Key takeaways
- AI only reflects the data you feed it — quality, representativeness and integrity beat model size.
- Data quality includes security: encrypt connections and protect data on public networks.
- OSINT and platform-provided signals are underused data assets that drive smarter decisions.
- Sequence AI projects data-first: audit, clean, secure, then govern continuously.
- A modest model on clean data beats a cutting-edge model on chaotic data.
There's a quiet truth behind every AI success story: the model is rarely the differentiator. The donnees feeding it are. Companies spend fortunes chasing the latest large language model while the data underneath remains fragmented, unsecured, or riddled with bias. From where we sit at Evocative, the businesses winning with automation aren't the ones with the flashiest models. They're the ones who treat data as an asset to be cleaned, protected and governed.
AI only reflects the data you give it
An AI system learns from what it's fed, and from the choices of the people who built it. If your training data represents only a slice of your customers, your outputs will faithfully reproduce that narrow view. The uncomfortable question isn't "how smart is the model" but "who and what does the data actually represent?" A powerful model trained on incomplete or skewed donnees simply makes bad decisions faster and more confidently.
This is why we tell clients to invest in their data foundation before their model selection. Swapping GPT-4 for a newer version won't fix hallucinations caused by contradictory records, duplicate customers, or missing context. Clean, well-labelled, representative data will.
The security dimension: data quality includes data protection
Data quality isn't only about accuracy. It's about integrity and confidentiality too. A dataset that leaks is a liability, not an asset. This is especially acute for freelancers and small teams working from cafes, coworking spaces and hotels, hopping across public Wi-Fi networks. Every unprotected connection is an opportunity for client data to be intercepted.
The good news is that the baseline protections are now cheap and simple to deploy:
- Encrypt connections on shared networks. A VPN is no longer an enterprise luxury; solutions start at under two euros a month and take minutes to set up.
- Separate professional and personal usage. Freelances juggling client accounts, online purchases and personal logins should compartmentalise to limit exposure.
- Make security frictionless for hybrid teams. If protection is hard to use, people bypass it. Choose tools your team will actually adopt.
These aren't glamorous, but they protect the raw material your AI and analytics depend on.
The data you ignore: OSINT and public signals
Most companies obsess over locking down sensitive internal data while completely underestimating the strategic value of information already public. Open-source intelligence (OSINT) — what's visible about your company, competitors and market in the open web — is a goldmine that too few exploit. Attackers use it. Smart competitors use it. Your business should too, both to understand your exposure and to enrich your decision-making.
The same logic applies to the emerging category of platform-provided data. Publishers like Le Monde now receive deterministic donnees from OpenAI showing how their content performs inside ChatGPT. This is a preview of a wider shift: as generative engines become discovery channels, the businesses that capture and act on this new signal layer will own the GEO (Generative Engine Optimisation) advantage. Again, it's not the model — it's who has access to, and knows how to read, the right data.
What a data-first strategy looks like in practice
When we build automation and AI projects, we sequence the work deliberately:
- Audit first. Map where data lives, how accurate it is, and who can access it. You cannot improve what you haven't inventoried.
- Clean and consolidate. Deduplicate, standardise formats, fill gaps. This unglamorous step delivers more ROI than any model upgrade.
- Secure by default. Encrypt in transit and at rest, enforce access controls, and give mobile workers the tools to stay protected anywhere.
- Govern continuously. Assign ownership, monitor bias, and keep enriching with legitimate external and platform signals.
Models will keep getting better on their own — that's the vendor's job. Your job is the data. Get that right, and even a modest model outperforms a rival's cutting-edge system fed on chaos. That's where durable competitive advantage actually lives.
Sources
- Entre clients, achats en ligne et comptes personnels, Surfshark VPN sécurise la vie numérique des indépendants — Journal du Net
- Travail hybride, Wi-Fi public, données sensibles : CyberGhost VPN simplifie la sécurité des petites équipes — Journal du Net
- Réseaux publics, risques réels : Proton VPN apporte une réponse simple — Journal du Net
- L'IA apprend de nos données. Qui apprend à l'IA ? — Journal du Net
- "Voici ce que Le Monde peut savoir pour son GEO avec les données déterministes fournies par OpenAI" — Journal du Net
- Pourquoi les entreprises sous-estiment encore l'OSINT ? — Journal du Net
