IA Act & data reliability

Understand, act, measure: launch your AI projects without immobilizing your data

Since the entry into force of the IA Act, data is no longer just fuel for use cases: it is a critical asset What is needed qualify, trace and explain. Many teams conclude that it is necessary to wait for perfectly clean data before starting. In the field, it's often the other way around: Well positioned, AI makes data reliable. It detects discrepancies, brings references closer together, enriches what is missing, and documents each step. The human keeps the referee; the AI boosts and secures.

Sovereignty and reliability: two sides of the same requirement

Make reliable, is to transform a heterogeneous heritage into usable and auditable base. Concretely:

• correct errors and remove duplicates;
• harmonize formats and standards;
• fill in the missing data;
• keep the evidence of each correction.

sovereignty, it's keeping the mastery : where your data circulates, who accesses it, how they are processed, according to which rules and with what reversibility.
The link between the two is transparency (journals, decisions, model versions). Reliable and traceable data is sovereign data.

Do you need a “threshold” to get started?

It does not exist No magic percentage. The real prerequisite is organizational : to know where is your data, To what they serve, and dispose of referential, even partial.

Field example. A major manufacturer managed a product base of 600,000+ references with 60% completeness. In a few weeks of iterations with AI — under business control — the base grew to 100%. The challenge is not initial perfection, but to target intervention areas and to measure improvement.

Where should AI be placed in the reliability flow?

At Strat37, we start from a simple principle: “Where a human can describe the task, AI can automate it.”
In practice:

• When swallowed (invoices, PDFs, emails): the AI reads, extracts, normalizes; the code imposes the safeguards (dates, amounts, units).
• Semantic approximation : matching between heterogeneous systems, justification of decisions, ambiguous cases traced back to the professions.
• Grey areas : specialized agents classify, complete and remove ambiguities, in several passes if required.
• Ongoing monitoring : far from the “one-shot”, the AI detects drifts, re-harmonizes and makes massive corrections throughout the life cycle.

Operating method: start without waiting for perfection

The progress is following a clear and defined pace:

1. Map & align shared sources, owners and definitions (business/technical).
2. Control formats, uniqueness and referential integrity, with targeted inspection critical tables.
3. Orchestrate code + AI : Python/SQL for the automatable; LLM/agents for context cases.
4. Measure continuously: completeness, precision, reminder, F1 score, duplication rate, Trust score — and compare To one Gold Dataset (untouched base) to objectify progress.
5. Stabilize & trace : log prompt, input/output and arbitrations; transparency indispensable under the IA Act.

Field returns (Strat37)

Our findings are based on more than 20 million data processed, structured (ERP, CRM, PIM, CSV...) and unstructured (invoices, PDF, emails).

• Industry (cleaning & complex classification): based on technical and multi-reference data, our models exceed 80% accurate on a very large scale — a threshold that allows automating tasks that were previously manual and limited.
• Real estate & asset management (textual classification): on varied but non-technical data, we reach 92% reliability through the structuring, harmonization and systematic AI checks. The documented decision path facilitates audits.
• Distribution & networks (enrichment & inter-systems): 100% product sheets can be completed automatically with targeted validation trades. On the reconciliation between heterogeneous systems, we observe ≈ 90% of reliable matching ; the remaining cases, which are intrinsically interpretative, are arbitrated by the professions.

Tangible results in less than 60 days

Without immobilizing your projects, you get:

• A base more complete, consistent, and auditable ;
• Of quality indicators on the rise (completeness, F1, confidence), Duplicates are down, of reduced deadlines ;
• Teams that Take back the hand, and AI models whom perform better.

From reliability to action

Once the foundation is stabilized, we go from Diagnosis To theoperational impact. The AI then supports:

• TheAugmented analysis (correlations, clustering);
• The continuous detection (excesses, fraud, breakdowns);
• The predictions (sales, risks, stocks, churn);
• Thenatural interaction with your data (“chat with your data”).

Operational results feed into measurement, which reinforces reliability: the loop virtuous is in place.

To remember

Place the AI in the right place of the flow, entrust thearbitration to the human, Measure your progress on a Gold Dataset and Trace every step. In this context, compliance is not a barrier: it is an accelerator of sovereignty and performance.