1. Clean architectures and a unified data platform
When the first edition of The Data Science Book was published, we strongly advocated for a clean, streamlined data architecture. In this approach, KPI definitions are not embedded in reports or dashboards, nor are they calculated during data extraction from the source. Instead, they are meticulously documented in the data warehouse. Today, we’re seeing a growing preference among organizations for this type of clean, pure architecture.
In the past, organizations typically used one tool to generate and consume management information. Now, the data warehouse is a central hub, accessed in numerous ways.
This shift is largely driven by the increasing adoption of unified data platforms, such as Azure, Google Data Cloud, or AWS Analytics, which make all information accessible from a single source. The need for such architecture is also amplified as a wider range of users tap into the data warehouse. Beyond traditional BI tools like Tableau, SAS Visual Analytics, or Power BI, platforms such as websites, intranets, apps, Python, R, and other BI and AI tools now rely on the centrally defined data.
Even Web services like REST APIs are essential for ensuring accessibility for both internal and external stakeholders. This clean architecture, which we championed two decades ago, is finally getting the recognition it deserves. Dubbed the “headless BI architecture,” it’s set to take center stage in 2025.
2. Is this the end of the data vault?
We’ve never been strong advocates of the data vault for reporting and analytics, so this trend doesn’t surprise us. With the rise of unified architectures for (big) data storage, reporting, dashboarding, data analysis, and machine learning (see trend 1), the need for a costly data vault has all but disappeared—except in rare, exotic use cases. For many, this could signal the final death blow to the data vault concept.
The original aim of the data vault was to ensure organizations could retrieve and verify “the truth” at any point in time, even down to the minute from years ago. This was particularly appealing for industries like banking. However, modern platforms such as Microsoft Fabric now provide robust extraction, transformation, and long-term storage capabilities (e.g., Parquet, JSON) for both structured and unstructured data. These platforms include built-in support for versioning, lineage, and governance, ensuring traceability without the need for a separate data vault.
The persistent staging area essentially functions as an archive—a massive data lake that accommodates all forms of data. Whether it’s internal or external, ERP or CRM records, photos, sound clips, videos, or documents, it all finds a place in this architecture. For financial institutions or healthcare organizations that need to demonstrate compliance, you can easily create a view or report directly on the data lake. This setup allows for seamless integration of historical data from diverse sources and handles semantic drift, ensuring that attributes or definitions that change over time are properly addressed.
If you’re considering or re-evaluating a data vault, now is the time to get a second opinion on its relevance.
3. Cloud BI and AI becoming the standard
Cloud-based BI has been available for years, but only recently have we seen a surge in organizations planning to move both their BI and AI operations to the cloud in the coming years. The key drivers? Scalability, immense computing power, robust built-in security protocols (think zero trust), and the ability to rapidly develop BI and AI solutions. With the cloud BI and AI market growing at an annual rate of 27%, it’s clear that the cloud is no longer optional—it’s the future.
For example, we developed an AI system for a housing association that analyzed 35,000 floor plans and automatically calculated the areas of balconies and roof terraces. Using cloud infrastructure, we leveraged immense computing power and completed the analysis in just one hour—a task that would have taken months otherwise. The project not only saved significant time but also delivered other benefits, such as fewer tenant complaints and informed rent adjustments, all powered by AI. This is the kind of speed and efficiency the cloud makes possible. Don’t underestimate its potential.
Still prefer to keep your BI and AI on-premises? If so, be prepared to defend your choice with solid arguments: comply or explain. However, don’t ignore the environmental impact—cloud data centers are massive energy consumers and demand significant resources. Keep that in mind as you navigate your strategy.
4. AI comes out of the closet
Nowadays, you can pull machine learning models straight out of the closet via an API to the cloud. This is often referred to as Machine Learning as a Service (MLaaS). These pre-trained models come packed with standard functionality and intelligence (e.g., Google AI Platform, SageMaker AI). For instance, they can instantly classify a plant or animal in an image and even identify the species. They can also recognize objects like faces, chairs, or cars with impressive accuracy.
