Artificial intelligence is gradually becoming embedded in the everyday operations of central banks. What was once largely confined to statistical experimentation and specialised technological projects is becoming part of their analytical systems, supervisory capabilities and information management processes. This transformation promises substantial efficiency gains for institutions whose effectiveness increasingly depends on their ability to interpret vast quantities of economic and financial data.

It also introduces a new contradiction. Central banks, whose institutional independence is a cornerstone of monetary credibility, may become increasingly reliant on private technology providers to perform some of their functions. Their decisions remain legally autonomous, but the systems contributing to those decisions may depend on infrastructure, software and models whose design and operation lie largely beyond their direct control.

The report AI Readiness in Central Banks, published on 28 October 2026 by the Irving Fisher Committee, associated with the Bank for International Settlements (BIS), provides a particularly revealing picture of this transition. Conducted in early 2026 across 65 jurisdictions and institutions, the survey builds on research undertaken in 2024 and introduces an essential distinction between declared readiness for artificial intelligence and the capabilities actually established to deploy it.

This distinction moves the discussion beyond statements of intent. An institution may regard AI as strategically important without possessing the data, expertise, infrastructure or governance mechanisms required to deploy it responsibly. Conversely, it may introduce operational applications before fully establishing the rules governing their use.

It is precisely within this gap between ambition, adoption and control that new institutional risks are emerging.

I. From Experimentation to Production

The survey figures point to a significant acceleration. Nearly two-thirds of respondents now identify artificial intelligence as a strategic priority, compared with fewer than half in the 2024 survey. The proportion of institutions with an organisation-wide AI strategy has reached 38%, while another 31% are developing one. Only 8% report having no plans to establish such a strategy.

The shift is equally visible in deployment. More than two-thirds of reported AI projects are now in production, including applications operating on a limited scale or within specific departments, compared with approximately half in 2024.

These figures do not mean that two-thirds of central banks have automated their principal functions. They refer to reported projects rather than the entirety of institutional processes. The distinction matters: a document-processing application used daily may qualify as operational without playing any role in monetary policy decisions.

The significance lies elsewhere. Central banks are beginning to incorporate AI into established workflows rather than evaluating it exclusively in experimental environments. As a result, the requirements are changing. It is no longer sufficient for a model to produce promising results during testing. It must operate reliably over time, comply with security requirements, preserve auditable records and withstand technical failures.

Moving into production therefore transforms a question of innovation into one of governance.

It also exposes a difference in institutional timescales. Technological development advances according to the evolution of models, infrastructure and suppliers, while public institutions adapt through budgetary procedures, specialised recruitment, internal controls and, occasionally, regulatory revisions.

Central banks must consequently integrate systems whose development often proceeds faster than their own organisational cycles.

II. Adoption Strategies: Different Institutional Models

Central banks do not possess identical resources or operate under identical constraints. Their artificial intelligence strategies depend on the size of their teams, the maturity of their information systems, their supervisory responsibilities, their access to computing infrastructure and their ability to recruit specialised personnel.

Four broad strategic architectures can be distinguished. These are analytical models rather than a statistical classification of the 65 survey respondents.

The first relies on acquiring commercial solutions. Institutions use existing applications, potentially adapted to their requirements, to shorten deployment times. This approach can deliver rapid productivity gains, particularly in document retrieval, assisted drafting, data processing and textual analysis. It also exposes institutions to suppliers' pricing policies, technological modifications and contractual conditions.

The second strategy prioritises internal development. Central banks build or adapt their own models, retain control over critical infrastructure and directly manage processing rules. This approach offers greater scope for auditing and customisation but requires scarce expertise, recurring investment and long-term maintenance capabilities. It does not necessarily eliminate dependence on external software components, processors or development tools.

The third architecture combines internal resources with external services. Sensitive data and selected applications remain within controlled environments, while commercial capabilities are used for less critical tasks. This hybrid strategy seeks to balance efficiency, security and autonomy. Its effectiveness nevertheless depends on the quality of interfaces, the separation of operating environments and the practical ability to replace one supplier with another.

Finally, institutional cooperation offers a fourth approach. Central banks can share standards, tools, software libraries, evaluation methodologies and selected research capabilities. Such cooperation reduces fixed costs and limits the fragmentation of resources. It can also create new interdependencies if shared infrastructure becomes a single point of failure.

