Report: The AI implementation gap

In 2024 and 2025, every industry newsletter wrote about the "AI revolution" in companies. Meanwhile, the McKinsey Global AI Survey 2024 provides hard data: although 78% of companies worldwide declare that they use AI in at least one business function, only 1% consider themselves "AI mature" — that is, organisations that have truly redesigned their processes around artificial intelligence. This ninetyfold difference between "we use" and "we transform" is precisely the implementation gap.

In Polish SMEs the problem is even more acute. According to the report of the Polish Economic Institute (PIE — Polski Instytut Ekonomiczny, 2024), 65% of small and medium-sized enterprises declare an intention to deploy AI within 2 years, but only 18% have completed their existing projects with a measurable ROI. The rest are stuck at the pilot stage, in "proof of concept" or have abandoned the investment. This article diagnoses why this is happening — and what specifically can be done about it.

The text is addressed to three groups of readers: owners and managers of SMEs considering an AI deployment, local government and institutional decision-makers designing support programmes, and journalists and analysts mapping out the Polish market. All figures come from public sources — a full list is provided at the end of the report.

The AI implementation gap is the difference between the declared adoption of a technology (the company has bought a licence, launched a pilot) and the actual operational use of it (the technology brings measurable ROI in the company's day-to-day activity).

The best metaphor is a bridge over a river: on one side stands an SME that has just bought an AI tool (e.g. a voice assistant for customer service). On the other side awaits a real improvement — shorter handling time, lower costs, higher customer satisfaction. The bridge between these two banks is implementation: training employees, integrating with existing systems, changing processes, validating results. Most companies never finish building the bridge. The pilot runs for a few weeks, then the team returns to old habits.

Gartner, in its Hype Cycle for Generative AI 2024, places most Polish firms in the "Trough of Disillusionment" phase — just after the peak of expectations, in the valley of disappointment. This is a natural phase of adoption for any technology, but for SMEs it is particularly costly because budgets are small and the margin for error is narrow.

Three key indicators of the implementation gap that are worth monitoring:

(1) the time from pilot to production — if it exceeds 6 months, the project is probably stuck;

(2) the percentage of employees who use the tool every day — if it is less than 40%, the tool is "shelfware";

(3) measurable ROI within 12 months — if there is none, the project is an investment loss, not an investment.

McKinsey "The state of AI 2024": 78% of companies worldwide use AI in at least one business process (up from 55% in 2023). But only 23% report a material impact on EBIT at the organisational level. The rest are functional deployments (marketing, IT, customer service) that do not translate into company results.

Gartner AI Hype Cycle 2024: the average time from a Generative AI pilot to a full production deployment — 20 months. For comparison: for classic ERP systems — 9 months. AI is organisationally more difficult because it requires changes to processes, not just to the technology.

PIE (Polish Economic Institute) 2024: in Polish SMEs 62% of AI projects never advance beyond the pilot phase. 78% of deployments that did reach production have a measurable ROI below what was assumed. Only 5% of Polish SMEs integrate AI in more than 3 business processes simultaneously.

Eurostat Digital Intensity Index 2024: Poland ranks 18th out of 27 EU member states in terms of SME digitalisation. In the "AI use" category we are 22nd. Our closest competitors are Bulgaria, Romania and Slovakia. The leaders (Denmark, Finland, the Netherlands) have indicators 2-3 times higher than Poland.

MIT Sloan Management Review 2024: in a study of 5,000 companies worldwide, only 11% report that they have achieved "significant financial value" from AI. 52% report value lower than assumed. 37% are not able to measure value at all. This means that the problem of measuring ROI is more widespread than the problem of technology.

An analysis of over 600 case studies from the Polish and international markets (Harvard Business Review, MIT Sloan, PIE) makes it possible to identify seven recurring causes of deployment failures. Each of them is individually avoidable — but they often occur in combination, which doubles the risk.

Cause 1 — lack of business strategy. The most common trap: a company buys an AI tool because "everyone has one", without a clear answer to the question "what problem does it solve?" Without a problem to solve, there is no way to measure success. Sanity test: if you cannot describe in a single sentence what business decision the AI is meant to improve — pause the project.

Cause 2 — low data quality. AI is an algorithm fed by data. If the data is fragmented, inconsistent (e.g. the customer's name in three different formats), out of date or biased — the model learns from garbage. A classic ML engineering saying: garbage in, garbage out. Most SMEs have never inventoried their data, yet expect AI to "somehow figure it out".

Cause 3 — lack of in-house competence. After deployment, the company is usually left alone with the tool. If no one in the organisation can interpret the model's outputs, adapt it to business changes or replace faulty components — the system gradually loses value. Warning indicator: if external support ends with a one-off training session, the project will die in 6 months.

