And it is working.

The Toolbox: Copilot for Daily Work, Claude for the Heavy Lift

The council’s AI arsenal is deliberately small. Microsoft Copilot, integrated with the council’s existing M365 tenant, handles daily productivity tasks: transforming first-draft Word documents into polished PowerPoint presentations, summarizing meeting transcripts, converting technical council language into plain English, and analyzing spreadsheets. Because Copilot runs under organizational accounts, the council could enforce enterprise controls from day one, shutting out the risks of consumer-grade chatbots.

For the Long-Term Plan submissions, however, a different tool was needed. The council used Anthropic’s Claude, accessed under a commercial contract that explicitly prevented any submitted data from being used to train the vendor’s models. Claude ran thematic analysis across the mass of public comments, grouping them into categories aligned with the consultation document. Staff then validated every theme and summary.

“It was very, very accurate,” McCallum told the Otago Daily Times. The automated theming saved weeks of repetitive reading, but the human check remained non-negotiable—a hallmark of Waitaki’s approach.

Governance That Puts Humans in the Driver’s Seat

Since May 2023, when the council’s formal AI policy was adopted, all usage has been governed by a cross-functional AI Governance Group. Comprising members from IT, legal, communications, and service delivery, the group monitors risks, steers pilots, and ensures no unauthorized tools creep into council operations. Its mandate is clear: AI may assist, but never decide.

“We will always have a person checking whatever AI does,” McCallum says. “We’re not using AI in a fully automated runaway and [we] make decisions, so we don’t have AI make decisions for us.”

That human-in-the-loop rule is backed by rigorous data-protection measures. The council ran Data Protection Impact Assessments (DPIAs) before deploying AI on public submissions. Its contracts with vendors include non-training clauses—no council data is used to improve foundation models—and strict audit rights. Prompts, raw AI outputs, and the final human-edited versions are logged and retained, creating an audit trail that can be examined under freedom-of-information rules.

The council is transparent with the public, too. While AI is used only as an internal acceleration tool, the policy commits to proactive communication if autonomous decision-making were ever considered. For high-stakes decisions like planning approvals, the council publishes assurance statements detailing what the AI did, who reviewed it, and how results were validated. This reduces the “black box” perception that can erode democratic trust.

The Numbers: 52 Licences, $21,000, and a Fiscal Case Study

Cost discipline is a central pillar of Waitaki’s model. The council purchased 52 ring-fenced AI licences—covering both Copilot and Claude seats—for a total annual cost of NZ$21,000. That figure, less than half the salary of one additional staff member, was deliberately presented to elected members and ratepayers as a value-for-money comparison.

“It’s about showing that we’re not spending recklessly,” McCallum says. “This is an efficiency lever, and the numbers prove it.” While the price tag should not be treated as a universal benchmark—vendor pricing varies by contract and scope—the approach of framing AI investment against staff costs has helped secure budgetary approval. The ring-fenced licences also guarantee that no personal data or council documents feed into external training pipelines, a stipulation often missing from consumer versions of the same tools.

The Risks: Hallucination, Bias, and the Fine Print

Despite the council’s careful controls, risks inherent to large language models remain. Hallucination—where AI produces fluent but incorrect information—is a well-documented pitfall. Even with human review, errors can slip through, particularly when summaries are used to inform policy papers. Waitaki mitigates this by maintaining raw submission data so that councillors and the public can verify AI-derived summaries. But the burden of auditability is ongoing.

Bias is another thorny issue. AI summarization relying on frequency counts risks privileging mass template submissions over unique technical comments from expert consultees. The council must calibrate weighting rules and ensure that statutory submissions are flagged separately. This is a design choice, not an AI shortcoming, and it requires transparent public communication.

Data-handling nuances demand more than vendor FAQs. While Microsoft’s enterprise Copilot and Anthropic’s Claude for Work both advertise that customer data is not used for training, the council’s contracts must lock in guarantees on data residence, deletion rights, and log retention periods. The fine print matters—a lesson Waitaki’s cross-functional governance group is well placed to enforce.

Public communication, too, must evolve. The council informs citizens that AI assists with summarization, but proactive publication of audit trails and correction mechanisms would further bolster trust. As McCallum notes, any shift toward autonomous use would require full public disclosure.

Fast Follower, Not Lone Wolf: Sharing the Blueprint

Waitaki’s “fast follower” posture—adopting proven enterprise tools and sharing policy templates—has resonated across New Zealand’s local government sector. Several councils in Canterbury and the South Island have adopted the AI policy drafted by McCallum’s team, and cross-council working groups now routinely exchange lessons on what works and what doesn’t.

“We’re not all reinventing the wheel ourselves,” McCallum says. “We’re using this collective approach, and we’re talking about how we’re using it, what’s working, what’s not working.”

This mirrors emerging best practice internationally. In the UK, Buckinghamshire Council has measured significant time savings from AI in contact centers, while emphasizing the need for governance frameworks and staff training. These case studies converge on a central message: AI can deliver real productivity gains in the public sector, but only when wrapped in policies that prioritize due process and public accountability.

A Checklist for Other Councils

For small to medium-sized councils eyeing a similar path, Waitaki’s experience offers a practical checklist:

  • Establish an AI Governance Group with cross-functional membership (IT, legal, communications, service leads).
  • Pick one or two pilot tasks with clear metrics: time saved, accuracy targets, and auditability measures.
  • Use enterprise-grade vendor offerings and insist on contractual non-training clauses for public-data uses.
  • Run DPIAs and publish a short public statement (what AI did, who reviewed it, how errors are corrected).
  • Require human sign-off for any AI-derived text that informs decisions or public communications.
  • Keep an audit trail of prompts, raw AI outputs and the final human-edited documents.
  • Invest in staff training and a champion program to build adoption and literacy.

Conclusion: Augmentation, Not Automation

Waitaki District Council’s AI rollout is a case study in pragmatic governance. By starting small, insisting on human oversight, and tying every licence to contractual privacy protections, the council has saved money, absorbed a tenfold spike in public feedback, and shared its policy freely with peers. It has not replaced human judgment; it has freed staff to exercise more of it.

As McCallum sums up: “We don’t have AI make decisions for us.” That principle—simple, non-negotiable, and backed by a paper trail—may be the single most important lesson for any public body venturing into AI. Waitaki’s blueprint proves that even the smallest council can harness cutting-edge technology without losing the human touch that underpins democratic service.