A Spoonful of Tar in the AI Barrel

This report analyses the fundamental discrepancy between the promises of a "tenfold increase in productivity" and the actual data pointing to gains of around 10%, examines the instability of the business model of industry leaders, and points to sovereign, European and open-source alternatives that ensure data security and cost predictability.

For chief technology officers (CTOs) in small and medium-sized enterprises (SMEs) and non-governmental organisations (NGOs), the key task is to separate the aggressive marketing of corporate vendors from real operational value.

Real Productivity in the Age of AI

The central point of the marketing narrative around artificial intelligence is the promise of radically accelerating intellectual work. However, independent studies conducted in 2024–2026 cast a shadow over these claims. An analysis carried out by the DX organisation on a group of 121,000 developers at more than 450 companies shows that despite the widespread adoption of AI assistants (used by 93% of developers), the real productivity gain measured by Pull Request (PR) throughput has plateaued at just 10%.

This "productivity plateau" stands in stark contradiction to the claims of vendors such as GitHub or Google, which in their own studies declare time savings of up to 55%. The cause of this phenomenon is that optimistic reports omit system-level costs. While generating code syntax itself is faster, the time spent reviewing, debugging and integrating machine-generated code has grown disproportionately. Senior developers now spend an average of 4.3 minutes analysing a single AI suggestion, whereas reviewing human-written code takes them 1.2 minutes.

Engineering analysis points to Amdahl's law as a fundamental barrier to the mythical 10x growth. If writing code accounts for only 20–30% of a developer's total working time (the rest being design, meetings, testing and documentation), then even infinitely speeding up the coding phase allows for a maximum overall process improvement of around 22–30%. The mathematical model of process acceleration ($S$) can be described by the equation:

where $p$ is the fraction of time affected by AI, and $s$ is the acceleration factor for that fraction. With $p = 0.25$ (25% of time spent coding) and $s = 10$ (tenfold acceleration of writing), the resulting $S$ is only 1.29, i.e. a 29% overall improvement, which is still an optimistic result that does not account for the growth of technical debt.

In organisations struggling with internal dysfunctions, AI does not fix processes but exposes their flaws, acting as a catalyst for chaos. The "perception gap" phenomenon is particularly dangerous here: 69% of developers declare that they feel faster, even though objective data indicates a slowdown in their work. This stems from the reduction of cognitive effort on routine tasks, which creates a false sense of efficiency, while actual project progress is hampered by growing complexity and logic errors generated by the models.

Intelligence...? The ARC-AGI v3 Benchmark

A critical approach to AI requires understanding that current language models (LLMs) do not possess intelligence in the sense of the ability to reason and generalise to unfamiliar tasks. François Chollet, a distinguished researcher and creator of the ARC-AGI benchmark, defines intelligence as the efficiency of acquiring new skills. Most popular tests (e.g. MMLU) measure only "crystallised intelligence" – that is, knowledge that could have been memorised during training on billions of web pages.

The ARC-AGI (Abstraction and Reasoning Corpus) benchmark was designed to exclude the possibility of using memory and statistical pattern matching. It consists of geometric tasks that are intuitive for children but require "fluid intelligence" – the ability to discover rules from just a few examples.

Data from March 2026 shows that despite enormous spending on scaling models, AI systems still fail at tasks requiring the discovery of new symbolic transformations. AI's successes in other fields are often the result of so-called "scaffolding" – complex instructions and external mechanisms built by humans that mask the model's lack of genuine cognitive autonomy. For CTOs in SMEs, this means that AI will perform excellently in repetitive and well-documented tasks, but will fail in situations unique to a given organisation, where there are no ready-made patterns in the training data.

The Economics of AI in a Bubble

The AI sector in 2025 and 2026 is characterised by an extreme disproportion between investment and profits. Global venture capital spending on AI reached 258.7 billion dollars in 2025, accounting for 61% of the total value of VC investments worldwide. At the same time, leading players are recording record operating losses. OpenAI, despite reaching 20-25 billion dollars in annual recurring revenue (ARR) in 2025/2026, forecasts a loss of 14 billion dollars in 2026 alone.