Cause 4 — choosing the wrong use case. Companies select the most spectacular applications (e.g. "a chatbot on the homepage") instead of those that actually return the investment (e.g. "automating the preparation of offers for B2B customers"). The rule: the best first project is one in which ROI is measurable after 90 days and the cost of failure is bearable.

Cause 5 — underestimating infrastructure costs. The licence for the AI model itself is often only 30% of the total cost of deployment. The rest is integration with company systems (CRM, ERP, email), security, monitoring, operational API costs (every query to the model costs money). Companies regularly plan a budget of 100,000 PLN and spend 350,000 PLN.

Cause 6 — organisational resistance. AI evokes employees' fears: "will it replace my position?", "will I be controlled by an algorithm?". Without a deliberate internal communication strategy and the involvement of the team in the project at an early stage, resistance sabotages the deployment. An employee who does not use the tool makes the tool useless — no matter how good the model is.

Cause 7 — compliance, GDPR and the AI Act. The year 2024 brought the AI Act (EU Regulation 2024/1689), which classifies AI systems by risk. "High-risk" deployments (e.g. AI in recruitment, in credit scoring) require audit, documentation and certification. Many SMEs deploy tools without realising that in 6-12 months they will need to meet compliance requirements — or withdraw the product.

The Polish SME market has three features that intensify the implementation gap compared with Western European markets.

The first feature — fragmentation: the average company in Poland has 3-4 employees, in Germany — 11. Smaller scale means fewer resources for experiments.

The second feature — linguistic: most commercial AI models (ChatGPT, Gemini, Claude) were trained mainly on English-language data. Quality in Polish is lower than in English — especially in niche industries (medicine, law, engineering). Polish models such as Bielik (NASK — the Polish national research institute) and PLLuM (a consortium of universities) are filling this gap, but they are still at an early stage of development.

The third feature — lack of industry data: global companies have access to large data sets in their sectors. Polish SMEs often operate on markets where no publicly available benchmarks exist — which makes training models and comparing results more difficult. The consequence: Polish AI deployments are less accurate and mature more slowly.

There is also a positive asymmetry: Polish SMEs have lower IT specialist labour costs than in the West, which theoretically allows them to deploy more for less. In practice, this potential is poorly exploited — because there are too few AI specialists in Poland and they are more expensive than classic developers.

Case 1 — a furniture factory in Pomerania (40 employees, manufacturing sector). The company invested 180,000 PLN in a predictive maintenance system — AI was to predict machine failures before they occurred. After a year: the system was working, but on historical data that no one in the company ever revisited. The lack of a process for reacting to alerts meant that when the AI signalled a risk, no one knew what to do. Lesson: technology without a decision-making process = waste of money.

Case 2 — an accounting office in Gdańsk (12 employees, B2B services). The company deployed a chatbot to answer customer queries. After 3 months customers complained that the bot replied "mechanically and did not understand the context". It turned out that the bot was connected to a generic model (GPT-3.5), with no training on the office's industry materials. Lesson: out-of-the-box AI rarely fits a niche business — you have to retrain the model or choose a different one.

Case 3 — an HR-tech start-up in Tricity (8 employees). The company created AI for CV filtering for corporate clients. After 6 months, one of the clients discovered that the model was discriminating against candidates with Polish surnames (a data-training bias effect). A post on LinkedIn = loss of trust, the client terminated the contract. Lesson: the AI Act and GDPR (in Polish: RODO) treat bias as a serious legal risk. A fairness audit should be built into the deployment process, not added after a failure.

Conclusion from the three case studies: in none of them was the problem technical. The models worked. The problem was organisational — lack of process, lack of retraining, lack of audit. This is good news: organisational mistakes are easier to fix than technological ones. All it takes is knowledge and methodology.

Analysis of successful deployments (PIE, MIT Sloan, McKinsey 2024) shows five patterns that consistently appear in projects with high ROI.

Pattern 1 — pilot instead of big bang. Start with a single concrete process (e.g. automating one report). Measure ROI after 90 days. If it works — scale. If not — close it down and draw conclusions. A pilot costs 20,000-50,000 PLN and lets you avoid a 500,000 PLN catastrophe.

Pattern 2 — competences before licences. Invest in training the team BEFORE you buy the tool. Employees who understand how AI works (at the conceptual, not technical level) are able to choose the right tool, negotiate a better contract with the supplier and maintain the system after deployment.

