Central banks are usually the last institutions anyone expects to move fast on new technology. They hold the most sensitive data in a financial system, their mistakes move markets, and their legitimacy depends on decisions that can be explained. That is exactly why the way they adopt artificial intelligence is worth watching. In a little over two weeks, from 28 October to 13 November 2026, four publications gave a rare, evidence-based view of AI in central banks: a global survey from the Bank for International Settlements (BIS), a BIS working paper on an AI agent for econometric research, a National Bank of Ukraine paper on big data and language models in monetary policy, and a Czech National Bank review of how its machine-learning GDP nowcasts have performed. Read together, they say a lot about what "AI readiness" means for any regulated financial institution, not only for central banks.
What the BIS survey says about central bank AI readiness
On 28 October 2026 the Irving Fisher Committee on Central Bank Statistics (IFC), which sits at the BIS, published IFC Report No 19, AI readiness in central banks. It is based on a survey conducted in early 2026 across 65 jurisdictions and compares the results with a similar survey from 2024. The authors define AI readiness as an institution's preparedness to adopt, adapt and advance AI in its operations, decision-making and objectives, and they split it into two parts: organisational readiness (governance, resources and capabilities) and data readiness (managing information so that AI can use it effectively, securely and responsibly). They also make a useful distinction between perceived readiness, how an institution rates itself, and actual readiness, what it has really done.
The headline findings:
- AI has moved up the agenda. Almost two thirds of central banks now treat AI as a high strategic priority, up from less than half in 2024.
- More AI is in production. The share of AI projects in production, including small-scale production, is over two thirds in 2026, compared with about half two years earlier. The number of respondents not using AI or machine learning at all has fallen to almost zero.
- General-purpose tools lead. Usage has shifted towards general-purpose AI such as productivity copilots, while agents and computer vision remain less used, which the report links to trade-offs in cost, security, integration and governance.
- Governance has caught up. The share of central banks without internal rules or recommendations for AI use has fallen to less than one tenth, from about one third in 2024.
- Infrastructure and people are in place, but thin. Almost two thirds report dedicated AI infrastructure, and almost three quarters employ AI experts, yet those experts often make up less than 2.5% of the workforce.
Where the risks are concentrating: cost, vendors and autonomy
The more interesting part of the report is what has not improved. Expectations about AI's impact remain high but have moderated in core areas such as monetary policy and financial stability, which the authors connect to the difficulty of moving from pilot to full-scale projects. Most central banks still spend less than 5% of their budget on AI, and many spent less than they had planned in 2024, even though they expect budgets to rise.
Three risk signals stand out:
- IT cost. The share of central banks flagging IT expenditure as a risk has jumped to almost 80%, covering computing, storage, cloud, integration with legacy systems, vendor and licensing fees, security hardening and compliance tooling.
- Vendor concentration. Central banks mostly use external or off-the-shelf models, with only about a fifth self-hosting or deploying models in-house only. The report notes high reliance on a few providers for general-purpose AI and a limited number of suppliers for frontier models, and argues for open, interoperable solutions and common protocols to avoid lock-in.
- Autonomy without finished controls. Governing increasingly autonomous AI while keeping effective human oversight is identified as a key challenge. The report cites controls such as human-in-the-loop mechanisms and "kill switches" to halt autonomous processes, yet nearly half of central banks are still in the early stages of implementing these mitigations, and fewer than 10% are approaching completion.
The recommendations are practical. The report suggests clear decision boundaries: where AI may be used, where it is unsuitable and where human control is essential. Its own example is that AI could assist with coding or information retrieval while human approval would remain mandatory for policy work. It also recommends diversifying reliance on model developers and cloud providers, adopting open standards and contingency arrangements, and managing cost through "AI tokenomics", a framework for tracking how generative models consume tokens. A box in the report discusses open-weight models, which it says reportedly represented over half of commercially available AI models as of 2025, as a way to reduce cost and dependence, while warning that deploying them securely needs infrastructure and expertise.
Agentic AI in central banking: MaRCo@BIS keeps the economist in charge
On 12 November 2026 the BIS published Working Paper 1395 by Marco Jacopo Lombardi, Taejin Park and Fernando Perez-Cruz, introducing MaRCo@BIS, the Macroeconometric Research Companion. It is a concrete example of what agentic AI in central banking looks like when it is designed for accountability.
MaRCo@BIS combines an off-the-shelf agentic coding environment with plain-text workflow files, a verified dictionary of published model specifications and a tested library of econometric methods. The agent proposes models and explains their economic rationale, retrieves and harmonises data, runs estimation and diagnostics, and logs every choice for replication. The researcher must confirm or override each consequential decision. The authors benchmarked the agent's replications of three widely used structural vector autoregressions (VARs) against the originals and ran five exercises, from a verified replication to an open design question.
The findings are measured. The agent reproduced published results where the specification and data were well defined, and substantially reduced the mechanical effort of research. In an open exercise on fiscal policy, short-run results were more robust than medium-run effects, which depended heavily on modelling choices. The authors conclude that AI should support rather than replace expert judgement, because of the risks of cherry-picking specifications, automation bias, data limitations and economically inappropriate interpretations. The framework is released as a public good for central banks and macroeconomic researchers.