None of these strategies guarantees complete autonomy. The appropriate choice depends on the criticality of the application and the level of control required. A commercial solution may be entirely suitable for an administrative task while being inappropriate, under the same conditions, for an essential financial surveillance function.

The strategic question is therefore not whether central banks should use external technologies, but which functions they can reasonably outsource without compromising their ability to exercise control.

III. The Data Paradox: Deploying Before Being Fully Prepared

One of the survey's most significant findings concerns data readiness. While artificial intelligence projects are multiplying, 52% of the institutions surveyed still consider themselves to be at an early stage of data preparation. Approximately 27% report ongoing work, while only 17% are approaching completion.

These figures suggest that operational AI adoption is advancing more rapidly than the consolidation of certain information infrastructures. They do not establish that existing applications are operating on inadequate data, but they reveal a tension between the pace of experimentation and the transformation of information management systems.

For a central bank, data quality is not a secondary technical consideration. Monetary statistics, prudential information, payment data, macroeconomic series and market indicators rely on different definitions, frequencies and collection methodologies.

Artificial intelligence can facilitate their integration, identify inconsistencies and accelerate their use. It can also amplify errors when sources are incomplete, metadata are insufficient or statistical transformations are poorly documented.

A powerful model applied to a poorly structured database does not necessarily produce better knowledge. It may simply generate a misleading representation of reality more quickly.

This difficulty is particularly important for data subject to revision. Estimates of growth, employment and inflation evolve as new information becomes available. A system that combines data from different publication dates may introduce information that was unavailable when a historical forecast would originally have been produced.

This phenomenon, known as temporal data leakage, can create an artificial impression of predictive accuracy.

Data readiness therefore requires more than centralisation. It demands rules governing provenance, version control, classification, quality and retention. Models must be evaluated using information genuinely available at the relevant point in time, rather than datasets reconstructed retrospectively.

As AI becomes operational, data governance consequently moves closer to the core of monetary governance.

IV. Applications: From Document Assistance to Systemic Surveillance

Artificial intelligence applications within central banks involve substantially different levels of risk.

Document-related applications constitute one category. They enable institutions to search extensive collections, compare regulatory texts, summarise publications, translate documents and assist software development. Their value lies primarily in time savings and improved access to institutional knowledge.

A second category concerns statistical production. Machine learning techniques can contribute to anomaly detection, information classification, data preparation and the exploitation of unstructured information. They can also improve short-term forecasting tools when conventional indicators become available only after a delay.

A third area involves prudential and financial supervision. AI can help analyse banking reports, identify concentrations of risk, reconcile heterogeneous information and flag developments requiring human investigation. Its use can expand the analytical coverage of supervisory authorities without implying that supervisory decisions are delegated to algorithms.

Finally, some applications may contribute more directly to macroeconomic research and policy assessment. They can explore nonlinear relationships, compare scenarios and process larger datasets than those traditionally used in conventional models.

However, improved predictive capabilities should not be confused with an automatic improvement in causal understanding. A model capable of accurately forecasting an indicator does not necessarily explain why that indicator changes.

Monetary policy is not simply about anticipating events. It requires an assessment of economic mechanisms, the effects of interventions and the possible consequences of decisions taken under uncertainty.

Artificial intelligence can enrich the available information without replacing the requirements of economic reasoning.

This distinction becomes essential when AI systems begin influencing the analyses presented to monetary policy committees. Even without automated decision-making, models can direct analysts' attention, select signals considered important and alter which scenarios appear plausible.

Algorithmic influence can therefore emerge well before any formal delegation of decision-making authority.

V. Technological Concentration: Dependence at Multiple Levels

Technological sovereignty cannot be reduced to the nationality of a supplier. It depends on market structure, infrastructure location, contractual rights, access to critical components and an institution's ability to continue operating when a service is disrupted.

The artificial intelligence value chain comprises several layers: semiconductors, data centres, cloud services, foundation models, development platforms and specialised applications.

A central bank may diversify its applications while retaining a common dependence on a single infrastructure provider. It may also use several model suppliers whose services ultimately rely on the same cloud platform.

Apparent diversification does not therefore guarantee genuine independence.

This concentration first creates operational risk. A prolonged interruption of an external service can render multiple applications unavailable simultaneously. The impact depends on the criticality of the functions involved, the duration of the disruption and the existence of alternative procedures.

