AI in business: why so many Belgian companies tap only a fraction of its power
The real obstacle is not the tool, but the lack of a data foundation
Many Belgian companies now have access to the same artificial intelligence tools as the largest technology groups. Yet most of them use only a tiny fraction of their real potential. This is not for lack of ambition, but for lack of having built the data foundation that would allow AI to produce anything other than a generic result.
The finding: AI reduced to conventional automation
In most Belgian companies, artificial intelligence is currently confined to tasks that conventional business software could just as easily have performed: sorting invoices according to a fixed rule, sending a template e-mail after an order, filing documents in a predetermined folder. These are well-defined, routine processes, automation in the most traditional sense of the term, simply dressed up with the words artificial intelligence because the tool used to carry them out contains the technology. The confusion is real: a system that applies a fixed rule, a condition that always triggers the same action, learns nothing and does not improve with use, unlike a model trained on the company’s own data, capable of spotting patterns that no rule written in advance could have anticipated.
The causes: no data foundation and little understanding of the technical fundamentals
The first obstacle is purely quantitative. For an AI model to produce a reliable analysis or prediction about a given process, a type of sale, a customer profile or an ordering pattern, you generally need several thousand examples of that same process. Yet most Belgian companies, across all sectors, do not even reach a thousand structured records for one and the same sales principle: the data exists, but it is scattered across mailboxes, isolated spreadsheets and business software that does not talk to each other.
The second obstacle lies in the number of variables, not just the number of rows. A company that has a thousand orders but has never recorded the acquisition channel, the length of the sales cycle or the reason for a refusal gives AI only a flat picture of its business. Without these variables, the tool can neither detect why some customers buy more nor anticipate a change in behaviour. It simply repeats the average of what it sees, for lack of additional angles of analysis to cross-reference.
The third obstacle is a lack of understanding of the very technical fundamentals that make it possible to build this foundation. Ask most management teams what a Git repository is, the tool used precisely to organise, version and centralise a project’s data and code, and the answer is usually silence. The same goes for the notions of iteration, which consists of retraining a model as new data comes in, data cleaning, which consists of removing duplicates and correcting inconsistent formats, or a multi-agent system, in which several specialised AIs share out a complex task. Without this first level of technical culture, it is structurally impossible to prepare a usable database, whatever the budget devoted to the tool itself.
The consequences: computing power reduced to a simple time-saver
This lack of a foundation sustains a mistaken but widespread perception: AI is supposedly just a gadget, a toy that saves a few minutes on a repetitive task, such as writing an e-mail, summarising a document or generating an image. This view is not wrong in itself, but it reduces a high-powered computing tool to the role of a slightly more sophisticated spellchecker.
The reality is the opposite: when it is properly fed, AI does not merely automate what a company already does, it serves to increase its volume, with more customers identified, more orders anticipated and more sales opportunities detected before a competitor sees them. The difference lies solely in the quantity and quality of the data supplied upstream: it is a growth lever, not a convenience tool.
The solution: build the data foundation first, before the tool
In practical terms, the first step is not to choose an AI tool, but to set up a Git repository to centralise and structure the company’s data, and then to identify the two best sources already available to feed it. For virtually every company, these are the complete history of e-mails exchanged with customers and suppliers, and the history of orders, purchases and sales: two deposits that almost every company has been accumulating for years without ever exploiting them. Where they exist, transcripts of telephone calls, obtained by converting speech to text, usefully complement this foundation, but they remain rare: few Belgian companies systematically record their voice communications. E-mails, orders and official communications are already enough to build a solid database, without heavy technical investment, enabling AI to start creating relevant FAQs or to better understand the company’s real activity.
This is precisely the approach Selion puts in place for its clients. An initial audit identifies the data actually available and its condition. This data is then collected and integrated into structured pipelines, then analysed at scale, including across several million conversations where that volume exists. This analysis makes it possible to extract the most useful information for building FAQs, both for internal use and for external customers, and to better understand how information flows within the company in order to improve it continuously.
This work is not limited to a one-off set-up. The aim is to make collection and analysis recurring and automated, so as to continuously feed the company’s knowledge base. This knowledge base is then used to train internal staff and newcomers, and to answer customers across all channels: e-mail, the website and sometimes telephony when the company wishes to go that far. The ultimate goal is to document the bulk of the company’s activity so that it can be shared both internally and externally.
A significant part of this work consists of filtering out what must never leave the company. Passwords, security keys and any sensitive information are excluded before any use of the data, as are exchanges that do not concern general questions intended to be shared.
In principle, this project is simple to get started: all it takes is collecting the existing data. Companies that do not have the in-house skills to go further can call on Selion, which takes charge of the entire process, with results generally delivered within a month.
📋 In summary
| Issue identified | Observed reality | Course of action |
|---|---|---|
| AI reduced to automation | Confusion with a fixed-rule system that does not learn | Clearly distinguish between automated rules and a model trained on data |
| Insufficient data foundation | Often fewer than 1,000 structured entries per process, few variables recorded | Centralise e-mails, orders, purchases/sales in a structured repository |
| Little understanding of technical fundamentals | Git, iterations, data cleaning, multi-agent: concepts unknown in-house | Get trained or get support on these basics before investing in the tool |
| AI seen as a time-saver | Perceived as a mere convenience tool rather than a growth lever | Steer usage towards increasing customer/order volume |