Pattern 3 — ROI in 12-24 months, not 6. A realistic horizon for return on investment in AI is 18 months for medium-sized projects. Shorter pilot projects may have ROI in 3-6 months, but the full transformation of a process requires time for adoption.

Pattern 4 — human-in-the-loop. Do not try to automate completely. The best AI deployments in SMEs are those in which AI proposes, the human decides. The model does the heavy lifting of preparation (data analysis, draft replies, suggested priorities), and the employee verifies, modifies and approves. This increases trust, reduces risk and allows problems to be detected quickly.

Pattern 5 — fairness and bias audit at the very beginning. Before deploying a model concerning people (recruitment, evaluation, customer segmentation), run a bias audit. The test: does the model treat different demographic groups equally? The AI Act requires this legally for high-risk systems — but for any deployment it is simply good practice.

An SME does not have to go through the AI transformation on its own. The Polish National Recovery Plan (KPO — Krajowy Plan Odbudowy) allocates more than 2 billion PLN to the digitalisation of small and medium-sized companies in 2024-2026. European Funds for Pomerania 2021-2027 has a specific priority axis devoted to the digital transformation of SMEs.

The second pillar of support is EDIH (European Digital Innovation Hubs) — a network of innovation centres financed by the European Commission. Each EDIH offers a free digital audit for SMEs, pilot AI deployments and training support. In Poland there are 7 EDIHs operating, including one in Pomerania.

The third pillar — non-governmental organisations (NGOs). Foundations such as Baltic Digital Institute offer impartiality, which no company selling specific software has. An NGO does not earn from a licence, so it recommends the tool best suited to the client's needs — even if it is open-source software at zero cost.

For an SME planning an AI deployment, the best sequence is:

(1) an EDIH audit (free of charge, gives an objective map of processes),

(2) team training from an NGO/university (free or low-cost),

(3) a pilot financed by KPO/EU funds (a grant of 50-70% of the cost),

(4) scaling with a commercial partner or independently. This sequence minimises financial risk and maximises the chances of success.

Five concrete recommendations for local government decision-makers, regulators and creators of support programmes — based on data from the Polish and international markets.

Recommendation 1 — AI support programmes for SMEs should finance COMPETENCES, not only technology. Currently 80% of grants go to the purchase of software. Meanwhile, the data show that the shortage of competences is a greater blocker than the shortage of licences. Proposed change: a minimum of 30% of the grant budget for training and mentoring.

Recommendation 2 — introduce an obligation to document ROI in AI grants. Beneficiaries should measure and report the impact of the project on specific business indicators (revenues, costs, handling time). This disciplines companies to choose sensible applications and creates a public knowledge base of what works in Polish SMEs.

Recommendation 3 — support Polish AI models (Bielik, PLLuM) as an alternative to global commercial models. The availability of local models increases technological sovereignty and improves the quality of AI in the Polish language. Public investment in R&D of language models gives a long-term economic advantage.

Recommendation 4 — a network of regional digital mentors for SMEs. Individual training sessions and one-off audits are insufficient. Polish SMEs need a long-term partner (NGO, EDIH, university) in a 2-3 year relationship. Public programmes should finance not projects but mentoring relationships.

Recommendation 5 — simplifying AI Act compliance for SMEs. The AI Act introduces obligations that small companies are unable to meet without support. Proposed change: create a simplified track for SMEs (e.g. companies of up to 50 employees use standard templates of documentation, audits and certifications made available by the state free of charge).

The AI implementation gap in Polish SMEs is a real phenomenon, measured in hard numbers: 65% of companies want to deploy AI, 18% scale it with measurable ROI. The difference between these two figures is 47 percentage points of lost value — and thousands of hours of work in projects that never return on investment.

The seven causes of failure (lack of strategy, poor data quality, lack of competences, wrong pilot scope, underestimating costs, organisational resistance, compliance) are well documented and repeatable. This means that they can be consciously avoided — provided that the company approaches AI methodically, not reactively.

The five patterns of success (pilot instead of big bang, competences before licences, realistic ROI horizon, human-in-the-loop, fairness audit) are a roadmap developed on the basis of thousands of successful deployments. None of them requires a revolutionary technology — all of them require organisational discipline.

The role of public support (KPO, EU funds, EDIH) and non-governmental organisations is irreplaceable in the Polish context. An SME with a grant, training and long-term mentoring has several times higher chances of success than a company acting on its own. Decision-makers designing support programmes should rely on data: finance competences, require ROI measurements, support Polish models and build a network of mentoring relationships rather than single grants.

The AI implementation gap is not inevitable. It is diagnosed and measurable — and that means it is solvable. The subject is not "whether", but "how quickly and in what order".