The finance ministry is giving the grunt work of election analysis to AI. The design is worth copying.

Since 2007 the Ministry of Finance has costed the election promises of every major party before each Riigikogu election. The work is slow and manual: read each platform, pull out every promise, sort it by policy area, check whether it already sits in a development plan or a budget, then put a price on it. For the spring 2027 election, ERR reported on 7 October, the ministry will hand the first part of that chain to an AI tool.

Raoul Lättemäe, who heads the ministry's fiscal policy department, described the split to ERR. The tool breaks platforms into individual promises, assigns each to a policy area and searches development plans, budgets and the ministry's own earlier analyses to find where the topic has already been dealt with. That is the preparatory work. The costing and the verdict on whether a promise is feasible stay with the analysts, who review and add to everything the tool produces.

What the ministry got right

Three design choices stand out, and none of them is specific to politics.

• ✳The AI does the part that is tedious and checkable: decomposition, classification and retrieval. People do the part that needs judgement and carries the signature.
• ✳The tool answers from an indexed set of documents and cites them. Lättemäe told ERR that a connected AI would point to the specific development plan, promise or platform where a topic appears, instead of answering from its training data alone.
• ✳The scope is limited on purpose. Only parties that field a full list of candidates get analysed, because otherwise the workload becomes unmanageable, AI or not.

There is a fourth point that matters more in Estonia than elsewhere. The tool is built for Estonian-language text and has to find the right document even when the question is phrased differently from the source. Anyone who has tried to search a folder of Estonian contracts with an off-the-shelf tool knows why that requirement is in the specification.

The money and the timeline

The project is one of seven that the Ministry of Justice and Digital Affairs funded in August from its support measure for AI solutions in the public sector. The finance ministry received 119,000 euros. The ministry's own competition page puts the expected project cost at up to 140,000 euros and invited research institutions and other capable teams to apply by 7 October. One team will be chosen to build it. The analysis itself will be published and presented before the election.

If the build succeeds, the ministry plans a public search tool on the same index. Journalists, researchers or anyone else could connect their own AI assistant to it and get answers that cite development plans, state budgets, the previous five promise analyses and the current platforms. That is a sensible way to make public data usable by AI without pretending the data lives inside the model.

The same measure funded six other projects

The August round is a useful catalogue of what Estonian public bodies think AI is for right now. The agricultural registers agency is building a workflow that checks subsidy applications for completeness and legal compliance before they are submitted. The geological survey is building an open-source retrieval system over its archive of survey reports. The transport authority is prototyping automatic analysis of free-text customer feedback from several channels. The national library is testing open and local language models for describing digital heritage. In total the first round gave out 1,323,862 euros and left 676,138 euros in the pot, with applications accepted on a rolling basis until the budget runs out.

Notice the pattern. Validate incoming documents, search a large archive, summarise feedback. None of these is a chatbot. All of them are a specific step in an existing process where the input is text and the output is checked by a person.

What an Estonian company should take from this

The ministry's measure is for public bodies, so a private company cannot apply to it. We wrote in August about the EIS grants that private companies can use for AI projects. The design, though, transfers directly. If your team reads tenders, contracts, audit reports, customer complaints or supplier documents, you have the same problem the finance ministry has: a lot of text, a small number of people, and a result someone has to sign.

• ✳Pick one document-heavy process and write down the steps. Mark which steps are decomposition, sorting or lookup, and which ones need a decision. Give only the first group to AI.
• ✳Build on your own documents, not on the model's memory. Index the contracts, plans or reports the answer should come from, and require every answer to cite the page it came from.
• ✳Test with Estonian text from the start. A tool that works on an English demo and fails on your Estonian archive has not been tested.
• ✳Limit the scope before you start, the way the ministry limited itself to full candidate lists. A narrow tool that finishes beats a wide one that does not.
• ✳Keep the signature with a person. The ministry publishes the analysis under its own name and its analysts check every line. Your contract reviewer or tender manager should do the same.

The finance ministry has been doing this analysis by hand for five election cycles. It did not replace the analysts. It took the part of their week that nobody will miss and gave it to a machine, with a citation on every answer. That is the version of AI adoption most Estonian companies can start next week.