Not a crystal ball: how AI can help governments manage debt

Rising debt costs are stalling development, UN Trade and Development (UNCTAD) warned this month in its latest World of Debt report. Global public debt reached $111 trillion in 2025, up from $49 trillion in 2010. Developing countries’ interest payments on public debt rose from $363 billion in 2010 to nearly $1 trillion in 2025, and in 51 developing countries, home to 3.7 billion people, interest payments now exceed government spending on health or education.

The cost gap is a large part of the story. Developing countries paid an average interest rate of 5.2% on public debt, compared with 2.2% in developed countries. Since 2022, interest payments on external public debt have exceeded net new lending to developing countries, which means external borrowing no longer adds resources for development. UNCTAD estimates that narrowing the borrowing-cost gap could save developing countries about $500 billion a year in interest.

The pressure is not confined to developing economies. The OECD’s Global Debt Report 2026 expects governments and companies to borrow a record $29 trillion from bond markets in 2026, 17% more than in 2024. Some 78% of OECD government borrowing this year will go to refinancing existing debt, and the shift toward shorter maturities lowers interest costs today while raising near-term refinancing risk. IMF Managing Director Kristalina Georgieva said in October that the growth needed to shrink debt ratios without budget cuts or tax increases is out of reach in the near term.

Those top-line figures, however, obscure a quieter problem. In many countries, debt managers must navigate this market with data that is incomplete, inconsistent or poorly structured. Artificial intelligence is increasingly offered as a fix. The evidence says it can help in some areas, with conditions, and not in others.

Why traditional debt-management tools fall short

At the country level, UNCTAD calls for stronger macroeconomic management and public institutions, improved debt portfolio structures and greater use of innovative financial instruments, and at the global level for more technical assistance to strengthen national capacities. Better data and analytics could plausibly support these goals, but only if the basics are in place. Three weaknesses stand out, especially in emerging markets and developing economies.

• Fragmented and inconsistent data: The IMF warns that many developing countries run public financial management systems that generate inconsistent, incomplete or poorly structured data, which is incompatible with machine learning requirements.
• Fragile forecasts: The World Bank’s DeMPA framework notes that debt managers need reasonably reliable and timely forecasts of cash balances to plan borrowing. Yet in many low-income countries, it observes, budgets are not always reliable, with revenues often overestimated and important expenditures underestimated.
• Incomplete coverage of liabilities: UNCTAD’s DMFAS programme has been extending debt-data coverage to non-traditional debt, and the conclusions of its 2025 advisory group refer to contingent liabilities, public-private partnerships and integration with budget and accounting systems, areas that debt systems are being extended to cover.

A warning score is a prompt for analysis, not a forecast to bet the budget on.

Where AI can help debt managers

None of the following is proven to lower borrowing costs. They are the areas where the institutions’ own work points to plausible value.

• Early-warning signals: A 2021 IMF working paper by Hellwig tested econometric and machine learning techniques for predicting fiscal crises out of sample. It found that machine learning can improve predictions, while stressing that the models identify predictors, not causes.
• Cash and liquidity forecasting: The OECD notes that machine learning forecasting models can increase accuracy and speed, but are black boxes whose assumptions and calculations cannot be fully understood. Explainability is therefore part of the design, not an afterthought.
• Cleaner, integrated debt data: Because models depend on data, the first AI project in a debt office is likely to be a data project: validating, standardising and linking records across agencies.
• Investor communication: A 2026 World Bank guide for debt managers finds that AI is beginning to influence investor relations mainly through internal enablement, and that ESG considerations now shape investor allocation in sovereign markets.

What public finance already shows

The OECD’s review of 200 real-world government AI use cases finds that applications are foreseen in virtually all areas of public financial management, but so far concentrate on a few domains such as forecasting, budget monitoring and financial management. The examples below are not debt-office applications, but they show what maturing systems look like:

• Türkiye applies anomaly detection to government accounting data, comparing spending units with similar units and with their own history. Its October 2025 presentation to PEMPAL reported around 100,000 spending units and 4,000 accounting units, with 75.5% of cases showing no anomaly, 22.5% a trend anomaly, 1.1% a cluster anomaly and 0.9% both. In principle, this lets reviewers concentrate on the small share of unusual cases.
• France’s tax administration reported that 56% of tax audits in 2023 were targeted using AI.
• Georgia processes about 70% of payments through a risk-based “Green Corridor” and is testing AI to identify risky transactions. The World Bank notes that centralising government accounts established the foundation for using AI.
• The pattern is consistent: foundations first, intelligence after.