This computer vision is very powerful and is gaining tremendous popularity because you don’t have to train the models (anymore). After all, the appearance of a flower, animal, or human face is unlikely to change substantially in the next few centuries. However, more dynamic patterns, such as the behavior of criminals or fraudsters, evolve as they adapt to avoid detection. Even for these scenarios, cloud-based models are improving and will become more readily available in the coming years. Organizations will still need to fine-tune these models with their own data periodically, but the overall speed of developing automated decisions will see a dramatic boost.
This development is making AI and machine learning significantly more accessible.
The key takeaway is that, for many use cases, organizations no longer need to train their AI systems entirely from scratch. This can help avoid potential issues with privacy and data quality during training. That said, there will still be instances where additional training is necessary, especially for applications where MLaaS is not yet available. For more on this topic, check out our insights on the trend ‘Photography becomes the new universal language‘ or reach out to book an AI inspiration session.
The Artificial Intelligence handbook The Artificial Intelligence book provides an in-depth look at the latest developments in AI and Business Intelligence. It covers current trends such as clean data architectures, the rise of cloud BI and AI, and the integration of AI into daily business processes. Additionally, it explores the impact of generative AI and the importance of data quality.
5. Generative AI overhyped and reaching limits
The introduction of ChatGPT in November 2022 created a societal uproar: a computer that could listen, converse, and generate entire documents and poems? It was groundbreaking. Startups sprang up overnight, and billions of dollars poured into the sector. But now, it seems the generative AI bubble is bursting. Overhyped from the start, the coming year is poised to bring widespread disillusionment.
The cracks are already showing. Generative AI often produces unreliable answers, introduces biases, and lacks depth. Responses tend to be long-winded, repetitive, and not particularly creative. Using tools like ChatGPT or Gemini sometimes requires dozens of prompts to reach a conclusion—and even then, the results often leave users feeling unsatisfied. While GenAI is an excellent mimic, a powerful interactive encyclopedia, a creative brainstorming tool, a skilled programmer, and an efficient summarizer, its limitations are glaring. Despite its appearance of intelligence, it remains inherently shallow and algorithmically constrained.
Adding to the challenges, generative AI has essentially “taught itself out” by exhausting the vast troves of data on the Internet. With little new information to learn from, the momentum has stalled. Compounding this is the emergence of AI-generated content being used to train new models—leading to a self-perpetuating cycle of “tainted” machine learning. This degradation is causing models to falter. While tech companies are exploring solutions like slow, linear learning (inspired by traditional education), it remains unclear whether these efforts will improve model performance. The core issue lies in generative AI’s reliance on probabilities—it predicts the best next word, punctuation, or pixel, based on prompts and context, using machine learning algorithms that have existed for over two decades.
For specific applications, GenAI can outperform the average human, which secures its place in the future of technology. However, for high-quality outcomes, human expertise is still irreplaceable. The hype surrounding generative AI may have passed, but this marks an opportunity for the technology to mature and become genuinely useful. Smaller, more specialized models (Small Language Models, or SLMs) could be a promising direction for the next phase of development.
6. Narrow language models
With generative AI reaching its limits and exhausting its sources of learning, a promising alternative is emerging: narrow language models, or Small Language Models (SLMs). Unlike large language models like ChatGPT, which have combed through the entire Internet, SLMs are trained on focused, highly specialized datasets. This specialization allows them to produce answers that are more specific, relevant, and tailored to their intended applications. While they lack the broad general knowledge and linguistic versatility of large models, this isn’t always necessary for certain use cases.
Consider a help desk system that has amassed and documented years of solutions to computer problems, or a dataset designed specifically for generating construction quotes or managing complex IT projects. In these scenarios, SLMs can shine by delivering highly targeted insights and solutions.
SLMs are also lightweight, fast, and capable of running directly on smartphones without requiring an Internet connection. This makes them particularly appealing for on-the-go applications. In theory, SLMs have the potential to outperform general language models for niche tasks, but their practicality remains to be proven. As the technology is still in its infancy, there are few success stories to draw from—but the potential for narrow language models to revolutionize specialized applications is undeniable.
