Many organisations and development teams have already experienced failures with artificial intelligence (AI) projects. Research bodies such as McKinsey and Gartner note that the chance of an AI project's failure can be many times its chance of success — the article cites roughly a twentyfold difference. Nonetheless, there are identifiable areas where deploying AI makes economic sense.
Where does AI pay off? Concrete figures
To judge feasibility, companies must know precisely where and how much time employees spend on particular processes. If a task consumes 10 percent of working time and can be optimised by 20 percent, the net improvement across the whole process is only 2 percent — typically not enough to justify a major investment. Conversely, if a task occupies 60–70 percent of staff time, a 20 percent efficiency gain translates to about a 14 percent overall improvement, which often makes costly developments profitable.
This scale effect explains why a system-driven approach — focusing on specific process parts and measurable gains — usually yields better returns and lower risk than an AI-driven strategy, where AI is adopted on the hope it will generally improve operations.
AI-driven versus system-driven deployment
Data show AI-driven projects tend to be riskier. When an AI model receives a well-defined question and the necessary data, its outputs are of good quality. If, however, the model is given unstructured document collections and asked to ‘find something’, its performance is likely to be weaker. Therefore, system-driven implementations that target concrete tasks carry less uncertainty.
Data security and regulation
One must carefully consider which data can be entrusted to public models. Certain categories — for example, customers’ health data — usually cannot be shared externally. System design must account for such restrictions: an available integration (e.g., with OpenAI) is useless if the required data cannot be supplied.
The regulatory environment reinforces this caution: the European Union’s AI regulation (EU AI Act) entered into force in August of this year, and organisations must also observe cybersecurity directives such as NIS2. Increasingly frequent data-leak incidents show that once data leave an internal system control is often lost. For sensitive data, building local, closed systems is therefore recommended.
According to in-house measurements cited in the article, locally run models typically achieve around 80–90 percent of the performance of comparable cloud solutions while keeping sensitive data inside the organisation. The article mentions that Quattrosoft uses an NVIDIA DGX Spark device with 128 gigabytes of memory, capable of running substantial models on-premises.
Rapid change and continuing risks
AI is a very fast-moving field: decisions such as which model to use must be made in an environment where a significantly better model may appear within weeks. Another concern is the reliability of AI-assisted development: the most optimistic estimates suggest code generated by AI contains about 1.5 times as many bugs as human-written code; other estimates indicate the factor could be much higher, even up to ten times.
The volume of AI-generated content online is also growing; its average quality is slightly lower than human-created content. Feeding such AI-produced material back into training data may degrade, rather than improve, future model quality. For this reason, some initiatives collect works written before 2022, which are less likely to contain AI-generated content.
Conclusion: seize opportunities with controls
Despite uncertainties and rapid evolution, the article argues that organisations should take advantage of the opportunities offered by AI. The recommended approach is to use AI where it is suited — for example, to optimise systems and processes — and to carry out deployments in a system-driven manner with appropriate data governance and regulatory controls.
