The year 2024 has reshuffled the technical cards on several fronts simultaneously. Between the European regulatory framework coming into effect, RAG architectures moving from prototype to production, and quantum computing emerging from laboratories, companies’ technological choices have become more complex. Here, we analyze the structuring axes that are redefining stacks and strategies.
RAG Pipelines in Production: The Technical Challenge Behind Generative AI in Business
The conversation around generative AI has shifted. The issue is no longer access to language models, but their reliable connection to internal data. IT departments that have moved beyond the pilot stage in 2024 are working on industrialized RAG pipelines, meaning architectures where a vector search engine queries document databases (ERP, CRM, PLM, SharePoint) before injecting the relevant context into the prompt sent to the LLM.
Further reading : The Latest Trends in Kitchen Utensils
Open source components like LangChain, LlamaIndex, or Haystack form the foundation of these processing chains. On the managed services side, Azure AI Search and Vertex AI Search are capturing an increasing share of enterprise deployments. Several major European industrial and banking groups have announced pilots that have moved into production over the year.
We regularly find in Starlight Infos’ tech articles detailed analyses of these hybrid architectures and their sectoral adoption.
You may also like : The best outing and activity ideas to discover near you
The main challenge remains the quality of retrieval. A poorly calibrated vector index returns document noise to the model, degrading the final response. The performance of a RAG system depends more on indexing than on the model itself. Data engineering teams are therefore investing heavily in semantic chunking, re-ranking, and metadata filtering strategies.
European AI Act and Obligations for General-Purpose Models
The European regulation on AI (AI Act), formally adopted in spring 2024, directly targets providers of generative AI models deemed general-purpose. The obligations focus on several axes:
- Transparency regarding the datasets used for training, with documentation of sources and filtering procedures
- Explicit management of copyright, a major friction point since French and German press publishers have initiated legal actions or concluded licensing agreements in 2023-2024
- Mandatory labeling of generated content in certain use cases, particularly misinformation and deepfakes
- Documentation of systemic risks for models exceeding a certain capacity threshold
This regulation reshapes the business model of large language models. Compliance costs (dataset audits, license management, labeling) are passed on to commercial offerings. Free public versions could become limited in the medium term.
For user companies, the impact is concrete: any deployment of a high-risk AI system now requires documented compliance assessment. Legal departments and data teams must work together from the design phase.
Quantum Computing and Post-Quantum Cybersecurity
Quantum computing reached a new level of visibility in 2024, not through a spectacular breakthrough in the number of qubits, but through the structuring of an industrial ecosystem. The concept of cyberquantum summarizes the convergence between quantum computing power and threats to current cryptographic systems.
The risk is identified: asymmetric encryption algorithms (RSA, ECC) on which the majority of secure exchanges rely become vulnerable to a sufficiently powerful quantum computer. Data encrypted today and intercepted by an attacker could be decrypted in a few years, a strategy known as “harvest now, decrypt later.”
We observe that major companies in the banking and defense sectors have begun migrating to post-quantum algorithms as early as 2024. The cryptographic transition cannot wait for the arrival of a large-scale functional quantum computer. The recently published NIST standards accelerate this migration.
Priorities for IT Departments
The cryptographic inventory is the first step. Identifying where RSA and ECC are used in the infrastructure, mapping certificates, and then planning the migration to resistant algorithms. This technical project mobilizes security, network, and application teams simultaneously.
GreenTech and Digital Sobriety: Constraint or Technical Lever
The pressure on the environmental footprint of digital technology translated in 2024 into concrete requirements. Sustainable technologies are no longer just part of CSR discourse: they become a selection criterion in tenders and supplier audits.
The paradox is direct. The training and inference of generative AI models consume considerable resources in computation and energy. Data centers are seeing their electricity consumption increase while companies’ climate commitments tighten.
Technical responses involve optimizing models (quantization, distillation, pruning) and choosing less resource-intensive architectures. Compact and specialized language models are gaining ground against massive generalist models, precisely because they offer a better performance-to-consumption ratio for targeted use cases.
Private 5G also plays a role in this equation. By bringing data processing closer to its source (edge computing), it reduces back-and-forth trips to the central cloud. Industrial deployments in factories and logistics are accelerating.
The technological landscape of 2024 is characterized by this productive tension between increased system power and regulatory, energy, and security constraints. Companies that manage to excel are those that approach these three dimensions as a single architectural problem, not as siloed subjects.