adminAI in games business intelligence could help publishers and studios make sense of market shifts faster, assess investment opportunities and explore possible outcomes. Joost van Dreunen, co-founder of SuperData Research and founder of analysis firm Aldora, argues that small language models (SLMs) may be particularly useful because they can focus on a company’s specific questions and data.
Image credit: Jakub Żerdzicki
This article is part of AI Week.
Why games companies need faster business intelligence
Traditional market research can involve lengthy planning, surveys, vendor reviews and multiple rounds of analysis. By the time a report is ready, audience interests and commercial conditions may already have changed. Van Dreunen sees this as a structural challenge: established research processes can struggle to keep pace with a fast-moving games market.
Sudden hits such as Grow A Garden and Steal A Brainrot illustrate why early signals matter. Smaller specialists may notice emerging trends, while larger organisations can find it harder to identify them in time.
Practical takeaway: Treat market research as an ongoing activity, not a one-off report. Regularly reviewing new data can help teams spot changes sooner, while still allowing time to check whether a signal is meaningful.
How AI in games business intelligence could work
Rather than focusing mainly on generating content, Van Dreunen describes AI as a way to make information easier to access and use. A team could ask questions, compare scenarios and revisit an analysis as new information becomes available, instead of waiting weeks for a completed research project.
SLMs are smaller than large language models (LLMs) and can be designed for a narrower set of tasks. A studio making mobile shooter games, for example, may need answers about its own category rather than broad coverage of every part of the games industry. Van Dreunen argues that a more focused model may be less costly to maintain and more energy efficient. It may also be possible to run some models locally, including on a smartphone, rather than relying on a continuous online connection.
He contrasts a model with 7 billion parameters, which may be sufficient for a specific task, with a much larger model of 700 billion parameters. The appropriate size depends on the job, the quality of available data and the level of expertise required.
Practical takeaway: Define the questions the system needs to answer before selecting a model. A focused use case makes it easier to assess whether the tool is useful, secure and worth maintaining.
Using internal data while protecting sensitive information
A localized SLM could combine external market information with a company’s private data without sending proprietary project details to a cloud service, according to Van Dreunen. In his proposed approach, users could examine a planned project or launch against industry research while keeping internal information within their organisation.
The model could also be trained or refined with company-specific material. Over time, it might help users identify the underlying question behind a request—for instance, distinguishing a general user-acquisition concern from a more specific question about weapon variety in a shooter game.
Van Dreunen says Aldora had spent two years aggregating structured and unstructured information for its training data. Sources included earnings reports, executive transcripts, academic articles, case studies, interviews and other material relevant to games businesses. He described a product rollout as planned for Q2 or Q3, with Q3 his expectation and Q2 his hope at the time of the interview.
Practical takeaway: Before using private information, confirm where data is stored, who can access it and whether it is used for training. Local deployment can support privacy, but security still depends on the system’s configuration and governance.
Can AI help publishers predict investment decisions?
Van Dreunen’s approach also involves looking at the people and organisations shaping the industry. Public statements, career histories and past decisions may provide useful context for understanding what a company or leader could prioritise next. That does not make an AI system an oracle: past behaviour can inform a forecast, but it cannot guarantee a future decision.
He points to the limited number of people who make influential choices across the games industry as one reason this analysis may be valuable. The aim is to help publishers, investors and other businesses examine likely scenarios—not to replace judgement or claim certainty.
This could also help companies outside the games sector understand the market. Brand managers, private equity firms and venture capital investors may be interested in games but lack detailed industry knowledge. A tool that combines market data with relevant analysis could help them ask more informed questions before committing to a project or investment.
Practical takeaway: Use AI-generated forecasts as decision support. Compare the output with the underlying evidence, consider alternative explanations and involve people who understand the relevant market.
Why human review remains essential
AI systems can produce confident but incorrect answers. Van Dreunen says human operators remain necessary, even if their involvement adds time and cost. Users need to understand the data behind a result and investigate how the system reached its conclusion.
He compares this to business intelligence tools such as Excel and SQL: sophisticated analysis is not valuable if the input data has been organised incorrectly. In the same way, a model’s answer can be misleading if its sources are incomplete, poorly structured or out of date.
For that reason, people using an SLM may need training in how to ask precise questions, check supporting data and recognise uncertainty. A smaller model can make information more accessible, but it does not remove the need for analytical expertise.
Conclusion: AI in games business intelligence is decision support
AI in games business intelligence could make specialist market insights quicker to access and easier to apply to investment and product decisions. Small language models may be a good fit for focused tasks, especially where organisations want to use private data in a controlled environment. Their forecasts should still be treated as informed possibilities rather than predictions of what will definitely happen. Human review, reliable data and clear security practices remain essential.
Joost van Dreunen | Image credit: THIS 2023/Alexander Håkansson
The interview has been edited for length and clarity.
