Valtiontalouden tarkastusvirasto VTV

Central government has no comprehensive and up-to-date overall picture of how, and how widely, AI is used (p.20). The AI situational review put before the ministerial working group in December 2025 'does not, in the audit office's assessment, give a realistic overall picture of the situation in public administration': it consisted largely of a list of individual measures whose weight in the whole remains unclear. A monitoring review of the digital compass implementation plan, compiled for the same working group a month earlier, recorded that the measure 'use of AI in public administration' was progressing 'entirely as planned', but offered 'only very sparse justifications' for that assessment. The ministry that compiled the AI review has itself recognised the need and intends to develop the picture during 2026. A Monitoring weakness of a particular kind: the reporting exists, and what it mainly establishes is that reporting took place. The audit identified positive examples of two ministries promoting AI within their respective administrative sectors. (p.21)

• Function: Governance
• Value: Domain knowledge
• Stakeholders: Policy setter
• Quality: Reliable, integrated information base Functioning oversight and governance

Cross-administrative coordination belongs to a standing interministerial group with four working groups of its own, and two separate AI coordination projects were set up alongside it, one by each of two ministries (pp.17-19). The projects had an agreed division of labour, yet 'seen from outside they may nevertheless have appeared to overlap, since forming the overall and situational picture belonged to the tasks of both' (p.19). In the survey nearly three quarters of respondents were dissatisfied with the clarity of the division of labour, roles and responsibilities between authorities, and with whether central government has working procedures for leading and coordinating development across administrative branches (pp.18-19); only a little over a quarter thought the steering supported the development of shared solutions. The group carrying the coordination duty does not decide on the allocation of resources, and the government programme's own commitment to pool digital project funding into a joint development budget has not been carried out in practice. A Coordination weakness in which responsibility is assigned everywhere and the means are held somewhere else.

• Function: Governance, Organizational structure
• Stakeholders: Policy setter
• Quality: Avoidance of duplication and fragmentation Functioning oversight and governance

Both the ministry's investment programme and the innovation fund's calls require applicants to show that the solution can be replicated and scaled for other organisations' use. 'The problem, however, is that central government and the wellbeing services counties have no ready operating models or structures for developing, productising, procuring or sharing jointly used AI solutions, which emerged in the audit as a clear shortcoming' (p.30). The audit office 'considers it peculiar that organisations applying for funding must be able to set out in their applications how the solutions would be replicable, even though central government has no ready structures and operating models for this' (p.31), and holds that the bodies coordinating the calls should first describe centrally what conditions they offer for productising shared solutions, who is responsible for them, and how developing organisations may use them. A shared customer-service solution explored under one agency's lead had clear benefit potential, but how it would in practice be made available to interested organisations was unresolved (p.30). A Design weakness in the funding instrument itself: a condition was written that the system provides no way to satisfy.

• Function: Governance, Planning
• Stakeholders: Funder
• Quality: Avoidance of duplication and fragmentation Cost control and value for money

Both funders require applicants to describe the benefits of their project, with economic and productivity benefits to the fore. 'Central government has, however, no common assessment criteria for benefits, so every organisation describes and assesses them in its own way. In addition it is probable that funding applications rather overestimate than underestimate the attainable benefits, which may to a large extent be calculated and not necessarily directly realisable as savings, for example' (p.29). Different kinds of use case are hard to compare because their benefits are formed by different logics: AI-assisted interpreting has an alternative cost that can be computed fairly reliably, whereas a use case said to save an expert half an hour a day yields no economic benefit at all unless the content or organisation of the work is changed as well (p.29). Benefits may also accrue to organisations other than the developer (p.30). A Benefit-management weakness: the figure that decides the allocation is the one nobody has defined how to produce.

• Function: Finance, Planning
• Stakeholders: Funder
• Quality: Reliable, integrated information base Cost control and value for money

The message 'that point-wise development should give way to a comprehensive reform of processes with the help of AI, because only in that way are significant benefits obtained' recurred throughout the audit (p.5). What the evidence shows is the opposite: a great many experiments are under way - the wellbeing services counties alone listed some 200 development items in November 2025 - but 'both the experiments and the use cases taken further appear point-wise, and AI use has not yet reached a comprehensive reform of processes' (p.28). Organisations began with the clearest and easiest cases, and the challenge is that the pilots do not remain temporary but reach production use (p.27). Interviewees held that the productivity leap AI makes possible does not arise from people's own experiments or from point-wise development, but from redesigning operations and processes, while understanding how AI changes human work; realising the benefits ultimately demands successful change management (p.28).