It also creates economic risk. Computing, storage, data transfer and model usage costs can change rapidly. A solution that initially appears attractive may become more expensive as usage increases and institutional processes adapt around it.

A third risk concerns technological lock-in. When applications are built around proprietary interfaces, specialised formats or services that are difficult to reproduce, switching suppliers may require substantial system redevelopment.

The cost of exit can consequently become as important a source of dependence as the initial acquisition price.

Legal considerations introduce another complication. The physical location of data, the laws applicable to a service provider, disclosure obligations, access to technical logs and audit rights do not necessarily coincide.

Infrastructure hosted within national territory is not automatically insulated from every form of foreign legal obligation.

It would nevertheless be excessive to equate every technological dependence with a loss of sovereignty. Central banks have long relied on software, equipment and networks supplied by private companies.

What is new is the potential depth of AI integration into analytical processes, together with the concentration of certain technological capabilities.

Sovereignty must therefore be assessed through operational criteria: auditability, control over data, contractual reversibility, service continuity and the retention of sufficient internal expertise to challenge or replace the results produced.

VI. Monetary Independence and Algorithmic Intermediation

Central bank independence is generally examined through legal mandates, the appointment of senior officials, financial resources and protection from political interference.

Artificial intelligence introduces an additional dimension: autonomy over the infrastructure through which institutional knowledge is produced.

An institution may enjoy complete legal independence while using systems whose performance, operating rules or conditions of access depend on decisions made by external companies.

This does not mean technology suppliers determine monetary policy. Rather, they may indirectly influence the instruments used to prepare decisions.

A model update, a modification of its parameters, a change in filtering rules or a deterioration in performance can affect an application's results.

If analysts lack adequate validation and comparison procedures, such changes may be interpreted as developments in economic data rather than alterations in the analytical tool itself.

The issue becomes particularly sensitive when model outputs feed dashboards, economic assessments or surveillance systems.

Central banks must therefore preserve their ability to reproduce analyses, identify model versions and compare conclusions against independent methodologies.

This requirement suggests three distinct forms of autonomy.

Legal autonomy protects decision-making authority. Operational autonomy ensures the continuity of institutional functions. Analytical autonomy allows institutions to understand, verify and challenge the instruments contributing to their judgement.

The third does not replace the first two, but it may increasingly become a practical condition for preserving them.

VII. The Systemic Risks of Shared Artificial Intelligence

The adoption of AI by central banks forms part of a broader transformation of the financial system. Commercial banks, insurers, asset managers, market infrastructures and supervisory authorities are also developing algorithmic tools.

This diffusion creates new interdependencies.

When several institutions use similar models, trained on comparable information or supplied through the same platforms, their analyses may become more homogeneous.

Such convergence is not necessarily problematic under normal conditions. It may even facilitate the collective identification of certain risks.

During periods of crisis, however, it could amplify simultaneous reactions if systems lead different actors to prioritise the same signals or responses.

The risk is particularly relevant in activities where algorithmic decisions can trigger rapid adjustments in liquidity, portfolios or hedging positions. Correlated behaviour could then reinforce market movements.

For central banks, the problem has two dimensions. They must monitor risks arising from AI use by financial institutions while assessing their own technological dependencies.

A failure affecting a common provider could consequently disrupt supervised institutions and some of the tools used by their supervisors at the same time.

This is a risk scenario, not an outcome established by the 2026 survey. Its significance will depend on actual system architecture, model diversity and the availability of continuity mechanisms.

Nevertheless, the possibility supports an expansion of resilience testing.

Traditional cybersecurity and business continuity exercises could be supplemented with scenarios involving the unavailability of an AI provider, corrupted analytical outputs, manipulated input data or an inability to reconstruct a model-assisted assessment.

The objective is not to eliminate every technological risk, but to prevent a single failure from spreading across several essential functions.

VIII. Budgets, Skills and Economic Trade-offs

Artificial intelligence also changes the expenditure structure of central banks.

Visible costs include licences, computing services, storage infrastructure, maintenance contracts and development teams.

Yet the full cost of an AI system also includes data preparation, audits, security controls, testing, documentation, training and continuity mechanisms.

An application that is inexpensive to experiment with can become costly when required to meet the standards of a monetary institution.

Budgetary decisions must therefore extend beyond immediate comparisons between internal development and commercial subscriptions.

