You found the use case, and you know where artificial intelligence could help you save time, automate certain tasks, or improve your decision-making.
All that remains is to choose the technology and launch the project.
Pause. That’s not entirely true.
Before talking about models, agents, or architecture, there is a much less exciting but far more important question:
Can you trust the data that will power your AI?
Artificial intelligence can analyze your data at impressive speed, detect trends, and produce results in seconds. However, it cannot make unreliable data reliable.
And this is not limited to AI. The same applies to BI and, more broadly, to any initiative that relies on your data.
When a small data problem becomes a big AI problem
If you look at your customer data, you might encounter issues like these and their impact:
- The same customer appears twice, so dashboards count four customers instead of three.
- An address is missing, so our customer’s neighbor receives our offer (or worse, their personal data).
- “Sherbrooke” is written in several different ways, skewing statistics by city.
- A date of birth shows a future year, so our AI excludes the data.
- Each team has created its own indicators, so the different totals do not produce the same results.
In a dashboard, these issues can already lead to inaccurate indicators, inconsistent categories, and different results from one team to another. With AI, they can also influence the predictions, recommendations, or responses produced by the model. In short, your data can send you down the wrong path and cost you time, energy, and money.
The risk becomes harder to detect with AI: a convincing result is not necessarily a reliable one.
“A model’s task is to provide the best possible answer based on its training and source data. It will therefore provide an answer even when its confidence is low.” Sébastien Fleury, COFOMO
In other words, just because AI gives you a clear, well-formulated answer does not mean the data behind that answer is reliable enough to be trusted without validation.
Your AI only knows the reality you show it
Let’s take a concrete example from a credit project.
An organization wants to use AI to predict the risk associated with a request for a cash advance. It has historical data on its customers and credit decisions.
There is one problem, however: it primarily retains data on approved applications.
The model can therefore learn to recognize the characteristics of approved applications, but it has far less information about the factors that led to rejected ones. In a way, we are asking it to understand the whole story while only telling it part of it.
The amount of data alone is therefore not enough:
“Before investing in a use case, you need to know what data you are using (essential data), where it comes from (availability), what condition it is in (quality level), what it means (business definitions), and who is responsible for it (data domain owner).” Yassine Laadidi, COFOMO
The signs are often already there
You do not necessarily need to launch an AI project to discover that your data has certain weaknesses.
Instead, ask yourself a few questions:
- Do your teams regularly correct data manually before using it?
- Do two departments get different results for the same indicator?
- Are there multiple sources for the same information without a clearly identified official source?
- Is your data poorly documented?
- Is essential information frequently missing?
These situations are already familiar in business intelligence: conflicting KPIs, manual corrections, problematic data loads, and a loss of trust in dashboards.
AI does not necessarily create these problems, but it can amplify them.
All AI depends on data
The consequences also vary depending on the use case.
Predictive AI relies heavily on the quality and representativeness of the historical data used to develop its models.
Generative AI, particularly when using a RAG (Retrieval-Augmented Generation) architecture, depends on the quality of the information it is given access to.
And with agentic AI, the stakes go even further: incorrect information can influence not only a response, but potentially a decision or an action carried out by the agent.
The more AI is involved in an important decision, the higher the expected level of trust in the data should be.
Should you wait until your data is perfect?
No, and fortunately so!
Trying to make all of an organization’s data flawless before getting started is more likely to turn your initiative into a never-ending project.
The most pragmatic approach is instead to start with a specific use case.
What data is actually required? Is it reliable enough? Where does it come from? Is it representative? And what would the consequences be if it were wrong?
A prototype or pilot project, validated by employees who understand the business context well, can help uncover certain gaps before going any further.
“It is better to work with a limited amount of high-quality data than with a large amount of unreliable data.” Sébastien Fleury, COFOMO
Before asking AI, question your data
This is where data and AI governance become particularly important.
Documenting data sources and their transformations, clarifying responsibilities, establishing common definitions, and measuring the quality of critical data help build stronger foundations for AI.
Before your next project, Yassine Laadidi recommends starting with five questions:
- Do we have the right data for our use case?
- Do we know where it comes from?
- Is it sufficiently reliable and representative?
- Do teams agree on what it means?
- Do we know who is responsible for it?
The goal is not to have perfect data everywhere.
It is to know which data truly matters for your project, what condition it is in, and how much you can trust it.
Because before deciding what your AI can do, it is better to know what it can truly rely on.


