Questions about Artificial Intelligence in the Tax Agency

According to Spanish regulations (RD 817/2023), an AI system is a system designed to operate with a certain autonomy and that, based on data provided by machines or people, infers how to achieve objectives based on machine learning, logic and knowledge, ultimately generating content, predictions, recommendations or decisions that can influence the environment in which these AI systems operate.

In contrast, systems that use rules defined solely by people to execute operations automatically are not AI. Therefore, massive data processing, network or graph analysis, risk analysis systems and the robotization or automation of certain actions are not AI, provided that these systems operate in a deterministic manner, based on rules set by humans, and without using the predictive or generative capabilities of AI.

The National Artificial Intelligence Strategy (ENIA) includes among its strategic objectives the incorporation of AI as a factor for improving productivity and efficiency in the Administration and the operation of public services, as well as in the interfaces of relations with companies and citizens. This is also reflected in the Digital Plan drawn up by the Ministry for Digital Transformation and the Civil Service, in the context of the promotion of AI in the public sector included in Lever 5 of the ENIA.

The Tax Agency seeks to use AI to increase the effectiveness and efficiency of tax and customs actions, prioritizing its use in terms of information and assistance to taxpayers and in the area of prevention.

The use of AI will help to promote and simplify compliance with tax obligations and will improve the quality of data declared by taxpayers. In this way, many verification procedures that originate from errors made when filing can be avoided, thereby also avoiding possible adjustments. The use of AI will help to tackle these problems at their source, seeking to prevent non-filing of declarations during the voluntary period, and promoting the possibility of solving deficiencies in declarations when they are submitted.

The Tax Agency has provided itself with a strategy for the use of AI, as well as a methodology for its practical application.

The aim is to ensure that the development of AI in the Tax Agency takes place within a framework of security and protection of taxpayers' rights, with strict respect for the current legal system and the principles and values that govern all actions of the Tax Agency.

First of all, it is an AI that is always supervised by people (this is the 'human in the loop' principle). It is people who will review, question and validate or not the recommendations, suggestions or predictions proposed by the AI system.

People will be present at all stages of AI development and use, from the first moment of data collection and verification of its quality; Next, in the information processing phase, the choice of AI techniques to be used in each project, the training process and the quality of the system after training; and, finally, in monitoring and evaluating the operation of each project. Ultimately, people will control and be responsible for the operation of the AI system as a whole and in each of its phases.

The AI system at the Tax Agency has a second objective, that of being respectful of people ('human-centric' approach), applying ethical and transparency principles in the use of AI. For example, if the Agency uses AI to interact directly with citizens, it will expressly report this.

The methodology for the practical application of AI in specific Agency projects has the essential function of ensuring the quality and governance of these projects. Because it is a methodology that covers the entire project life cycle.

Initially, the methodology will outline the specification of requirements, the risk analysis and the cybersecurity and data protection measures to be considered, a feasibility analysis, design decisions, validation of results and implementation into production.

From that moment on, the same methodology will serve as a guide for monitoring, evaluating and maintaining the project that is already operating in the production environment.

In each phase of the life cycle of each project, the actors who should participate are identified, the roles they should play and the functions they should perform at each moment. Similarly, the result to be obtained in each of the stages and based on each project is determined.

And throughout this process, the participation of the competent units of the Tax Agency in matters of security and control is ensured, as well as possible external audits to be carried out by the Spanish Agency for Supervision of Artificial Intelligence (AESIA).

Any AI system must comply with EU and national data protection regulations regarding the collection and processing of information.

The Tax Agency will verify that the data sources are reliable and contain current, updated and unbiased information.

Next, extreme care will be taken in cleaning and processing the data after it has been collected, before it is finally validated by humans.

During the development of each project, it will also be verified that the data is complete and represents, in the appropriate proportion, the entire universe on which the AI system will act, that it is accurate and does not contain errors, and that it has the required temporal validity.

Likewise, work will be done to prevent, detect and correct any biases that may be contained in the data, also guaranteeing the privacy and security of the data used, especially when they may involve personal information.

The Tax Agency has already been taking advantage of the opportunities offered by technological advances in AI to apply them to better develop its mission, which consists of promoting voluntary compliance with tax obligations. Current examples of use in this regard are chatbots (virtual assistants), some of the IRPF warning messages for possible errors when modifying information on work income offered by Renta Web, or the warnings offered to potential non-IRPF filers during the voluntary filing period.

The automated administrative actions issued by the Tax Agency will in no case be based exclusively on the results obtained from an AI system. Under the 'human in the loop' principle, human intervention is always guaranteed, supervising and validating or not the decisions proposed by the system.

Limitations on the use of AI will be determined by the legal-regulatory framework applicable at any given time, under the supervision of AESIA, as well as by the principles of proper use of AI and the Tax Agency's own values.

There is currently no decisive use of AI in control procedures. There are indeed cases of use in specific management procedures initiated at the request of the taxpayer itself and in which predictive analyses based on AI are only allowed to have any effective relevance if their proposal is favourable to the taxpayer.

Therefore, there are no AI techniques being used for the processing or instruction of inspection or limited verification procedures (managers).

As regards the taxpayer selection phase, there are mass analysis tools, which are not AI, and which help officials in the risk analysis tasks prior to control procedures, their use being limited to the automated processing of data which, in any case, must be assessed by officials together with the rest of the information obtained, before making a decision regarding an inspection verification action.