They should incorporate expenditure over several years, switching costs, investments in reversibility and the consequences of service interruptions.

Skills are equally important.

Central banks need specialists capable of developing models, but also economists, statisticians, lawyers, cybersecurity professionals and auditors sufficiently familiar with AI to evaluate its applications.

Recruiting a small number of engineers does not, by itself, create institutional control over artificial intelligence.

Dependence can become cognitive as well as technical. When an organisation gradually loses the ability to reproduce analyses generated by external providers, it becomes less capable of challenging their conclusions or changing methodologies.

Preserving internal expertise should therefore be treated as a strategic investment.

It maintains institutional knowledge, enables comparisons between available solutions and strengthens negotiating capacity by providing a genuine understanding of suppliers' constraints.

From this perspective, AI budgets should not be assessed solely through productivity gains. They should also reflect an institution's ability to preserve essential functions without exclusive dependence on an external actor.

IX. Growing Asymmetries Between Central Banks

The transition towards artificial intelligence may reinforce existing differences between institutions.

Larger central banks generally possess greater financial resources, research capabilities and recruitment networks.

They can experiment with several technologies, maintain specialised teams and develop infrastructure adapted to their needs.

Smaller institutions face different constraints. Fixed investments may represent a larger share of their budgets, while competition for specialised talent can limit their ability to establish permanent teams.

These asymmetries do not automatically determine strategic quality.

A smaller institution may achieve excellent results by concentrating resources on carefully selected applications, while a larger organisation may accumulate poorly coordinated projects.

They nevertheless influence the range of available choices.

For central banks in emerging economies, access to commercial solutions can accelerate statistical modernisation and strengthen certain supervisory capabilities.

At the same time, such strategies may increase exposure to foreign suppliers, fluctuations in computing costs and restrictions affecting cross-border data transfers.

Language presents another challenge.

Models developed primarily from internationally available datasets may perform unevenly across languages, regulatory terminology and institutional contexts.

A strategy adapted to emerging economies should therefore prioritise national data quality, control over critical applications, the development of local expertise and cooperation between institutions.

Technological sovereignty does not necessarily require recreating the entire global AI industrial value chain.

It can instead rest on the selective ability to control essential applications, protect sensitive data and maintain credible alternatives.

X. Towards Risk-Based Governance

The multiplication of projects makes risk classification indispensable.

An application summarising public documents does not present the same challenges as a system processing confidential prudential information.

Similarly, an internal programming assistant should not be subject to identical controls as a system whose outputs contribute to monitoring a payment infrastructure.

Effective governance should therefore classify applications according to their criticality, degree of automation, data sensitivity and potential consequences.

Lower-risk applications can benefit from relatively flexible experimentation procedures, subject to basic security and confidentiality requirements.

Applications involved in sensitive processes require stronger validation, comprehensive documentation and clearly defined human oversight.

For the most critical uses, several safeguards become necessary: reproducible results, comparisons with benchmark methods, continuous performance monitoring, model version control and the ability to deactivate a system without materially disrupting the relevant institutional function.

Institutional accountability must remain identifiable.

When a system produces an incorrect recommendation, the organisation should be able to determine whether the error originated in the data, the model, its integration or human interpretation.

Governance focused exclusively on formal compliance would be insufficient.

Risks evolve as models and operating conditions change. Control mechanisms must therefore be continuous and proportionate to the importance of the functions involved.

Technological independence is established less through declarations of principle than through verifiable capabilities for control, auditing and substitution.

XI. Three Possible Trajectories

The developments observed between 2024 and 2026 suggest several possible institutional trajectories.

Under the first scenario, central banks progressively integrate AI within controlled environments.

They strengthen data capabilities, diversify suppliers, develop internal expertise and adapt governance mechanisms.

Productivity gains are accompanied by improvements in operational resilience.

Under a second scenario, adoption continues to advance more rapidly than institutional control.

Applications proliferate, infrastructure remains fragmented and contractual dependencies accumulate.

Institutions benefit from immediate efficiency gains but face increasing difficulties auditing their systems, controlling costs and organising replacements.

A third scenario would involve stronger institutional cooperation.

Central banks could pool selected tools, develop common standards and share evaluation methodologies.

Such an approach could reduce resource asymmetries and improve collective negotiating capabilities in dealings with technology providers.

