A tax adviser handling twenty or thirty clients a day cannot spend half a working day tracking whether the Directorate General of Taxes has published a binding ruling that affects the case in front of them. The AEAT Tax Doctrine Search Engine holds tens of thousands of resolutions; locating them, reading them and extracting the applicable criterion requires time and expertise that, under daily pressure, often simply does not exist. Artificial intelligence changes that equation: with a well-implemented system, the advisory firm can retrieve relevant resolutions in seconds, obtain an executive summary of the current criterion, and identify any contradictory doctrine across different rulings. This article explains how it works, what its limitations are, and what any firm should require before deploying it.
What are AEAT binding rulings and why do they matter
Binding rulings (Article 89 of the General Tax Act) are resolutions issued by the Directorate General of Taxes (DGT) in response to specific questions from taxpayers. They are binding on the Tax Administration when the applicant follows the criterion received, and they serve as doctrinal reference for analogous cases. In practice, they are the most reliable source of tax interpretation before adopting a complex fiscal position.
The problem is their volume. The DGT publishes several thousand resolutions per year. The public database of the AEAT Doctrine Search Engine (petete.tributos.hacienda.gob.es) contains more than 90,000 rulings accumulated since the 1990s. Finding those that apply to a specific case — by tax type, by taxable event, by sector, by date — requires legal expertise and patience. AI can provide speed and coverage; the legal judgement remains with the professional.
How an AI system for querying DGT doctrine works
The technical architecture underlying these systems is called RAG (Retrieval-Augmented Generation). The basic scheme has three layers:
- Semantic indexing: the binding rulings published by the DGT are processed and stored in a vector database. Each resolution is converted into a numerical representation that captures its meaning, not just its exact words.
- Similarity retrieval: when the professional types a question in natural language («Is VAT on catering services for employees deductible when the company has its own canteen?»), the system searches for the most semantically similar resolutions, not just those containing those exact words.
- Synthesis with traceability: the language model drafts a summary of the criterion contained in the retrieved resolutions, always citing the ruling number (V-XXXX-XX) and the year of issue, so the professional can verify the original source.
This same approach can be applied to other complementary doctrinal sources: resolutions of the Central Economic-Administrative Court (TEAC), Supreme Court judgements on tax matters, AEAT circulars, and even the firm's own internal documentation (reports, internal criteria, agreements with the tax inspectorate). The result is a tax copilot that searches all those sources simultaneously and presents the consolidated criterion with precise references.
If you want to see how Summum IA deploys this type of assistant in professional firms, you can review our Copilot for tax and legal firms service, where we detail the deployment process and technical requirements.
Specific use cases in a tax advisory firm
1. Quick resolution of queries on routine transactions
The professional receives an invoice from a client with an unusual transaction — a property swap between private individuals, a debt assignment, a related-party transaction between shareholders — and needs to know whether there is DGT doctrine on the treatment for personal income tax or corporation tax. Instead of manually navigating the Tributos search engine with several different terms, they type the question in natural language and within thirty seconds have a summary with the three or four most relevant resolutions from the past five years, ranked by proximity to the case.
2. Criterion change alert
The DGT has no obligation to notify when it changes its criterion from one ruling to the next. A well-configured AI system can automatically compare new doctrine published each week against what the firm already has indexed, and flag rulings where the criterion varies. This is particularly relevant in areas such as deductible expenses in the personal income tax of self-employed individuals, where doctrine has evolved significantly in recent years.
3. Preparation of complex tax returns
In corporation tax returns with significant extra-accounting adjustments (related-party transactions, accelerated depreciation, R&D&I deductions), the system can retrieve the rulings that justify each adjustment and generate a doctrinal support sheet that is archived alongside the client's file. In the event of a tax audit, the firm can demonstrate that it acted with due diligence.
4. Internal team training
Junior technicians at the firm can use the system to learn how to argue fiscal positions with doctrinal backing. The copilot does not just provide the answer: it explains the line of reasoning the DGT follows in its resolutions, which accelerates learning of the most common interpretation criteria.
