Publication date: November 02, 2026
A growing number of businesses are investing in artificial intelligence and expect to recover the cost of those projects through the research and development (R&D) tax relief. The Polish tax authorities and the Supreme Administrative Court, however, draw a clear line: integrating an off-the-shelf AI model or implementing a commercially available system – however advanced or “self-learning” – is not R&D activity. The relief is available only to those who create new technology themselves and bear genuine research risk. We explain how the tax authorities’ “two-step test” works, what the Frascati Manual and the case law of the Supreme Administrative Court say, which costs in IT projects qualify and which are traps – and how to document the work so that the deduction survives a tax audit.
In the era of digitalisation, artificial intelligence (AI) is becoming an integral part of how modern businesses operate. Deploying advanced algorithms, automating processes and building intelligent analytical systems all involve substantial investment. It is natural for businesses to look for ways to optimise that expenditure by claiming the research and development (R&D) tax relief. The practice of the tax authorities and the case law of the administrative courts show, however, that the line between an innovative implementation and qualifying R&D activity is a fine one, and crossing it requires meeting strict statutory criteria.
The R&D relief, governed by Article 18d of the Corporate Income Tax Act[1] (the “CIT Act”) and, correspondingly, Article 26e of the Personal Income Tax Act[2] (the “PIT Act”), allows specific categories of expenditure, known as eligible costs, to be deducted from the tax base. Although the mechanism is almost identical in both statutes – with minor adjustments such as the right of an “enterprise in succession” to continue deductions under the PIT Act – the key to claiming the relief lies in a proper understanding of the statutory definition of R&D activity itself.
Under Article 4a(26) of the CIT Act, research and development activity means creative activity comprising scientific research or development work, undertaken in a systematic manner in order to increase the stock of knowledge and to use that stock of knowledge to devise new applications. For an AI project to be covered by the relief, it must therefore cumulatively be creative in nature, be conducted methodically according to an established schedule, and generate new knowledge or lead to a new application of knowledge. The requirement to “increase the stock of knowledge” does not mean that the knowledge must be new on a global scale. According to the explanatory notes issued by the Minister of Finance, it is sufficient to develop a new solution or a new application of knowledge from the perspective of the enterprise, provided that the work is genuinely creative and involves real technological uncertainty[3].
When verifying entitlement to the deduction, the tax authorities apply a rigorous mechanism that may be described as a “two-step test”. First, they assess the nature of the work itself, examining whether it substantively meets the statutory definition of R&D. Only after this stage has been passed does the second step open up: financial qualification, i.e. verification of whether the expenditure incurred falls within the closed catalogue of eligible costs.
When interpreting these provisions, the Polish tax authorities and courts refer to the guidelines contained in the OECD’s Frascati Manual[4], which is widely used as an interpretative standard for assessing R&D activity. The Manual sets out five criteria that R&D activity must satisfy cumulatively. The work must be novel, i.e. aimed at findings that go beyond the existing knowledge in the relevant industry. It must be creative, based on original concepts rather than mechanical replication. The third, critically important element is uncertainty – at the outset of the project, the final outcome, the time required and the costs must be difficult to predict because of technological risk. The activity as a whole must be systematic (planned and budgeted), and its results – including negative ones – must be reproducible or transferable thanks to appropriate codification and documentation.
For a business’s tax security, it is fundamental to separate the implementation of ready-made solutions from genuine R&D work. Implementing commercially available tools such as ERP systems, SaaS platforms, RFID technology or pre-built AI models consists solely in adapting them to the individual needs of the enterprise. Such activities are not creative in nature and do not generate new resources of objective technical knowledge, and the risk involved is operational rather than research-related.
One can speak of R&D activity in the field of artificial intelligence only where the business acts as the creator of the technology rather than its user. Qualifying work will therefore include designing and training proprietary network architectures from scratch, modifying the core of algorithms to achieve properties previously unknown on the market, or building prototypes in an experimental environment to verify unique hypotheses. It should be stressed, however, that the use of off-the-shelf AI models does not automatically rule out the R&D relief. Where a business makes significant modifications to the architecture of such a model, develops its own training methods or devises new optimisation mechanisms, such work may well satisfy the criteria of R&D activity. The Director of the National Tax Information Office has already addressed AI-based projects directly in individual tax rulings, assessing them against the same criteria that apply to “conventional” IT projects[5].
An example of an implementation that does not qualify a business for the R&D relief is purchasing API access to a ready-made large language model and integrating it with the company’s own customer database in order to build a simple website chatbot. The work is then limited to parameterisation, adjusting the user interface (UX/UI) and preparing prompts – there is no technological uncertainty, only the operational risk of whether the implementation will be completed on time and within budget.
An example of creation, by contrast, is a situation in which a business uses an open-source AI model but its engineers substantially alter the model’s architecture, design a unique data-weighting algorithm to minimise so-called hallucinations in a highly specialised medical field, and run a series of experiments without knowing at the outset whether the model will achieve the precision required by the market. Here the risk is research-related: the project may fail, and yet its result – knowledge of which approaches do not work – still constitutes new technical knowledge.