The subscription-based economic model (SaaS), with current inference (response generation) costs, is deeply deficit-ridden. It is estimated that OpenAI spends an average of 3.30 USD for every 1.00 USD of revenue, which means a loss of 2.30 USD on every interaction, even for paying users.

The market is in a state of an "AI subprime crisis", where startups building on the APIs of large corporations hand over all the capital they raise to them, and the model providers in turn transfer these funds to cloud providers (Microsoft, Amazon, Google) and chip makers (Nvidia). OpenAI has declared infrastructure spending of around 1.4 trillion dollars over the next eight years, financing it mainly with debt, without a clear path to profitability before 2029-2030.

Analysts, including Ray Dalio of Bridgewater Associates, point to a striking similarity between the current situation and the dot-com bubble at the turn of the century. For smaller organisations, this means a high risk of "vendor lock-in", which at any moment could drastically raise prices or change the terms of service to improve margins. A CTO's strategy should therefore assume diversification and the use of open-weight models that can be run on their own infrastructure.

The Hidden Costs of AI: Cognitive Debt and Code Degradation

The greatest long-term threat to organisations implementing AI is growing technical debt. Research indicates that AI-assisted code is of lower quality in every measurable respect: it lacks structure, contains duplicates and often ignores the specific context of the project.

A new phenomenon is emerging – Comprehension Debt. This is the future cost an organisation will incur to understand, fix and secure code that was machine-generated and that no one on the team fully understands. Unlike traditional technical debt, which is the result of a developer's conscious compromise ("I'll do it quickly, I'll fix it later"), AI debt accumulates invisibly. The developer is often unaware that the AI applied a faulty pattern or a dangerous workaround.

Data from the GitClear platform confirms these concerns:

• The amount of "copy-paste" code rose from 8.3% to 12.3% between 2020 and 2024.
• The amount of code undergoing healthy refactoring (improving structure) fell by 39.9%.
• The percentage of code requiring correction within just two weeks of being written is constantly growing.

For SMEs, this means that the initial time savings on writing code can be multiplied away by later failures and the inability to scale the system. The solution is to change Code Review processes: developers must be assessed not on the amount of code delivered, but on its readability and their understanding of what the AI generated on their behalf.

How to Approach This Wisely?

Mistral and the European AI Ecosystem

In response to the dominance of American corporations, the European AI market has developed an alternative based on transparency and data sovereignty. France's Mistral AI has become a symbol of this approach, offering "open-weights" models that allow organisations to host locally, customise and maintain full control over infrastructure.

The key advantages of Mistral models for SMEs and NGOs are:

• Data privacy: Sensitive information about donors, beneficiaries or clients does not leave local servers or a private cloud within the EU.
• Cost efficiency: Models such as Mistral 7B or Mixtral are optimised for performance, allowing them to be run on standard hardware without the need to invest in GPU clusters costing millions of euros.
• GDPR compliance: Using European models makes it easier to meet stringent personal data protection requirements, eliminating the risks associated with transferring data to the USA (e.g. under the US Cloud Act).

Alongside Mistral AI, Germany's Aleph Alpha (the PhariaAI model) positions itself as a leader in "Explainable AI", which is critical for NGOs and public institutions requiring full transparency of decision-making processes. This ecosystem is complemented by the Hugging Face platform, which serves as a central repository for thousands of open-source models (e.g. Llama 3, Command R+), enabling easy testing and deployment of solutions tailored to specific market niches without tying oneself to a single corporate ecosystem.

n8n as the Foundation of Automation

In the area of process automation, instead of the Microsoft-promoted Copilot tool, which ties an organisation to the Azure and Office 365 ecosystem, the SME and NGO sector is increasingly turning to n8n. It is a "fair-code" platform that combines the simplicity of a visual interface (low-code) with the power of scripting in JavaScript or Python.

