Approach: back to basics means spreadsheets
A simple database of audit recommendations was suggested. The logic was to demonstrate that, while reports might be delayed, recommendations were still relevant as they were not being addressed, despite being raised recurrently. The initial iteration of the tool was deliberately low-tech: a simple spreadsheet indexing all recommendations issued alongside updates on their implementation. It integrated an AI application capable of producing tailored reports for different audiences – parliamentary committees, civil society organisations and the media. Particular attention was given to ensuring that no specialised training was required. Kiribati was among the first and larger of the Pacific countries to adopt it.
In 2017 Kiribati adopted a new legislative framework for the Auditor General which reinforced their independence and capacity with secure lines of funding. The Kiribati Audit Office (KAO) was also operating in a political environment which put significant emphasis on anti-corruption and accountability; in 2020, the country hosted a regional forum on integrity which led to the adoption of the Teieniwa Vision, a regional anti-corruption framework.
From the outset, the Auditor General treated the idea of the audit recommendation tracking tool as a work in progress, not to be perfected in isolation before being released, but to be developed collaboratively with the actors whose behaviour it needed to shift. Working through the IT auditor, the KAO transformed the database into an interactive tool hosted on a web-based portal. Rather than waiting for a finished product, the Auditor General began producing audit reports and disseminating them in partnership with the Parliamentary Accounts Committee before the public-facing searchable version of the tool was released. Similarly, auditors were encouraged to liaise with audited entities to explain their recommendations and collect feedback before the client response function of the tool became operational.
The Auditor General’s explicit objective was to change the KAO’s reputation from punitive and compliance-focused to collaborative and advisory. He accepted, indeed planned for, the possibility that engaging more closely with auditees, parliament and the public would create demand for broader reforms within the KAO itself, and therefore placed the tool within the accountability ecosystem dynamics he aimed to change, not outside them.
In 2024 KAO representatives demonstrated the web-based upgrade of the tool at a regional event bringing together auditors and parliamentarians. The response was immediate: the Excel version, which had been carefully designed as a ‘no training required’ entry point, lost its appeal across Pacific SAIs overnight. Even teams that had invested heavily in the Excel database now focused exclusively on negotiating access to the web-based version, regardless of the fact that the first attempt to import the tool crashed the website of one SAI in the North Pacific, that another SAI was unable to afford the fees for basic proprietary subscriptions, and that the Kiribati team lost its IT auditor in 2025 and was forced to pause the public-facing search function of the tool.
Lessons: back to basics means planning for recalibration of positions and relationships
This dynamic is familiar to PFM reformers, and extensive guidance has been produced to resist the temptation of sophisticated tools, big bang transitions and overwhelming numbers of reforms when basics are not in place (Schick, 1998; Haque et al., 2016). Yet the approach of the KAO offers a more nuanced alternative reading. Demand for the web-based platform was not irrational or disconnected from constraints: it reflected a genuine understanding of what the more sophisticated functionality could accomplish, and was a direct response to documented challenges. Pausing the public-facing version of the web platform was not a failure for the Auditor General, because the office launched the PDF version of its citizen report, and continued to engage audited entities for collaborative work. This is a stark departure from digitalisation reforms where tool adoption is the end point and success measure – a metric that can encourage designers and reformers to design tools that limit disruption, or even reinforce power dynamics, to ensure adoption.
Feedback on the basic version of the tool, which focused on internal data management and reducing the cost of report production, was that it was missing functionalities that were key to the political dynamics at the root of the auditors’ challenge. This included dynamic and constructive engagement with audited entities on implementation and visibility of repeat recommendations. The important lesson was the intentional sequencing. The KAO’s approach was to build each functionality only when the relational conditions for its use were in place, but build with the outcome deliberately in view. The priorities set in the initial stage of reform should not be based on a linear understanding of PFM progress, but by a politically endorsed vision of the end goal.
AI-generated audit summaries: from simplification to relevance in Micronesia
This section draws on experience from the Federated States of Micronesia in using AI to produce accessible summaries of audit findings for legislators and citizens. It outlines implications for potential shifts in focus for technical assistance towards the strategic and relational elements highlighted by the Kiribati experience.
Problem: audit findings without audiences
Legislative assemblies across Micronesia play a prominent role in drafting and adjusting the budget. The flip side of this high engagement with the appropriation process is limited time and appetite for oversight. Public audit reports are rarely tabled and debated. Even where audit offices have high-quality reports including clear findings and recommendations, and have invested in accessible communication and risk-based audit selection to maximise impact, as is the case for the Office of the National Public Auditor (ONPA) (PASAI 2022), translating findings into follow-up action depends on channels that extend beyond the audit office itself.
Approach: from drafting to accompanying
A UNDP technical team supported ONPA on developing its manual audit recommendation tracking system. Alongside this, the UNDP team explored whether AI-generated summaries could complement the office’s existing work on accessibility. The assumption was that simplified plain-language versions of findings might make engagement less onerous for legislators, and thereby create conditions for follow-up action. Before the release of LLM tools to broad audiences, the role of traditional technical assistance teams was primarily around production. Consultation and engagement with stakeholders was mostly left to local institutions, who tended to hire consultants (often with limited exposure to audiences’ interests and perspectives) for the production of Citizen Budget Guides and other citizen-facing reports.
The original line of thinking was that AI would essentially do the technical drafting, replacing outside technical consultants. Yet the practical value of the AI tools turned out to be different, and more instructive. When the team used prompts such as ‘explain this audit to a non-expert’ or ‘translate this into plain language for a 15-year-old’, the outputs did indeed reduce technical complexity. At that point, practitioners with direct experience of civic engagement and audit communication were brought in, including a civil society activist from the Jamaica Accountability Meter Portal (JAMP) and former auditors from South Africa. Their feedback reframed the underlying question.
Effective engagement with legislators and citizens, they argued, depends not on making information simple, but on making it matter. The question that drives uptake is not ‘do you understand this?’ but ‘does this affect you, and how?’. The AI summaries, prompted towards simplification, produced outputs that stripped out contextual relevance and generalised findings that could have described any audit, anywhere. The problem was not that they were too technical for citizens. It was that citizens had no stake in them.
The UNDP team took this feedback as an opportunity to reconsider its own approach. In doing so, the process of outsourcing the drafting and seeking feedback, and the ease with which everyone involved agreed to reconsider the draft, revealed something unexpected. Because no one had written the AI-generated version, auditors and consultants engaged with it as a working object rather than a professional product, making it easier to criticise, question and receive feedback on. It also positioned contextual understanding much more prominently than in prior design experiences.
The technical team noticed that the distance AI created between the auditor and the document was shifting the conversation from document production towards audience understanding. This was something careful consultation processes had previously struggled to achieve, and contributed to the decision to present draft reports to legislators at a regional conference in 2025 before they were finalised, an act of institutional openness that would have been inconceivable under the previous model of polished consultations on finished outputs. The feedback from legislators was encouraging and ranged much wider than the drafts themselves, to include publication timing, collaboration on interpretation and emphasis and requests for increased contact time.