7. AI literacy becomes mandatory for everyone
The AI Act, passed by the European Union in May 2024, has officially become law. This legislation classifies AI systems into four risk levels—unacceptable, high, limited, and minimal/none—and introduces corresponding regulations to ensure responsible use of artificial intelligence.
Fostering AI literacy is a key objective of this legislation. People should be able to identify AI-generated results and critically assess them. For expert guidance on complying with the AI Act, contact us here.
8. Voice AI is not breaking through
Talking to computers isn’t new, but with advancements in generative AI and natural language processing, Voice AI has become more interactive and potentially useful. This technology shines in scenarios where typing isn’t feasible or practical, such as while driving. For instance, salespeople could dictate a conversation report on the go, with GenAI summarizing it and sending it directly to a CRM system. While some view Voice AI as a rising trend, we believe its widespread adoption remains unlikely.
The concept isn’t new. Back in 2015, I attended a SAS conference in Marbella where a simple form of Voice AI was demonstrated. Users could interact with their BI system via Alexa, asking questions like, “Show sales for the past 12 months broken down by product group,” and receiving answers neatly displayed on the screen. It was impressive, but as with many technologies, success depends on human adoption and acceptance.
While some enthusiasts embrace Voice AI, it seems unlikely that we’ll see a surge in its use for frequent or confidential conversations. Privacy concerns loom large—spoken conversations can be overheard, making them less secure. Additionally, if multiple people are using Voice AI in the same physical space, overlapping conversations can confuse the system, limiting its effectiveness.
There are certainly valuable applications for Voice AI, but we don’t expect it to revolutionize the way we communicate. Its use will likely remain niche rather than becoming a dominant force.
9. Seamless integration of BI & AI
There’s growing consensus that traditional business intelligence (BI)—including reporting, dashboards, and basic data analysis—and artificial intelligence (AI) should be seamlessly integrated. This trend is reflected not only in tools like Tableau, Cognos Analytics, and Power BI, where AI features are embedded into reports and dashboards, but also in customer discussions and the increasing preference for unified data platforms.
The integration of BI and AI is a natural progression, as AI essentially automates data analysis. In this sense, AI is a powerful extension of BI, and the two cannot be effectively separated. However, successful AI implementation requires more than technical expertise. While an AI expert might excel at building models, they often lack the broader knowledge of BI and data analysis necessary to apply AI effectively within organizations. This gap often leads to AI projects failing or never reaching production.
One common pitfall is deploying AI in isolation from BI, which disconnects AI applications from a foundational understanding of the data. For example, predicting pregnancy outcomes in India with AI, without deep knowledge of the underlying data, demonstrated the dangers of such disconnection.
Here’s the reasoning: the purpose of BI is to enable data-driven decision-making through data analysis. AI, in turn, automates that analysis. If AI is deployed without first leveraging BI to thoroughly understand the data, organizations risk poor results, loss of trust, and potentially harmful decisions. AI can generate extraordinary insights, but automation without deep knowledge is a recipe for failure. The mantra is clear: first understand (BI), then automate (AI).
Want to learn how to integrate BI and AI successfully? Consider our Business Intelligence training, designed to equip you with the tools and knowledge to bridge the gap.
10. AI as a big polluter
The data centers powering AI systems like ChatGPT and Gemini consume vast amounts of electricity and significant quantities of drinking water. While studies vary, it’s clear that electricity usage in data centers has doubled over the past three years, driven in part by the rising popularity of AI. A single query to ChatGPT consumes nearly 10 times more energy than a typical Google search. By 2030, data centers are projected to account for 3-4% of global electricity consumption.
This trend is troubling, particularly as we face the urgent need to reduce our ecological footprint. However, given that generative AI is still a relatively new phenomenon—and some believe its hype may subside—the current data might present a distorted picture. Legislation often lags behind technological advancements, but it’s likely that the unchecked growth of data centers will face stricter regulation in the coming years. For instance, companies like Microsoft are already exploring self-sufficiency in power generation for their data centers. The issue of water consumption, however, remains a significant challenge, as it cannot simply be “generated” like electricity.