For non-governmental organisations, n8n offers a unique opportunity to automate the "boredom audit" (the Boreout phenomenon) – that is, the elimination of the most tedious, repetitive administrative tasks. Examples of implementations include:

1. Donor management: Automatic synchronisation of data from web forms and payment gateways (e.g. Stripe, PayPal) directly with CiviCRM or Google Sheets, along with sending personalised thank-you messages.
2. Smart mail assistant: Using a local AI model (e.g. via n8n and Ollama) to read incoming e-mails, categorise them (e.g. urgent, offer, request for help) and generate summaries for the board.
3. Reporting and analysis: Automatic retrieval of data from Google Search Console and GA4, trend analysis by AI and generation of weekly reports in PDF format sent to Slack or e-mail.

Using n8n in the self-hosted model (e.g. on a Docker server) allows an NGO to maintain full technological sovereignty while drastically lowering operating costs compared to commercial solutions, which can cost an average company from 10,000 to 20,000 PLN per year.

Financing Transformation with Dig.IT

For Polish enterprises in the SME sector, a key support instrument in 2025–2026 is the Dig.IT – Digital Transformation of Polish SMEs programme, run by the Industrial Development Agency (ARP). This programme offers grants ranging from 150,000 PLN to 850,000 PLN for digitalisation-related projects, including AI and automation implementations.

To qualify for support, an enterprise must meet a series of rigorous financial and operational criteria:

• Industry affiliation: Activity in the manufacturing sector (section C of PKD codes) or services for industry.
• Market track record: A minimum of 5 full completed financial years before the date of application.
• Financial condition: The weighted average of the last 3 years must meet the following ratios:
• Net Sales Profitability (ROS) $\ge 4$.
• Quick Ratio $\in [0.7; 1.8]$.
• Total Debt to Assets $\le 68$.

The project budget structure in the Dig.IT programme is strictly defined:

1. Mandatory component (min. 60% of costs): Purchase of software licences, commissioned programming work and implementation of digital technologies (integrations, on-the-job training).
2. Optional component (max. 40% of costs): Purchase of fixed assets (servers, computers, control panels), materials and specialised training for employees.

Another important source of financing is the SMART Path (FENG), which supports research and development (R&D) projects and the implementation of innovations, with call dates scheduled throughout 2025. For companies from Eastern Poland, there is a dedicated automation and robotisation programme with a budget of 70 million PLN, supporting a business model of transformation towards a circular economy (CE).

Google Workspace for Nonprofits

Unlike the commercial sector, non-governmental organisations in Poland have access to unique resources under the Google Workspace for Nonprofits programme. Since 2025, this programme has been extended with Gemini (generative AI) features, which are offered free of charge with enterprise-class safeguards.

The key difference between the free consumer version and the NGO version is privacy protection: interactions with Gemini in the organisation version are not used to train public models, and files and data are not subject to review by Google's human moderators.

Organisations can register at google.com/nonprofits, which opens the way to a free plan for up to 2,000 users, including 100 TB of shared cloud storage. This is a powerful tool for NGOs that allows modern work without straining the budget, provided the organisation implements internal procedures for the safe use of these tools.

Data Security and GDPR in the Age of AI

Implementing AI in organisations subject to EU law requires conducting a Data Protection Impact Assessment (DPIA) in accordance with Art. 35 GDPR. The key legal challenges include:

1. The right to be forgotten (Art. 17): Removing personal data from an already-trained neural network model is technically extremely difficult, and sometimes impossible.
2. The principle of data minimisation: AI models require enormous amounts of data, which conflicts with the obligation to collect only the information necessary for a specific purpose.
3. Hallucinations and accuracy: The generation of false information about natural persons by AI can lead to infringement of personal rights and fines imposed by supervisory authorities (as in the case of the 15 million euro fine imposed on OpenAI by Italy's Garante, although later suspended).

For SME and NGO employees, a "safe prompting instruction" should be implemented, which categorically prohibits:

• Entering client, donor and employee data into public models.
• Pasting fragments of code containing API keys or passwords.
• Entering strategic plans and confidential documents into systems that do not have "Enterprise Data Protection" guarantees.