Big data and language models under pressure: the National Bank of Ukraine
On 6 November 2026 the National Bank of Ukraine (NBU) published an occasional paper by Volodymyr Lepushynskyi and colleagues on big data and AI as inputs into monetary policymaking. The context is extreme: after 2022 the war interrupted official statistics, and the Monetary Policy Committee still needed timely readings of prices and activity.
According to the paper, the NBU's assessments draw on millions of online prices, Google Trends, satellite night-light imagery, online job vacancies, fiscal data and indicators of military activity and infrastructure damage. More recently the bank has used large language models to analyse text, including news, social media and its own communication, and has built a prototype system of AI agents that simulates household survey respondents. The paper traces a decade of this work and sets out three principles for applying AI-generated signals to monetary policy decisions and communication. For a technology leader the lesson is that alternative data and language models become valuable when ordinary data fails, but each new source adds its own uncertainty that has to be managed explicitly.
AI nowcasting of GDP: the Czech National Bank's scorecard
On 13 November 2026 the Czech National Bank (CNB) published an updated blog post, timed to a lecture by Governor Aleš Michl at Kyoto University, on how its GDP nowcasting models have performed since 2023. It is one of the clearest public scorecards of AI nowcasting of GDP by a central bank.
The CNB runs a suite of models: its Rushin index built on high-frequency data, a dynamic factor model developed at the Federal Reserve Bank of New York, and machine-learning methods including LASSO, Ridge regression, Elastic Net, support vector machines and Random Forest. Since 2023, the CNB says, its economists, using all the models plus their own judgement, have missed quarterly GDP growth by 0.21 percentage points on average, while the best machine-learning model, Random Forest with no human input, missed by 0.22. The average error has fallen to about 0.2 percentage points, compared with below 0.7 before the pandemic, and the forecast is almost 35% more accurate than a simple autoregressive benchmark, against about 15% before the pandemic.
The bank's conclusion is careful: the human-adjusted forecast is still the most accurate, but only just, and the best result comes from combining models with human judgement. The CNB will now update model weights every quarter based on recent performance. It also describes building its own AI technology base: an innovation hub called CNB Lab, and servers with NVIDIA H200 graphics cards for working with data and code in a secure environment.
What technology leaders in banks should take from this
These four publications come from institutions with very different sizes and circumstances, yet they share a pattern that applies to commercial banks, payment firms and any regulated organisation:
- Production is the new baseline. The question is no longer whether to use AI but whether you can show where it is used, who approves its outputs and how it performs over time.
- Human sign-off is designed in, not bolted on. MaRCo@BIS makes the researcher confirm each consequential decision; the CNB keeps expert judgement on top of its models; the BIS report recommends explicit decision boundaries. This matches the direction US supervisors set out in the OCC AI supervisory framework, where autonomy and consequence drive the level of control.
- Vendor dependence is now a named risk. When a central bank survey ranks IT cost and concentration among rising risks, boards will ask the same questions of commercial banks. Hosting some models yourself is one option; I covered the trade-offs in Self-hosting LLMs.
- Scorecards beat demos. The CNB's published error comparison between humans and models is the kind of evidence that builds trust. As I argued in AI agent benchmarks in 2026, a benchmark result is only a hypothesis until it is confirmed on your own tasks and data.
A practical playbook for AI readiness in regulated finance
- Measure actual readiness, not perceived readiness. Keep an inventory of AI use cases with their stage (pilot, small-scale, production), owner, data sources and approval point. Self-assessments are useful only when checked against this list.
- Write decision boundaries down. For each business area, state where AI may act, where it may only assist and where human approval is mandatory. Review the boundaries when a model, tool or permission changes.
- Treat cost as a design input. Track token consumption and compute per use case from the start, so that scaling a pilot does not become a budget surprise.
- Plan for vendor exit. Prefer open interfaces, keep prompts, evaluation sets and logs portable, and test at least one alternative model, including an open-weight option, for critical workloads.
- Finish the controls before raising autonomy. Human-in-the-loop checkpoints, tested kill switches and full decision logs should be complete before an agent is allowed to act rather than suggest.
- Publish a scorecard internally. Compare AI outputs with human and baseline results over time, as the CNB does for nowcasts, and decide model weights or trust levels from evidence.
- Make data AI-ready. Classification, lineage, metadata and access rules determine what AI can safely use. The BIS report treats data readiness as half of AI readiness, and it is usually the slower half.
Closing thought
The most reassuring message from this fortnight of central bank publications is not that AI is getting better, although it is. It is that the institutions with the highest stakes are adopting it with explicit boundaries, logged decisions and published comparisons against human judgement. That is a model any bank or financial technology team can copy today, long before it needs to decide how much autonomy to give its first AI agent.
References
- AI readiness in central banks (IFC Report No 19), Rafael Schmidt, Olivier Sirello, Bruno Tissot and Juan Pablo Cova, Irving Fisher Committee on Central Bank Statistics, Bank for International Settlements, 28 October 2026 (PDF)
- MaRCo@BIS: AI-agentic econometric analysis for central bankers, Marco Jacopo Lombardi, Taejin Park and Fernando Perez-Cruz, BIS Working Paper 1395, 12 November 2026
- Big Data and Artificial Intelligence as Inputs into Monetary Policymaking at the National Bank of Ukraine, Volodymyr Lepushynskyi et al., Visnyk of the National Bank of Ukraine, Occasional Papers 2026/01, 6 November 2026
- How accurate are AI models at nowcasting GDP?, Czech National Bank, čnBlog, 13 November 2026