Interestingly, AI is now being singled out as a major polluter, even though social media platforms—many of which also integrate some AI—have consumed large amounts of energy for years. To bring nuance to the discussion, research shows that up to 80% of data center energy use is unrelated to AI and instead supports standard data processing and storage. With grid congestion already a pressing issue in many Western countries and urban areas, it’s time to collectively address the environmental impact of data centers, regardless of whether AI is the main contributor.
Organizations should start by asking critical questions: Do you really need certain AI applications or BI dashboards? Should ChatGPT be made available to all employees? When implementing AI, ensure it also supports sustainable practices. In many cases, AI can actually contribute to sustainability by reducing waste and increasing precision. With thoughtful application, AI has the potential to align with environmentally responsible operations rather than exacerbating their impact.
11. The sheen of Big Data is fading
With AI taking center stage and capturing most of the attention, the question arises: where does this leave the field of big data analytics? Interest in big data has been waning since November 2022 (as indicated by Google Trends), and its shine seems to have dulled. However, it’s important to remember that AI without big data is significantly less powerful.
Rather than focusing solely on artificial intelligence—particularly generative AI—it’s worth considering how big data applications can still provide value for your organization. As photography and visual data become increasingly dominant forms of communication (see the next trend), the need for robust big data strategies is far from over.
12. Photography is becoming the new universal language
When we talk about big data, we often refer to unstructured or massive volumes of data—videos, emails, social media posts, web pages, sound clips, and, increasingly, photographs. As early as 2012, Marvin Heiferman predicted that photographs, with their ability to encapsulate immense amounts of information, would emerge as the new universal language. Today, nearly everyone can interpret photos, and most people use them as a form of expression. As noted by the Council for Culture in 2024, society is transitioning from a text-based culture to a predominantly visual one. Within the fields of AI and big data, we’re seeing exciting applications that harness photos as valuable data sources.
However, the potential of photography in AI and big data analytics remains underutilized. For instance, we recently collaborated on a pilot project for a municipality aiming to classify litter in the downtown area using AI. Crowd control cameras had already provided the necessary images; the challenge was repurposing them to detect litter.
The power of photography lies in its versatility. A single photo can yield multiple insights. Detailed images of urban areas can detect litter, analyze public space usage, and even optimize leaf sweeping schedules, including workforce needs. Beyond that, photos can assess tree health, evaluate road and sidewalk conditions, identify faulty lighting, gauge neighborhood livability, and much more.
Photography is, without a doubt, the new universal language. If you’re interested in exploring innovative big data applications like these, let’s start the conversation. Contact us today.
13. Data quality rises with a bullet
As organizations rush to adopt AI, they are discovering that reliable outcomes depend on one critical factor: high-quality data. This has sparked a growing interest in data quality and broader data governance practices. Many companies are adjusting their organizational structures to address this need, creating roles like data stewards and Chief Data Officers to treat data as a valuable asset that requires active management and continuous quality improvement.
Fortunately, AI itself can play a role in addressing data quality challenges. Through techniques like data profiling, AI can quickly identify inconsistencies and even correct them. Missing data points can often be derived from related attributes, and generative AI can provide “best estimates” to fill in the gaps. While poor data quality in dashboards or reports tends to be visible and easier to detect, machine learning presents a unique challenge—most of its operations occur behind the scenes, making quality issues harder to spot.
As the demand for robust AI applications increases, the importance of data quality and governance is climbing the agenda for boards and executive teams. Without strong processes to ensure clean, reliable data, the promise of AI cannot be fully realized. Data quality is no longer just a technical issue—it’s a strategic priority.
14. Growing IoT networks provide big data boost
Many complex machines and devices have long lifespans—often 10, 15, or even 20 years. Older models typically lack embedded sensors (IoT) or robust data models essential for performing reliable data analysis. As these machines are replaced or upgraded, a new flow of big data is emerging. Modern equipment is increasingly equipped with advanced IoT capabilities, expanding IoT networks and enabling machines and devices to exchange data seamlessly.