Conclusions and Strategic Recommendations

In summary of the analysis of the AI technology market in 2026, technology leaders in the SME and NGO sector should adopt the stance of a "pragmatic sceptic". Artificial intelligence is a valuable tool, but its implementation must be based on solid engineering foundations, not on marketing promises.

Key steps for organisations:

• Operational realism: Accept that the real increase in team productivity will be around 10%, and the time saved should be spent on improving quality and security, not on increasing the number of tasks.
• Infrastructure sovereignty: Begin testing Mistral models and the n8n platform in a self-hosted model. This is the only path to full control over costs and data privacy in the long term.
• Use of financing: Concentrate efforts on obtaining grants from the Dig.IT or KPO programme, treating them as capital for building one's own competences and infrastructure, and not just for buying ready-made subscriptions.
• Debt management: Introduce rigorous standards for documenting and reviewing AI-generated code to avoid "comprehension debt", which could become a company's "silent killer" in the future.

Artificial intelligence is not a revolution that will replace engineering, but a new layer in the technology stack that requires CTOs to be even more disciplined in managing architecture and security. Success will come to those organisations that can use the 10% real growth to build lasting value, instead of chasing the mythical 10x, building on the shifting sands of the unstable economic models of large corporations.

Sources

1. This CTO Says 93% of Developers Use AI, but Productivity Is Still 10% - ShiftMag, accessed: March 26, 2026, https://shiftmag.dev/this-cto-says-93-of-developers-use-ai-but-productivity-is-still-10-8013/
2. AI productivity gains are 10%, not 10x | daily.dev, accessed: March 26, 2026, https://app.daily.dev/posts/vdhjcnh93
3. The 10x AI Developer is a Myth - Emergent Minds | paddo.dev, accessed: March 26, 2026, https://paddo.dev/blog/ai-developer-productivity-myth/
4. Comments - AI productivity gains are 10%, not 10x, accessed: March 26, 2026, https://open.substack.com/pub/abinoda/p/ai-productivity-gains-are-10-not?utm_source=post&comments=true&utm_medium=web
5. AI Technical Debt: How AI-Generated Code Creates Hidden Costs - Tembo.io, accessed: March 26, 2026, https://www.tembo.io/blog/ai-technical-debt
6. What is ARC-AGI? - ARC Prize, accessed: March 26, 2026, https://arcprize.org/arc-agi
7. arXiv:submit/7403127 [cs.AI] 24 Mar 2026 - ARC Prize, accessed: March 26, 2026, https://arcprize.org/media/ARC_AGI_3_Technical_Report.pdf
8. ARC Prize, accessed: March 26, 2026, https://arcprize.org/
9. ARC-AGI-3: Every AI Model Scored Under 1%, accessed: March 26, 2026, https://www.theneurondaily.com/p/play-the-puzzle-that-broke-every-ai-model
10. Full Report: Venture capital investments in artificial intelligence through 2025 | OECD, accessed: March 26, 2026, https://www.oecd.org/en/publications/venture-capital-investments-in-artificial-intelligence-through-2025_a13752f5-en/full-report.html
11. Why Are ChatGPT, Claude & Gemini Losing Billions? The Real Endgame | Maher Saham, accessed: March 26, 2026, https://www.mahersaham.com/blogs/why-ai-companies-losing-billions-endgame
12. Nvidia CEO says elite engineers and AI researchers should spend at least $250K on tokens annually, or he'll 'go ape' - R&D World, accessed: March 26, 2026, https://www.rdworldonline.com/nvidia-ceo-jensen-huang-says-spend-250k-on-ai-tokens-annually-or-hell-go-ape/