The line of case law of the Supreme Administrative Court (Naczelny Sąd Administracyjny, NSA) on the classification of IT implementations is unequivocal. Particularly instructive are the twin judgments of 25 August 2021, II FSK 138/19[6] and II FSK 139/19[7], in which the NSA examined the implementation of modern document workflow, warehouse management and sales systems. The Court firmly held that adapting existing, commercially available software to the needs of the taxpayer’s business does not constitute R&D activity, as it is not a creative process that expands the stock of knowledge.
The same issue was addressed in the earlier judgment of 9 October 2019, II FSK 3589/17[8], concerning the implementation of a highly advanced, “learning” APS-class production planning system. The taxpayer argued that the system made decisions autonomously and learned the specifics of the enterprise. The Court nevertheless held that purchasing even the most innovative, self-learning technology from a third party and configuring it over many months is still merely an implementation. The conclusion from the case law is clear: a purchased innovation does not turn the purchaser into an entity conducting R&D activity. The administrative courts consistently distinguish between acquiring and configuring ready-made solutions and independently creating new functionalities that actually lead to the creation of new technical knowledge. This is of particular importance in the context of AI, where the “intelligence” of the tool itself is often mistakenly equated with the innovativeness of the work performed by the taxpayer.
A positive substantive assessment of the project does not end the challenges. The next area of risk is the categorisation of expenditure. In practice, the most significant item for businesses claiming the relief is personnel costs. These cover the remuneration of employees under employment contracts and of persons performing R&D work under contracts of mandate or contracts for specific work, together with employer-funded social security contributions. Such expenditure is deducted in proportion to the time actually spent on R&D activity. It is worth remembering that these costs are deductible at 200%, which makes them the most valuable category of eligible costs[9]. Another important category is depreciation write-offs on fixed assets and intangible assets used in R&D activity.
It should also be borne in mind that the purchase of software cannot be treated as the acquisition of materials and raw materials within the meaning of the statute, since it concerns rights in intangible property. As regards the cost of cloud services (renting computing power), SaaS subscriptions and API access to AI models, the possibility of including them in the R&D relief is, as a rule, limited. In each case, this requires an analysis of the nature of the expenditure and the way it is recognised in the accounting books. Software may constitute an eligible cost essentially where the business acquires the economic copyright or a licence to it, recognises it in its records as an intangible asset and includes the depreciation write-offs in the relief. This is particularly relevant in AI projects, where the cost of computing power for model training is often one of the largest budget items – and, at the same time, one of the most difficult to claim under the relief.
Projects that make use of artificial intelligence open the way to the R&D relief provided that the business genuinely co-creates new technology, generates new knowledge through research experiments and bears real technological risk in doing so. The burden of demonstrating these circumstances lies with the taxpayer, so the documentation should be produced as the project progresses, not merely in response to a request from the tax authority.
To protect their interests, businesses should adopt several project management principles. First, the conceptual, uncertainty-laden research phase must be clearly separated from subsequent implementation and maintenance work – ideally at the level of the project structure and budget. Second, precise records of staff working time must be kept so that only the man-hours devoted to creative work are included in eligible costs. Third, the whole process must be supported by codification of the research process in line with the standard described in the Frascati Manual: from formulating the research hypothesis and describing the state of knowledge at the outset, through analysing the technological uncertainty and documenting the experiments performed (including unsuccessful ones), to a final report recording the newly acquired knowledge. In AI projects, versioned code repositories, model training logs and benchmark test protocols have particular evidential value, as they make it possible to demonstrate that the outcome was not known in advance.
Where there is doubt as to the classification of a specific project, it is worth considering applying for an individual tax ruling. A precise description of the facts, emphasising the element of uncertainty and the team’s creative contribution, is then crucial – it is precisely on this ground that most disputes over the R&D relief in IT projects are decided.
[1]Act of 15 February 1992 on Corporate Income Tax (consolidated text: Journal of Laws of 2026, item 554).
[2]Act of 26 July 1991 on Personal Income Tax (consolidated text: Journal of Laws of 2026, item 592).
[3]Tax explanatory notes of the Minister of Finance of 15 July 2019 on the preferential taxation of income generated by intellectual property rights (IP Box), paras 32–45 (definition of research and development activity common to the R&D relief and the IP Box regime).
[4]OECD (2015), Frascati Manual 2015: Guidelines for Collecting and Reporting Data on Research and Experimental Development, The Measurement of Scientific, Technological and Innovation Activities, OECD Publishing, Paris, https://doi.org/10.1787/9789264239012-en; Polish edition: OECD/GUS (2018), Warsaw.
[5]See, inter alia, individual tax rulings of the Director of the National Tax Information Office of 4 March 2024, 0111-KDIB1-3.4010.742.2023.2.JKU; of 25 April 2025, 0114-KDIP3-1.4011.236.2025.2.EC; and of 13 March 2026, 0111-KDIB1-3.4010.7.2026.3.AN.
[6]Judgment of the Supreme Administrative Court of 25 August 2021, II FSK 138/19, LEX no. 3265262.
[7]Judgment of the Supreme Administrative Court of 25 August 2021, II FSK 139/19, LEX no. 3252790.
[8]Judgment of the Supreme Administrative Court of 9 October 2019, II FSK 3589/17, LEX no. 2740353.
[9]Article 18d(7)(3) of the CIT Act and Article 26e(7)(3) of the PIT Act.