• Function: Work process design, Strategy
• Stakeholders: Management
• Quality: Streamlined, standardized processes Appropriate use of technology and automation

Legislation and the differing interpretations of it emerged as a central obstacle: over 80 percent of respondents had identified provisions that block or restrict AI use (p.21). Where interpretation is uncertain, that 'is reflected in development as the precautionary principle, so that development targets containing uncertain interpretations are easily left unimplemented' (p.22); authorities 'have begun using AI on subjects that are legally clear, and have willingly left use cases containing possible interpretation ambiguities waiting' (p.24). The ministry responsible for the general legislation on public administration opened a two-year preliminary study into the legal conditions for AI in administrative procedure only in December 2025; interviewees called the study indispensable but said it should have been carried out earlier or faster, while the ministry explained that it knew of no concrete use cases in which AI was to be used in decision-making and had therefore not considered the study urgent (pp.23-24). The audit office 'considers questionable the premise that legislative change needs could be examined only at the stage when there is knowledge of use cases that challenge the law in force' (p.24). A Guidance weakness with a clear direction of travel: the clarification waits for the practice, and the practice waits for the clarification.

• Function: Regulations, Governance
• Value: Regulatory system
• Stakeholders: Lawmaker, Policy setter
• Quality: Clear objectives and goal-setting Appropriate use of technology and automation

The role of the central ICT service centre and of the digital agency 'in developing and promoting AI in central government is to some extent unclear', and the ministry that steers them 'had no clear view either of what the role of these organisations is in promoting AI use in central government' (p.31). The service centre 'still lacks the AI platform services the agencies need'; respondents saw it as a natural enabler of shared solutions from the infrastructure, platform and information security angle, but 'it has no clear mandate, according to respondents or to itself, to act as the leader of AI development', nor sufficient development capability, its activity being concentrated on basic IT services (p.39). Its customers may nevertheless experience it as a bottleneck (p.31). Organisations fill the gap by procuring platform services themselves; a joint procurement framework for robotic process automation and AI ran from 2019 to 2025 without being used at scale, and because AI purchases can also be made inside other frameworks the central purchasing body has no view of what agencies have bought (p.39). A Responsibility weakness: a capability that exists institutionally, with no one authorised to use it.

• Function: IT, Outsourcing, Product/service purchase
• Stakeholders: Delivery partner, Supplier
• Quality: Avoidance of duplication and fragmentation Adequate resources and competences

Several authorities have produced models and support materials for AI use, made for different purposes and not forming a coherent whole; using them requires each organisation to apply and combine them on its own initiative, and the existing guidance 'emphasises risk management and compliance, while giving less support on questions of productivity, scalability, operational change or leadership' (p.35). The insufficiency of shared models and instructions came through in most of the 36 open answers on the subject (p.36). 'On the evidence of the audit material the support materials often remain at too general a level. Organisations want concrete, applicable models and examples' (p.36), and they are not felt to resolve the interpretation problems around the rules. 'Effective information steering should reduce differences of interpretation between organisations, but in this it has not yet succeeded' (p.37). The mutual relations of the materials are nowhere described, and no information is available about their maintenance or update schedules (p.36). A Guidance weakness whose cost falls unevenly: the present materials demand substantial interpretation of the rules from each organisation, which brings differences in competence and resources into play.

• Function: Governance
• Stakeholders: Policy setter
• Quality: Clear objectives and goal-setting Adequate resources and competences

A sufficient supply of good-quality, interoperable data is the basic precondition for AI use; in high-risk AI systems this is a statutory requirement, and in any AI system the effect of errors or bias 'multiplies rapidly and leads systematically to erroneous or discriminatory outcomes' (p.40). Only 40 percent of respondents regarded their own procedures for assuring the quality of the data their AI solutions use as adequate, and only 12 percent agreed that central government has shared operating models and methods for ensuring it (p.40). Interviewees gave examples of poor data quality preventing AI use outright: AI could not yet be used for processing state aid 'because of deficient and fragmented data' (p.42). Large volumes of data are still delivered as unstructured attachments in many file formats, metadata is often a weakness, terminologies differ between systems, and data on the same matter may sit in different systems in different forms across several decades (p.42). Organisations therefore start from the use cases that need no particular existing dataset, but wide use requires a data infrastructure that is in order (p.40). An Information-base weakness underneath everything else: the technology is ahead of the material it has to work on.

• Function: IT
• Value: Assets
• Quality: Reliable, integrated information base Interoperability that spares users redundant effort

Almost all organisations have trained staff in some way and about 90 percent agreed that measures had been taken, but the respondents found what has been done insufficient: only a third considered their staff's AI literacy adequate (pp.43-44). General online courses, webinars and support materials are plentiful, but 'they do not form a coherent whole', and their content and examples do not connect to the organisation's own work, so they do not necessarily advance the adoption of AI (p.44). Only about a fifth agreed that enough shared training matching their organisation's needs is on offer (p.44). One respondent put the gap concretely: 'in developing AI competence we have to get away from the level of beware of hallucination and offer real literacy and usage skills' (p.45). Models and services for identifying role- and task-specific competence requirements and for building study paths are as yet few, even though management, lawyers, buyers, developers and experts were all named as roles that need them (p.45). A Competence weakness with a distributional edge: rapidly developing technology makes competence hard to maintain and experts hard to find, and small organisations are least able to do either.

• Function: Human resources
• Value: Human capital
• Stakeholders: Staff, Management
• Quality: Adequate resources and competences Adequately staffed, skilled workforce