This growing IoT ecosystem is generating larger streams of data, offering unprecedented opportunities for real-time analysis with AI. Organizations can now act on insights immediately, driving efficiency and precision. This expansion of IoT networks is rekindling interest in big data analytics, which had been waning in recent years.
As IoT networks continue to grow, the integration of real-time data with AI analytics will enable organizations to uncover deeper insights and take more informed, proactive actions. This marks a significant turning point for big data and AI applications.
15. Realtime first: the Egg of Columbus
In the early days of BI, it was common practice to update a data warehouse weekly. That schedule worked for years, but as the pace of the world accelerates, our ability to keep up with trends and developments has become crucial. Today, refreshing data daily has become the standard for many organizations. However, even daily updates often fall short of what’s needed. We no longer want to live in the past, constantly staring at the rearview mirror of outdated reports and dashboards. Those days are behind us.
With the advancements in AI, real-time analytics has emerged as a game-changer—because prevention is always better than cure.
Detecting litter at the moment it appears and cleaning it up immediately. Replacing machine parts before they fail. Identifying and stopping criminals before they act. Dispatching street sweepers to remove slippery leaves before someone gets hurt. Enhancing neighborhood livability before it declines. These are just a few examples of the transformative potential of real-time analytics.
This is the Egg of Columbus. The BI and AI standard is shifting from daily updates to real-time-first operations. Sectors like e-commerce, logistics, and financial services are already leading the way. The technology is ready, the data is available, and its quality is improving. Processes are crying out for speed and efficiency. But the question remains: are we, as humans and organizations, ready to embrace the leap to real-time-first?
Do we have the courage to claim AI leadership and push the boundaries of what’s possible? Where there is a will, there is a way. In the end, progress and innovation will triumph over inertia.
Conclusion
This year, we’ve deliberately chosen not to cover topics like self-service BI or AI ethics and privacy. When it comes to self-service BI, the industry’s promises of “data democratization” have been echoing for over a decade. Yet, we see little evidence of real adoption within organizations. What has become clear is that self-service BI often results in the proliferation of poorly constructed statistics, reports, and KPIs created by individuals without sufficient data literacy. It’s time to pause, reflect, and arm ourselves against the nonsense the market sometimes promotes.
AI ethics and privacy, on the other hand, are hot topics in theory but rarely translate into meaningful action. Take the mandatory AI registry, for example—despite its existence, it features very few noteworthy AI systems. The Police have registered only one high-risk AI system, even though we know more are in use. Similarly, the Ministry of Justice and Security and other government entities conspicuously lack entries in the registry. Searching for algorithms used by military intelligence yields no results at all—hardly credible.
This points to widespread “AI-washing,” where discussions of ethics and privacy become more about appearances than substance. Major tech companies remain largely unchecked, growing ever more powerful. Even government websites continue to track visitors, circumventing the true spirit of privacy legislation. The mandatory cookie consent banners are the result of poorly conceived laws that fail to achieve meaningful outcomes. Wouldn’t outright banning the tracking of individual users have been more effective? After all, trust in a shopping street would vanish if shopkeepers secretly shared customer data.
Ethics and privacy, in practice, often exist only on paper. Real ethics, however, isn’t complex: “What you do not want done to you, do not do to others.” This principle involves scrutinizing proposed solutions and their ecosystem-wide impact, weighing the pros and cons, and deciding whether to proceed. As Johan Cruijff famously said, “Every advantage has its disadvantage.” The challenge lies in seeing those trade-offs clearly.
The trends we’ve shared here come from the heart and are grounded in observations, research, and decades of experience in BI and AI. However, even we can fall victim to bias or miss important trends—let us know if you’ve spotted one. If there’s a trend you believe we overlooked, we’d love to hear your thoughts in the comments.
On behalf of the entire Passionned team, we wish you happy holidays and a fantastic 2025!