The RAISE Summit in Paris has become a significant meeting point for the AI ecosystem and the wider data industry. The two-day Summit in July 2026 brought together more than 9,000 participants, over 2,000 companies and approximately 350 to 360 speakers. More than 80% of the attendees are either founders, executives or senior decision-makers. Across the wider RAISE Week, participation exceeded 15,000.
The significance of the event, however, was not simply its scale. What stood out this year was the absence of a single, agreed-upon vision for the future of AI.
Speakers presented different perspectives on what comes after today’s large language models (LLMs), who should control the technologies and infrastructure on which enterprises increasingly depend, how organizations should assess the economics of AI and where human judgment must remain central.
5 Decisions that Will Prepare You for the Next Phase of AI Adoption
1. Do not design the enterprise around today’s LLMs alone.
One of the most interesting disagreements at RAISE concerned what comes after the current generation of LLMs.
Yann LeCun, Exec Chairman and co-founder of AMI Labs, former Chief AI Scientist at Meta, argued that language models can manipulate language without possessing the physical intuition of a child or even a housecat. His proposed alternative is the development of “world models:” systems that learn how the world behaves and use that understanding to predict, reason and plan.
The immediate implication is not that enterprises should move away from generative AI or attempt to predict which technical approach will ultimately prevail. It is that today’s generative-AI architecture should not automatically be treated as the final destination. This is particularly relevant in manufacturing, robotics, mobility and other physical environments where AI systems may need to understand relationships between objects, anticipate consequences and operate within real-world constraints.
Enterprises should therefore consider whether their current architectures can accommodate future model types and methods of intelligence. The strategic advantage will not necessarily come from choosing the eventual winning model in advance but from avoiding architectural decisions that make future adaptation unnecessarily difficult.
2. Treat AI sovereignty as operational control.
Arthur Mensch, co-founder of Mistral AI and Mozilla’s Mark Surman how to position sovereignty as more than a compliance, regulatory or data-residency issue. The broader question was who ultimately controls the models, infrastructure and economics on which companies are becoming dependent.
This distinction matters. An enterprise may know where its data is stored while still having limited influence over model availability, pricing, portability, infrastructure resilience or the conditions under which a service continues to operate.
The irony that a discussion about open-source resilience continued on stage through a power interruption provided a useful reminder: sovereignty without resilient infrastructure remains largely theoretical.
For enterprises, AI sovereignty should therefore be considered in operational terms. It is not only a question of where data resides, but also of how much control the organization retains over the technologies, providers and infrastructure supporting critical AI capabilities.
As AI becomes more deeply embedded in business processes, enterprises will need to understand where they are creating dependencies and whether they retain sufficient options if technical, commercial or operational conditions change.
3. Measure the cost per successful AI outcome.
The infrastructure debate at RAISE became increasingly concrete. Sessions addressed inference clusters, heterogeneous hardware, energy-first cloud design and the financing of AI infrastructure. Together, these discussions showed that enterprise AI performance can no longer be evaluated by looking at the model alone.
The more relevant question is which combination of model, data location, chips, energy and network capacity can produce the required result at an acceptable unit cost.
For enterprises, this means moving beyond model accuracy as the dominant measure of success. A model may perform well in isolation but still prove expensive, difficult to scale or operationally inefficient once infrastructure, human intervention and unsuccessful outputs are considered.
A more useful metric is the cost per successful “AI outcome.” This shifts attention from the theoretical performance of the model to the economics of completing a business task successfully, such as resolving a customer request, detecting a defect or approving a transaction, to the required standard of accuracy, speed and compliance. It encourages organizations to consider the full system required to generate that outcome, rather than optimizing one technical component in isolation.
Over the next 12 months, greater transparency around these economics will become increasingly important. Enterprises will need to understand not only whether an AI system works, but whether it can deliver reliable outcomes at a cost the organization can sustain.
4. Design trust into the AI operating model.
A roundtable titled “The Trusted AI Enterprise: Driving Productivity at Scale, Securely” focused on the relationship between trust, productivity and enterprise-scale deployment.
The central takeaway was that trust cannot be added after deployment as a separate governance layer. It must be designed into the operating model in which the AI system functions. And this design needs to be initiated before the PoC. That requires organizations to answer several practical questions before systems are deployed at scale:
What data may the system access?
What actions is it permitted to take?
At what point must a person intervene?
How will the organization demonstrate that the system continues to behave as intended?
These are not abstract governance questions. They determine how AI operates within real business processes.
As AI systems become more capable and gain access to more data and workflows, trust will depend on clearly defined permissions, controls, intervention points and evidence. Policies alone will not be enough if they are not reflected in the design and operation of the system. The organizations that scale AI successfully will be those that treat trust as a property of the operating model rather than as a review performed after implementation.
5. Preserve human accountability as AI gains authority.
RAISE presented several competing visions of the future. LeCun described forms of intelligence that extend beyond language. Mensch argued for more open and controllable AI ecosystems. Mark Cuban, billionaire tech entrepreneur, Shark Tank investor and co-founder of Cost Plus Drug, emphasized the continuing importance of human judgment and consequences. This diversity was more valuable than another prediction that every company will simply deploy more copilots.
The next phase of enterprise AI will not only involve systems that generate information. It will increasingly involve systems that recommend decisions, initiate processes and take actions. The key question is therefore not simply what AI agents are technically capable of doing. It is what authority organizations are prepared to give them and who remains accountable for the consequences.
Enterprises will need to define the permissions given to agents, the circumstances in which human intervention is required and the individuals or functions responsible for outcomes. Human involvement should not be treated as a general safeguard added to every process. It should be designed deliberately around the level of risk, authority and consequence associated with each use case.
As agentic capabilities expand, clarity of accountability will become as important as technical capability.
4 questions for the next 12 months
RAISE 2026 was less about one dominant AI future than several competing ones. A concern discussed at lengthwas recursive self-improvement (RSI): the possibility that AI systems could increasingly contribute to improving their own tools, training processes and successor models, creating capability gains that may outpace existing methods of evaluation, oversight and control.
Enterprises do not need to determine today which technical or market vision will ultimately prevail. They do, however, need to make decisions that preserve their ability to respond as the landscape develops.
Over the next 12 months, enterprise leaders should pay particular attention to four questions:
What is our cost per successful AI outcome?
How much authority are we giving AI agents, and who remains accountable?
Where are we exposed to infrastructure or provider concentration?
Can our architecture accommodate approaches beyond today’s large language models?
The winning enterprise AI strategy may not be based on selecting one vision of the future. It may depend instead on building sufficient adaptability, operational control, economic transparency and trust to operate across several possible futures.
ISG helps enterprises navigate a rapidly changing AI market and decide how best to prepare. Contact us to find out how we can help.
About the author
