Balancing Innovation with Safety and Trust
Launching generative AI in the public sector posed challenges, including:
Pioneering Trustworthy AI in Government: Callaghan Innovation's GovGPT Journey
Callaghan Innovation partnered with HSO to launch GovGPT, a secure Azure OpenAI-powered chatbot that helps New Zealand businesses quickly access government information.
Callaghan Innovation, New Zealand’s innovation agency, set out to safely launch one of the country’s first generative AI pilots in the public sector. By partnering with HSO and leveraging Microsoft’s Azure OpenAI technology, they delivered GovGPT- a secure AI chatbot that rapidly connects businesses to government information. Follow their journey from idea and partner selection to implementation and outcomes.
Innovating an AI “Front Door” for Government
Callaghan Innovation supports Kiwi businesses through technology. In 2024, Callaghan aimed to simplify how businesses access government support information, creating a digital front door with GovGPT- a conversational AI assistant pulling information from multiple government websites. GovGPT allowed business owners to quickly access reliable information via natural-language queries.
Challenge
![]()
Ensuring accurate, reliable information to prevent misinformation.
![]()
Protecting against misuse, inappropriate outputs, and privacy breaches.
![]()
Building transparency and trust among users.
![]()
Satisfying rigorous government security standards
Red teaming is a comprehensive, goal-oriented security evaluation in which ethical hackers simulate real-world cyberattacks to identify vulnerabilities and evaluate an organization’s defenses and response mechanisms. In the case of GovGPT, HSO performed a focused red team exercise to probe the chatbot for weaknesses- evaluating its resilience to misuse, content manipulation, and data exposure. This process was essential to ensuring the AI system met the stringent safety, trust, and compliance standards required for public sector use.
Expert Partner Selection
Microsoft recommended HSO, a certified AI specialist. Sarah Sun, Head of AI & Digital, noted, “HSO was recommended by Microsoft as having the exact level of expertise we needed.”
Agile Red-Teaming and Risk Mitigation
HSO conducted rapid and thorough red teaming. Richie Atkinson, Solution Lead, praised HSO’s efficiency: “It was very fast and efficient - one of the fastest and most thorough assessments I have seen. Our CISO was very happy with it.” HSO uncovered vulnerabilities beyond Azure’s built-in protections and provided actionable solutions, addressing all major concerns within 24 hours.
Sarah Sun stated, “Having HSO’s technical expertise built into the pilot made us feel much safer and more comfortable.”
HSO also guided prompt engineering and implemented robust safeguards to ensure transparency and security. According to Richie Atkinson, “HSO’s agility meant we went from zero to done in just three or four days.”
Architecture and Implementation
GovGPT was deployed in a secure, isolated Azure environment. While Azure Blob Storage was used to hold curated public sector data, the real value came from how HSO indexed and enriched that data using Azure AI Search enabling fast, relevant retrieval through natural-language queries. This was then integrated with Azure OpenAI’s ChatGPT model to power intuitive, conversational interactions. HSO initially conducted a rigorous red team assessment to identify and mitigate risks before public launch, but their role extended well beyond initial testing guiding the architecture and ensuring the solution adhered to responsible AI principles.
In the second phase, HSO partnered with Callaghan Innovation to build a scalable, production-ready foundation for future reuse and public-sector alignment. This included:
- 1
A secure dual-environment Azure architecture
for production and UAT—completely isolated from internal networks to mitigate security risks.
- 2
A modular, multi-LLM architecture
enabling future extensibility beyond a single model or provider, while adhering to enterprise governance standards.
- 3
A formal evaluation framework
to monitor AI performance, flag anomalies, and assess outputs across multiple dimensions such as accuracy, tone, and relevance.
- 4
Advanced prompt engineering and safety mechanisms
including refusal logic for unsupported queries, embedded sourcing, and layered content filtering.
Together, these capabilities ensured GovGPT could deliver a safe, transparent user experience at scale—ready to serve as a reusable model for other government agencies.
Outcome
A Secure, Successful Launch
GovGPT’s safe, high-impact deployment set a precedent for responsible AI in government—transforming a bold idea into a validated, trustworthy solution.
GovGPT launched publicly in October 2024, achieving:
![]()
Processing over 65 million tokens
in the first 3.5 months
![]()
Engagement with over 30 public sector agencies
prompting reuse, code sharing, and broader exploration of generative AI.
![]()
Positive media coverage and strong political endorsement
including recognition as a model for “trustworthy AI” by New Zealand’s Minister for Science, Innovation and Technology.
![]()
Zero major incidents reported
thanks to HSO’s proactive red-teaming, robust architecture, and real-time evaluation framework.
Callaghan Innovations Testimonials
Sarah Sun - Head of Digital & AI
- “The first engagement really showed the team’s capability, and trust was a no-brainer. We knew they were absolutely the right people to do the work.”
- “Having HSO’s technical expertise built into the pilot made us feel much safer and more comfortable.”
Richie Atkinson - Solution Lead
- “It was very fast and efficient—one of the fastest and most thorough assessments I have seen. Our CISO was very happy with it.”
- “HSO’s guidance enabled us to quickly move forward. By Saturday afternoon we had checked all the boxes to satisfy the security team.”
Expertise, Agility, and Safety at Scale
HSO provided far more than a point-in-time assessment—they were a strategic co-development partner across architecture, governance, and implementation. Their contributions spanned:
- 1
Deep AI and Microsoft Azure expertise
enabling Callaghan to fast-track development while upholding public-sector standards.
- 2
Rapid delivery with agile execution
helping the team move from prototype to production in record time—without compromising safety.
- 3
Strategic solution design
including the multi-LLM framework, evaluation models, and content guardrails that made GovGPT resilient, scalable, and transparent.
- 4
Capacity building
equipping Callaghan’s internal team with the tools and knowledge to maintain and expand the platform independently.
By embedding expertise directly into the pilot codebase, HSO ensured that AI safety was not just conceptual—it was operational.
A Blueprint for Public Sector AI Adoption
Callaghan Innovation’s GovGPT pilot demonstrates what’s possible when governments pursue AI innovation responsibly, and at pace. By partnering with HSO and leveraging Microsoft’s trusted AI infrastructure, Callaghan delivered a transformative solution that was safe, transparent, and reusable.
The approach offers a repeatable blueprint: align on a bold vision, bring in specialized partners early, and embed governance into every phase of development. Today, dozens of agencies across New Zealand are following in Callaghan’s footsteps—building on the GovGPT foundation and accelerating public sector transformation.
Note:
The GovGPT pilot has since concluded, and the live chatbot is no longer publicly accessible. However, the full project—including architecture, components, and code—remains available via GitHub for learning and reuse by other public agencies.
HSO is proud to have played a key role in this journey delivering not just a successful launch, but a scalable framework for future government AI.
Connect with us!
Please get in touch to discuss your requirements with one of our Microsoft experts. We would be delighted to have the opportunity to talk to you about how HSO can help transform your business.
Learn More
HSO Knowledge Base