13. AI bubble - Wikipedia, accessed: March 26, 2026, https://en.wikipedia.org/wiki/AI_bubble
14. Why Everybody Is Losing Money On AI, accessed: March 26, 2026, https://www.wheresyoured.at/why-everybody-is-losing-money-on-ai/
15. The Hidden Costs of Building with AI: Why Maintenance and Operations Matter - ApyHub, accessed: March 26, 2026, https://apyhub.com/blog/hidden-costs-of-building-with-ai?ref=peerlist
16. Top 12 Free Generative AI Tools for Europe (2026) - Fueler, accessed: March 26, 2026, https://fueler.io/blog/top-free-generative-ai-tools-for-europe
17. The 9 Best European AI Companies Right Now - Noota, accessed: March 26, 2026, https://www.noota.io/en/best-european-ai-guide
18. Large Language Models (LLM) GDPR Compliance, accessed: March 26, 2026, https://gdprlocal.com/large-language-models-llm-gdpr/
19. European AI: 7 companies and models that decision makers should know - data:unplugged, accessed: March 26, 2026, https://www.data-unplugged.de/en/blog/european-ai-models
20. Top 10 n8n Alternatives for 2026 [Tested & Reviewed] - Lindy, accessed: March 26, 2026, https://www.lindy.ai/blog/n8n-alternatives
21. n8n for NGOs: a secure and efficient Automation Tool - iXiam Global Solutions, accessed: March 26, 2026, https://www.ixiam.com/en/blog/n8n-for-ngos-a-secure-and-efficient-automation-tool/
22. 15 Best n8n Alternatives in 2026 - Vellum, accessed: March 26, 2026, https://vellum.ai/blog/best-n8n-alternatives
23. Low-code i No-code – czy programowanie wkrótce stanie się zbędne? - Devstock Academy, accessed: March 26, 2026, https://devstockacademy.pl/blog/narzedzia-i-automatyzacja/lowcode-i-nocode-czy-programowanie-wkrotce-stanie-sie-zbedne/
24. GDPR compliance considerations for self-hosted n8n - LumaDock, accessed: March 26, 2026, https://lumadock.com/tutorials/n8n-gdpr-compliance
25. AI i Automatyzacja w n8n: Jak Zrewolucjonizować Workflow w 2025 Roku, accessed: March 26, 2026, https://kawalec.eu/ai-i-automatyzacja-w-n8n-jak-zrewolucjonizowac-workflow-w-2025-roku/
26. n8n w praktyce z Cognity – przykłady workflow, o które najczęściej pytają klienci, accessed: March 26, 2026, https://www.cognity.pl/n8n-w-praktyce-z-cognity-przyklady-workflow
27. AI w Biurze: 5 Sprytnych Automatyzacji w n8n, Które Oszczędzą Ci Czas - AI Droga, accessed: March 26, 2026, https://aidroga.pl/ai-w-biurze-5-sprytnych-automatyzacji-w-n8n-ktore-oszczedza-ci-czas
28. Jak wykorzystać automatyzację n8n w marketingu? - kingasroka.pl, accessed: March 26, 2026, https://kingasroka.pl/jak-wykorzystac-automatyzacje-n8n-w-marketingu/
29. n8n w biznesie – przykłady automatyzacji w marketingu i sprzedaży - Dokodu, accessed: March 26, 2026, https://dokodu.it/blog/n8n/przyklady-biznesowe
30. Is Airtable Too Expensive? 5 Self-Hosted Alternatives Compared by Cost & Features, accessed: March 26, 2026, https://www.nocobase.com/en/blog/5-self-hosted-airtable-alternatives
31. Dig.IT Transformacja Cyfrowa Polskich MŚP | ARP - ECDF Dotacje, accessed: March 26, 2026, https://ecdf.pl/dotacje/dig-it-transformacja-cyfrowa-polskich-msp/
32. Kto może ubiegać się o grant Dig.IT 2025? Kompletny przewodnik po kryteriach dla wnioskodawców, accessed: March 26, 2026, https://quantus.com.pl/kto-moze-ubiegac-sie-o-grant-dig-it-2025-kompletny-przewodnik-po-kryteriach-dla-wnioskodawcow/
33. Dig.IT - Wsparcie transformacji cyfrowej polskich MŚP - Sekwencja.eu, accessed: March 26, 2026, https://www.sekwencja.eu/dotacja/dig-it-transformacja-cyfrowa-polskich-msp/
