The observation is now shared by data departments, business units, and software vendors: the era of massive accumulation is over. In AI systems, raw data no longer has value in itself. It must be structured, qualified, governed, and situated within a precise context of use. Two converging perspectives demonstrate this: the DQE white paper on Data Quality applied to AI and the CIO Online article titled “When it comes to AI, not all data is created equal” (source).
This convergence illuminates an ongoing transformation: AI does not value all data equally. It selects, prioritizes, exposes weaknesses, and reveals inconsistencies in invisible architectures.
From Volume to Value: Moving Beyond the Myth of “More Is Better”
The data-driven fantasy has long rested on the idea that everything must be stored, everything kept, “just in case.” This belief now collides with an operational principle: it is not volume that creates value, but the ability to identify trustworthy data.
In the DQE white paper, the demonstration is clinical. When customer data is unstable, models struggle to produce reliable recommendations. When data is fragmented across channels, personalization becomes approximate. When duplicates proliferate, indicators lose their meaning. The result: AI projects that fail, POCs that stagnate, and technological promises that never reach industrial scale.
CIO Online confirms this shift. The generative moment is here. The models are ready. But the essential work remains: “sorting, cleaning, ensuring reliability, governing.” Not all data is suitable for training a model. Not all data is legitimate to circulate in an automated organization.
Reliability, Freshness, Legitimacy: A New Data Hierarchy
The practical cases presented by DQE reveal an implicit hierarchy. Identity data (the “golden record”) towers above all others: without reliable linkage, there is neither personalization nor coherent behavioral analysis. An identification error contaminates the entire chain.
Next come transactions (contracts, payments), behavioral signals (navigation, clicks), and unstructured verbatims (tickets, calls, comments). The further we move from the core, the more unstable, noisy, and contextual the data becomes. It’s not that it’s useless. It’s that it requires verification mechanisms, contextualization, and algorithmic accountability.
This is where the convergence with CIO Online becomes fruitful: a data point’s value depends less on its raw accuracy than on its capacity to be used appropriately. Its relevance is a function of use case, timing, user… or AI agent.
AI Agents and Reinforced Governance: Human Permissions Are No Longer Sufficient
The CIO article introduces a rarely addressed point. The risk amplifies when it is no longer humans accessing data, but automated agents. The AI agent doesn’t read as we do: it processes massively, rapidly, without emotional context or implicit discernment.
This demands a profound transformation of governance. “Human” permissions no longer guarantee security or compliance. Control layers adapted to machine usage must be designed. This means: restricted perimeters, exhaustive traceability, explicit access limits, auditability.
At BPCE, this drift is anticipated by blocking errors at the point of entry. The logic is preventive: if an email is misspelled, it must not enter the system. In contrast, Baccarat accepts minimal friction in stores, but then centralizes correction and reliability assurance. Two models, one shared requirement: making data worthy of use.
From POC to Industrialization: AI Does Not Compensate for Shaky Foundations
Several cases analyzed in the white paper show that AI performance cannot compensate for systemic flaws. At ByMyCar, customer data is distributed across six ERPs and nine CRMs. Consolidation in Salesforce, combined with deduplication and enrichment solutions, finally enables reliable customer journeys. But as long as data remains unstable, the generative model has nothing solid to work with.
The same observation at Pledg, a fintech specializing in split payments: scoring happens in one second. Real-time reliability becomes a survival condition. AI is not a layer on top. It’s a vital organ. But it only functions if incoming signals are consistent, fresh, and traceable.
Select, Contractualize, Monitor: A Three-Act Strategy
This cross-analysis leads to a structuring recommendation: thinking about data usage through a contractual value chain.
- Select: define useful perimeters, and exclude redundant, suspect, or unstable data.
- Contractualize: formalize quality requirements (freshness, completeness, traceability) according to use cases and associated models.
- Monitor: track drift over time, because all clean data becomes obsolete if not maintained.
“Quality” ceases to be a fixed property. It becomes a dynamic state, managed, measured, correlated to an intended use. This approach transforms AI into an operational lever, but above all into a credibility factor.
Toward Credible AI: Explainability, Frugality, Sovereignty
The central question, ultimately, is not whether AI works. It’s whether it can be trusted. The white paper rightly cites concerns voiced by L’Oréal, Crédit Agricole, Cosmo Tech, and France Travail. Too many AI projects remain peripheral because core data is poorly governed or legally unclear.
In this context, digital sovereignty becomes a strategic lever once again. The challenge is not to reject cloud or large models. It is to understand what deserves to be automated, and under what technical, economic, and ethical conditions.