What is not yet proven, and the guardrails

The evidence does not yet show that AI can reliably predict a sovereign crisis, and there is no established evidence that it lowers borrowing costs. Nor are results well documented: the OECD notes that impacts are hard to assess because the costs, performance and outputs of early AI projects are rarely disclosed, and that initial projects have shown significant risks of errors. A 2026 IMF technical note on emerging digital technologies makes a similar point: they could significantly transform public financial management, but technology alone will not.

The OECD lists ethical risks from skewed data and weak transparency, operational risks such as over-reliance on AI, and public resistance when scandals reduce trust. Australia’s Robodebt scheme is a cautionary case. It was an automated income-matching system rather than AI, but its lesson transfers: the Royal Commission’s 2023 report found that it produced inaccurate results and recommended stronger review and oversight of automated decision-making.

Four safeguards follow: a qualified official accountable for each consequential decision, a record of why it was made, a way to challenge it, and transparency toward the public.

Egypt is ready to move: a three-layer blueprint for AI-ready debt management

Egypt is well placed to turn these lessons into practice. The 2026/27 budget sets a clear fiscal direction, targeting a primary surplus of 5% of GDP and a debt-to-GDP ratio of 78% by June 2027. According to the OECD’s review of budgeting in Egypt, the law provides for a Fiscal Strategy that reconciles debt sustainability, revenue mobilisation, expenditure projections and fiscal risk assessments within a unified framework. And the Ministry of Finance has already published a medium-term debt strategy built on ministry of finance and planning data covering both domestic and external debt.

A legal framework, a debt strategy and official debt data are the foundations the international experience above points to. That means Egypt can move quickly from strategy to pilot. A realistic blueprint has three layers:

1. Data layer: Build on the data behind the medium-term debt strategy by unifying standards for debt and contingent-liability data across agencies, so that every liability is recorded once, completely and on time.
2. Analytics layer: Run a narrow pilot of cash forecasting and early-warning indicators for a limited set of entities, with human review and published performance metrics.
3. Governance layer: Require human oversight of consequential decisions, explainable models, a record of every decision and independent review of data quality and models, and expand only where the pilot shows measurable gains in forecast accuracy or reporting speed.

The link to sustainable finance is direct. Green bonds, sustainability-linked debt and tokenised instruments all rely on investors’ confidence that the state knows what it owes and where the money goes. A debt office with clean, integrated data on direct and contingent liabilities can issue with more credibility and respond faster when conditions change. As a design idea rather than an established practice, governments could also link green-bond proceeds to payment data, so that spending that departs from its earmarked purpose is flagged for review before the money moves.

Why this is the moment to act

Three trends are converging: debt costs are crowding out development spending, the IMF, OECD and World Bank are all treating AI as part of public financial management modernisation, and there is growing recognition that data foundations must come first. Three steps could turn that convergence into progress:

• Publish pilot results, including failures: The OECD’s call for better reporting on AI project metrics should apply to debt offices too. The Borrowers’ Platform, which UNCTAD supports as secretariat and whose ministerial-level Governing Council meets for the first time in Bangkok on 12 October, offers a natural venue for borrowing countries to share what works.
• Fund data foundations first: Debt-data quality and integration are what any model depends on.
• Make governance the entry ticket: Human oversight, explainability and independent review should be conditions of every pilot.

The next era of debt management will not be won by the government with the most advanced model. It will be won by the one with the cleanest data, the clearest accountability and the discipline to test before it scales. With its foundations in place, Egypt is well positioned to show what a disciplined pilot looks like.

illuminem Voices is a democratic space presenting the opinions of leading Sustainability Thought Leaders, their views do not necessarily represent those of illuminem.

The world needs sustainability knowledge. At illuminem, no interest group or shareholder can influence our work. Thank you for supporting our mission to make high-quality and independent sustainability information free for all. Every contribution helps. Thank you for donating today.