These scenarios are neither mutually exclusive nor predictive.

Different combinations may emerge depending on the institution and the functions concerned.

The eventual trajectory will depend in particular on whether current experiments can be transformed into durable infrastructure without allowing deployment objectives to consistently override governance requirements.

XII. Measuring Genuine AI Maturity

Future surveys would benefit from moving beyond the number of adopted strategies and the proportion of projects in production.

Maturity could be assessed across several complementary dimensions: the share of critical applications independently audited, the practical ability to change suppliers, the proportion of sensitive data processed within controlled environments, the coverage of continuity testing and the existence of human validation procedures.

Budgets should also be examined according to their composition.

An increase in expenditure on external services does not carry the same implications as investment in internal expertise, data quality or infrastructure resilience.

Supplier concentration deserves particular attention.

It would be useful to distinguish the apparent diversity of applications from the actual diversity of the infrastructure supporting them.

Finally, model performance should be evaluated under operational conditions, including changes in economic regimes, statistical revisions and market disruptions.

A central bank genuinely prepared for artificial intelligence is not necessarily the institution using the largest number of applications.

It is the institution that understands the conditions under which its systems operate, recognises their limitations and knows how to continue fulfilling its mandate when those systems become unavailable or unreliable.

The Irving Fisher Committee survey marks an important stage in understanding the technological transformation of central banks.

Artificial intelligence is no longer a distant prospect. It is entering everyday operations, shaping institutional strategies and introducing new trade-offs involving resources, data and governance.

The figures nevertheless reveal an incomplete transition.

The expansion of operational projects coexists with data preparation that remains insufficient or unfinished across much of the surveyed population.

Strategic ambitions are advancing faster than some organisational capabilities.

This does not undermine the case for adopting artificial intelligence.

It demonstrates that productivity gains cannot be the sole measure of success in a transformation of this magnitude.

For central banks, the issue extends beyond administrative efficiency.

It concerns the quality of the knowledge used to conduct monetary policy, the resilience of supervisory systems and the capacity of public institutions to retain control over their essential functions.

Technological dependence is not new in the history of central banking.

Artificial intelligence, however, gives it a distinctive significance by becoming involved not only in the execution of tasks but also in the production and organisation of the information that informs decisions.

Central banks were established to preserve monetary credibility and protect economies against certain forms of instability.

They must now learn to govern a technological infrastructure whose suppliers, costs and conditions of access can evolve independently of their mandates.

Monetary independence will continue to be defined by law and institutional arrangements.

Its practical preservation will also depend, increasingly, on the ability to understand and control the machines contributing to economic judgement.

The question is therefore no longer simply how extensively central banks will adopt artificial intelligence, but how far they can do so without transferring, alongside their technological tools, part of their operational autonomy.

Principal Sources

  • Schmidt, Rafael; Sirello, Olivier; Tissot, Bruno; Cova, Juan Pablo (2026). AI Readiness in Central Banks. Irving Fisher Committee, IFC Report No. 19, Bank for International Settlements, 28 October 2026.
  • Godoy de Araujo, Douglas Kiarelly et al. (2025). Governance and Implementation of Artificial Intelligence in Central Banks. IFC Report No. 18, BIS, 28 April 2025.
  • Bank for International Settlements (2025). Governance of AI Adoption in Central Banks. Consultative Group on Risk Management, 29 January 2025.
  • Bank for International Settlements (2025). The Use of Artificial Intelligence for Policy Purposes. Report prepared for the G20.
  • Financial Stability Institute (2026). In Data We Trust? Emerging Policy and Supervisory Approaches to AI Data Use in Financial Services. FSI Insights No. 73, 26 March 2026.
  • Financial Stability Institute (2026). When Machines Attack: Frontier AI Cyber Threats and Policy Responses in the Financial Sector. FSI Paper No. 28.
  • Bank for International Settlements (2026). AI and the Global Economy: Implications for Central Banks. BIS Bulletin No. 130, 28 July 2026.

Methodological note: The 2026 survey statistics are reproduced from the figures supplied for this report. The discussions of technological sovereignty, strategic models, institutional scenarios and maturity indicators constitute independent editorial analysis rather than statistical conclusions directly established by the survey. Self-reported findings alone cannot establish the actual effectiveness of deployed applications or the degree of technological dependence of individual institutions.