Croatia drew an AI plan to 2032. The budget needs an outcome ledger

THE SIGNAL IN ONE SENTENCE

A strategy can point toward 2032. A citizen still has to deal with Tuesday. Croatia has now placed those two timescales beside each other. On October 16, Minister of Justice, Public Administration and Digital Transformation Damir Habijan presented the country's National AI Development Plan through 2032 and an accompanying action plan through 2028. The government says the pair should connect Croatian knowledge and technical capacity with European cooperation, speed responsible AI adoption across the economy and public sector, build skills, improve infrastructure and data, protect rights, and help domestic research and business ideas become products and services. The direction is broad by design. Earlier government material organized the plan around four objectives: human potential; infrastructure and data; ethical and lawful use; and innovation and the economy. Priority areas included public administration, healthcare, agriculture and the environment, energy, transport and tourism. A February presentation said 128 experts from 55 institutions were participating in the work. The newly presented plan also points toward practical public-service uses. The government named an assistant in the e-Citizens system and tools that could support judicial officials. Croatia already has experience with AI-assisted public information and court-document work, so this is not a plan written on a perfectly blank page. That is the good news. The harder news is that a national plan can collect worthy nouns faster than a ministry can produce better outcomes. Skills. Trust. Infrastructure. Innovation. Sovereignty. Competitiveness. Every one matters. None tells a resident whether a permit took less time, a farmer received better advice, a patient was routed safely, a small business reached affordable compute or a person could challenge a machine-assisted public decision. The plain signal is that Croatia needs an outcome ledger. Not another glossy dashboard with green arrows. A public, project-level record that connects every funded AI measure to a real owner, a baseline, a budget, a delivery date, an affected population, a measurable service outcome and a correction path. The action plan through 2028 is the natural place to start. Each measure should appear as one row in the ledger. The row should name the responsible ministry or agency, the delivery partner, the legal authority, the amount committed, the amount spent, the service being changed and the result the public is supposed to experience. If the measure funds training, the ledger should show who completed it, what capability was assessed and whether the skill was later used. Attendance is not an outcome. A room full of certificates may still leave an agency unable to evaluate a vendor's model, document an incident or explain an automated recommendation. If the measure funds infrastructure, the ledger should show who can actually use it. Compute capacity is not public value while it remains a headline number. Can a university team book it? Can a small company outside Zagreb reach it without a procurement department? Are public-interest projects priced differently? What is the waiting time, energy use and regional distribution? If the measure funds a public-service assistant, the ledger should start with the service problem rather than the model. How many people abandon the process today? How long does it take to find the right form? Which languages and accessibility needs must be supported? How often does the assistant send a person to the wrong service? How quickly can a civil servant correct an answer? The e-Citizens example makes the need concrete. An assistant can help someone navigate passports, residence changes, benefits or other life events. It can also provide a confident answer based on stale guidance. A useful measure is not the number of questions answered. It is the share of people who reached the correct service, completed the task and did not have to repair an avoidable error. The ledger should publish answer-quality samples, correction time, escalation rates and the date on which the underlying government guidance was last checked. It should make clear when the assistant is providing information and when it is helping initiate an official transaction. Justice requires an even higher bar. Croatian officials have discussed AI support for organizing judgments, summarizing large case files and helping people understand whether prior case law may be relevant to a dispute. Those uses are not equivalent. Document organization can reduce clerical burden. A summary can omit the one paragraph that matters. A prediction about litigation can shape whether someone seeks a remedy at all. The ledger should therefore label each judicial use by consequence. It should record the source set, the date range, the error tests, the human review requirement and the groups likely to be affected. It should report contested outputs and corrections without exposing confidential case information. A tool that assists a judge or court employee needs a different control surface from a public information chatbot. The same discipline belongs in the economic programme. The government wants to turn research and entrepreneurial ideas into useful products and services. That is a reasonable ambition for a smaller European economy. The danger is measuring activity instead of additional value. A grant awarded is activity. A pilot announced is activity. A conference badge is certainly activity, and the badge printer will be thrilled. Outcomes are harder. Did a supported company reach paying customers? Did a research result become a maintained tool? Did a traditional manufacturer improve quality or reduce waste? Did a tourism deployment improve access without turning a town into a surveillance experiment? Did public funding create a reusable capability inside Croatia, or merely purchase a temporary layer from an outside supplier? Every business-support entry should include the public contribution, private co-investment, procurement route, ownership of resulting intellectual property, interoperability requirements and an exit plan. If a vendor disappears or changes terms, the service should not become a historical reenactment of good intentions. Regional distribution matters too. National AI policy tends to pool around capitals, universities and already successful firms. Croatia's ledger should show which counties receive money, infrastructure, training and pilots. It should separate the location of a contractor from the location of the people receiving the benefit. That does not mean forcing identical projects into every region. It means making concentration visible and explaining it. The ethical and lawful objective needs operational measures of its own. A statement that a system complies with the EU AI Act is not a substitute for evidence. The ledger should link to impact assessments where publication is lawful, identify the risk category, name the human decision-maker, state the data-retention period and provide the route for complaint or appeal. For higher-impact systems, it should report false positive and false negative rates across relevant groups, accessibility testing, security incidents, overrides and service interruptions. If data cannot be published, the agency should explain the limitation and identify the independent body that reviewed it. Public reporting also protects good projects from political weather. A programme with a clear baseline, owner and outcome can survive a change of minister more easily than a promise attached to one press conference. A failed pilot can still produce value if the evidence is public and the lesson changes the next procurement. Failure should be a ledger state, not a scandal-management strategy. Croatia's plan deserves credit for tying a long horizon to a nearer action plan. The government also appears to understand that skills, infrastructure, rights and innovation must move together. An AI economy without skills imports judgment. Public services without infrastructure stall. Infrastructure without rights becomes a very efficient way to make a bad decision. Now the programme needs connective tissue between the categories. For every measure, publish the money, owner, baseline, target, evidence date, affected community, incident record and appeal route. Update the row even when the news is awkward. Let researchers, businesses, civil servants and residents see which investments are producing capability and which are producing slide decks. The strategy can keep the long view. The ledger should answer the Tuesday question: what changed for the person standing in front of the service today?