Comparison: manual AEAT search vs. AI system
| Criterion | AEAT search engine (manual) | AI system (RAG on doctrine) |
|---|---|---|
| Average search time | 15–45 minutes per case | 30–90 seconds |
| Source coverage | One source at a time (AEAT, TEAC, TS…) | All indexed sources in parallel |
| Detection of criterion changes | Manual, dependent on the professional | Automatable with weekly alerts |
| Source traceability | High (professional reads the original) | High if the system cites the ruling number |
| Risk of search bias | Medium (depends on terms used) | Low (semantic search, not literal) |
| Cost per search | Professional's time (~20–60 €) | Fraction of a cent (model API) |
| Doctrinal base update | Instant (public source) | Schedulable (weekly/monthly reindexing) |
What AI can do and what it cannot do
It is essential that the firm understands the limits before deploying any solution. AI does not make tax decisions: it retrieves and synthesises existing doctrine, but the responsibility for the fiscal position remains with the professional who signs it. That said, the real risks are manageable if the system is configured correctly:
- Hallucination (the model inventing a ruling that does not exist): mitigated by requiring the system to synthesise only from the documents it has actually retrieved, never «from memory». A correct RAG design guarantees this.
- Outdated doctrine: if the vector database is not reindexed regularly, the most recent rulings will not be available. A good system includes an automated periodic update process from the public source.
- Out-of-context interpretation: a binding ruling is specific to the case consulted; the professional must assess whether their client's facts are genuinely analogous. AI can flag this, but it does not replace legal judgement.
From the perspective of the European AI Act (EU Regulation 2024/1689, applicable from August 2026 in its main provisions), AI systems that assist in legal-tax interpretation fall, in principle, outside the «high-risk» categories regulated by Annex III, provided they are not the ones making the final decision. Responsibility lies with the professional who signs.
Technical and data requirements for deploying it in your firm
A correct deployment in a mid-sized advisory firm (five to twenty professionals) needs to address the following points:
- Data source: the AEAT Doctrine Search Engine is publicly accessible, but bulk downloading for internal use must comply with the terms of use published on the electronic headquarters (sede.agenciatributaria.gob.es). There are providers who already commercialise normalised databases of Spanish tax jurisprudence and doctrine with usage licences.
- Privacy of firm data: if the system is to include internal documentation (reports, client files), it must be guaranteed that data does not leave the corporate perimeter. This can be achieved with models running on the firm's own server, or with appropriate data processor agreements with the cloud provider.
- Integration with management software: the productivity gain multiplies when the copilot is invoked directly from the tax ERP (A3, Sage Despachos, Wolters Kluwer a3asesor…) without switching to a separate application.
- Traceability and auditing: each system response must be logged with the professional's query, the retrieved resolutions and the synthesised text. This protects the firm against claims and facilitates internal quality review.
At Summum IA we have been accompanying professional firms and SMEs in technology adoption since 2007. Our Copilot for firms service includes the design of the RAG architecture, the selection and ingestion of doctrinal sources, integration with existing management software, and team training.
Frequently asked questions
Can the AI system access in real time the new rulings published by the DGT?
Yes, provided an automatic update process is configured to periodically download and index new resolutions from the AEAT Doctrine Search Engine. The usual frequency is weekly or fortnightly, which is sufficient given that the DGT does not publish resolutions daily. The firm must ensure that its provider includes that refresh process in the system maintenance.
Can the AI system incorrectly cite a binding ruling that does not exist?
This is the risk known as «hallucination». In a well-built RAG system, the model only synthesises from documents it has actually retrieved from the indexed database; it does not generate references from memory. If the system is configured correctly, each reference to a ruling (V-XXXX-XX) corresponds to a real document that the professional can verify. The risk is not zero, but it is technically controllable and far lower than with general-purpose models that respond without access to a verified documentary base.
Does a firm using an AI copilot assume greater liability before the tax authorities?
Not in legal terms. Responsibility for the fiscal position belongs to the professional who signs it and the taxpayer who adopts it. The use of a decision-support system neither removes nor increases that responsibility. In fact, the documented use of a system that evidences the search for DGT doctrine before adopting a position can be a favourable argument in an audit procedure, as it demonstrates the firm's due diligence.
Is this type of system compatible with GDPR if we handle client data?
It depends on the architecture. If the copilot only accesses the public doctrinal base (DGT rulings, TEAC resolutions, Supreme Court judgements) without cross-referencing personal client data, the data processing involved is minimal. If the system also indexes internal firm documents containing client personal data, it must be guaranteed that processing takes place within EU or EEA territory, with a data processing agreement signed with the technology provider, and with appropriate security measures pursuant to Article 32 of the GDPR. Summum Consultoría can advise you on the applicable obligations.