Data from ISORA confirms that artificial intelligence has gone from being perhaps an experimental curiosity to a central piece of tax modernization in less than a decade.

The Tax Administration 2025 document, recently published by the OECD, shows that around 69% of the administrations evaluated already use artificial intelligence and another 24% are in the implementation phase, with cases ranging from advanced data analysis to virtual assistants and services based on natural language.

Behind many of these projects, there is a common pattern: strong analytical capabilities supported by cloud services, which allow for scaling, better cost control, and rapid experimentation.

We would like to highlight here some practical cases reported in the document that we believe contribute to leaps in efficiency within tax administrations.

Internal analytics and machine learning platforms

In Norway, the administration has created a cloud-based data and analytics platform that provides analytics services, data science/machine learning capabilities, dashboards, and reports for multiple units. This same platform is used to train machine learning models and begin exploring the management of unstructured documents.

In Australia, the Australian TAX Office (ATO) has implemented the Advanced Analytics Platform in the Cloud (AAP Cloud), hosted in a private cloud with secure access to tax data. It relies on open-source services, which reduce costs and provide technological flexibility, and leverages cloud elasticity to scale computing power when needed. The models running on this platform are used to identify potential non-compliance and better manage fraud in contexts where risk patterns change quickly.

Japan demonstrates how, on top of these capabilities, predictive models can be built and applied to revenue collection: the NTA uses AI to choose the most effective channel to contact delinquent taxpayers and to predict the time of day with the highest likelihood of response, achieving significant increases in contact rates. All of this would be much more expensive and more rigid without cloud infrastructures that allow paying only for what is used and adapting resources to processing peaks.

Large Language Model (LLM) [1] to improve developers’ work

The Swedish Tax Agency is developing an AI-based code assistant to support its developers and analysts. This assistant automates repetitive tasks such as code generation and debugging, speeds up development cycles, and reduces human error, freeing up time to solve complex problems and improving the quality of the systems that support tax collection and taxpayer services.

The logic is supported by contextual suggestions, code completion, and assistance in identifying and correcting errors. Additionally, integrating these assistants into cloud-hosted development environments makes it easier for teams to scale tests, integrations, and deployments with greater agility.

LLM to analyze, classify, and assign queries

Ireland offers a truly clear example of the use of generative AI to better manage the flow of inquiries. According to the document, on the MyEnquiries portal, taxpayers used to choose the wrong categories from drop-down menus; many inquiries arrived as “general” or misclassified. Today, AI models and natural language processing analyze the free text, classify the inquiry with 97% accuracy, and automatically route it to the appropriate expert, reducing the average referral time by more than 24 hours.

In Canada, the CRA collected more than 90,000 user comments on its website and used generative AI to summarize and group them into key themes. Based on these analyses, ten areas for improvement were prioritized, resulting in a 158% increase in user success in registration and login tasks. This type of use—summarizing, classifying, and detecting patterns in large volumes of text—is at the heart of what LLMs enable today, and becomes especially attractive when data and models can be processed economically and securely in the cloud.

LLM to support the taxpayer service staff

Several examples show how AI does not replace human agents but rather empowers them. Singapore developed IRASearch, a tool based on LLM and GenAI that acts as an assistant for officials when responding to inquiries. IRASearch refines searches, helps find accurate information, and uses generative AI to propose draft responses and explain complex tax concepts in plain language, tailored to the taxpayer’s specific case.

In Sweden, another assistant in development automates the generation of responses to emails about taxes and registrations, allowing for a higher volume of queries to be managed without sacrificing quality, and with continuous feedback to improve responses.

France, for its part, has created Caradoc, an assistant that combines GenAI and advanced document search to help officials to find relevant regulations, guidelines, and procedures more quickly. The system allows users to upload individual documents or entire collections and ask complex questions; the model responds and highlights the text fragments used, providing transparency and traceability. This solution is closely aligned with modern “LLM + cloud search” architectures.

Virtual assistants that interact directly with taxpayers

At the most visible end for citizens are the assistants who speak directly with taxpayers. Austria has created the Federation of Bots, a network of specialized chatbots (e.g., one for taxes and others for general government services) that share knowledge with each other. Initial results indicate greater efficiency, less burden on support teams, and high user acceptance.

Korea went one step further with its AI Tax Helpline: a voice-recognition-based telephone service and virtual assistant that answers queries, sends links, guides, and videos during the call, and allows taxpayers to go directly to the tax return section. In 2024, 98% of calls were successfully managed (compared to 26% the previous year), which is equivalent, in terms of service capacity, to hiring a thousand additional employees.

We believe there is a trend toward using artificial intelligence, both machine learning and deep learning to incorporate classification and regression mechanisms into processes, and generative AI, particularly LLMs aimed, at least initially, at improving the work of the tax agents, reducing timeframes, multiplying civil servants’ capabilities, and achieving significant increases in efficiency.

The growth reported in ISORA by the countries themselves leads us to believe that AI initiatives will increase in the near future. And with that comes the growing challenge of managing so much power without crashing on the curves. Perhaps that is why the Spanish Tax Agency has established a specific methodology to support the development of its AI projects, strengthening its governance by including in its analysis, in addition to data sources, the treatment of possible biases, the implications for stakeholders, the monitoring of results, and its retraining. How about that?

Best regards and good luck.

References:

[1] LLM Large Language Models are generative AI instances that include well-known systems such as ChatGPT (OpenAI), Claude (Anthropic), Copilot (Microsoft), DeepSeek LLM (DeepSeek), Gemini (Google), Llama (Meta), Mistral Large (Mistral AI), Grok (xAI), among others.

11 total views, 11 views today