34. Rusza pierwszy nabór wniosków w programie „Dig.IT – Transformacja Cyfrowa Polskich MŚP” - Euro-Funding Poland, accessed: March 26, 2026, https://euro-funding.com/pl/blog/rusza-pierwszy-nabor-wnioskow-w-programie-dig-it-transformacja-cyfrowa-polskich-msp/
35. Dig.IT - Digit, accessed: March 26, 2026, https://digit.arp.pl/
36. Dotacje na Cyfryzację w 2025 roku - Digital-Centre, accessed: March 26, 2026, https://digital-centre.pl/aktualnosci/dotacje-na-cyfryzacje-w-2025-roku/
37. Aktualizacja mapy dotacji na 2025 rok – kompleksowe informacje o nadchodzących konkursach dotacyjnych w jednym miejscu - Nowe Dotacje Unijne 2021-2027, accessed: March 26, 2026, https://nowedotacjeunijne.eu/aktualizacja-mapy-dotacji-na-2025-rok-kompleksowe-informacje-o-nadchodzacych-konkursach-dotacyjnych-w-jednym-miejscu/
38. Sprawdź ofertę Funduszy Europejskich i KPO dostępną we wrześniu, accessed: March 26, 2026, https://www.funduszeunijne.gov.pl/strony/wiadomosci/sprawdz-oferte-funduszy-europejskich-i-kpo-dostepna-we-wrzesniu-2025/
39. Workspace for Nonprofits: No-Cost AI Tools, accessed: March 26, 2026, https://workspace.google.com/learning/google-workspace-for-nonprofits-with-gemini
40. New AI features coming to Workspace for Nonprofits - Google Blog, accessed: March 26, 2026, https://blog.google/company-news/outreach-and-initiatives/google-org/gemini-google-workspace-nonprofits/
41. Use Gemini for nonprofits - Google Workspace Learning Centre, accessed: March 26, 2026, https://support.google.com/a/users/answer/14571258?hl=en
42. Compare Google Workspace features for nonprofits | Billing & subscriptions, accessed: March 26, 2026, https://knowledge.workspace.google.com/admin/billing/compare-google-workspace-features-for-nonprofits
43. Google Workspace for Nonprofits edition | Getting started, accessed: March 26, 2026, https://knowledge.workspace.google.com/admin/getting-started/editions/google-workspace-for-nonprofits-edition
44. Why AI-Generated Code Costs More to Maintain Than Human-Written Code | by AlterSquare, accessed: March 26, 2026, https://altersquare.medium.com/why-ai-generated-code-costs-more-to-maintain-than-human-written-code-91b57256bd6a
45. True Cost of AI-Generated Code. A Strategic Analysis of "Comprehension… | by Justin Hamade | Medium, accessed: March 26, 2026, https://medium.com/@justhamade/true-cost-of-ai-generated-code-f4362391790c
46. Processing Personal Data in the Context of AI Models: EDPB's Opinion 28/2024, accessed: March 26, 2026, https://www.europeanpapers.eu/europeanforum/protecting-personal-data-in-context-of-ai-models
47. Large language models (LLM) | European Data Protection Supervisor, accessed: March 26, 2026, https://www.edps.europa.eu/data-protection/technology-monitoring/techsonar/large-language-models-llm_en
48. Generative AI and GDPR Enforcement in Europe: A Lot of Noise, One Fine, Zero Survivors, accessed: March 26, 2026, https://www.crossborderdataforum.org/generative-ai-and-gdpr-enforcement-in-europe-a-lot-of-noise-one-fine-zero-survivors/
49. Training AI models – European Data Protection Board's opinion and recent developments, accessed: March 26, 2026, https://www.kennedyslaw.com/en/thought-leadership/article/2025/training-ai-models-european-data-protection-board-s-opinion-and-recent-developments/
50. Kompletny Przewodnik Po Promptach AI – Ponad 100 Gotowych Rozwiązań, accessed: March 26, 2026, https://securitybeztabu.pl/100-rozwiozan-ai-sec/