Contact us

Almost three years after the hype : where is the adoption of AI in business?

AI : Between dazzling promises and strategic uncertainties

Since the emergence of ChatGPT at the end of 2022, artificial intelligence has become an obsession for boards of directors. CEOs are demanding that CIOs “integrate AI into processes,” “assist business lines in their AI efforts,” and “quickly deliver value.” Generative AI is no longer an innovation gimmick, but a symbol of competitiveness. In a climate of economic slowdown, it is seen as an essential… and immediate lever.

For CIOs, the pressure is immense : adopting AI can make or break their careers within the company. Like the web twenty years ago, it is emerging as an unavoidable revolution, with this difference: expectations are much higher and the timeframes for delivering results are much shorter.

Three years later, the time has come to ask ourselves: Is AI truly delivering on its promise for the company? Are the costs incurred sustainable and justified by a tangible return on investment? Should we build our own AI internally or rely on market solutions ? How far can we go with sensitive data without compromising the organization’s sovereignty? And finally, how should we manage this adoption: centralize or let business lines experiment, regulate shadow AI or encourage it?

In this article, we will cover :

  • Key figures from MIT and Gartner, which measure the gap between expectations and reality.
  • The main obstacles to adoption: costs, lack of ROI, reliability, organizational integration.
  • The dilemma of in-house AI versus market solutions, and its implications for CIOs.
  • Possible attitudes toward this revolution: strategic partnerships, governance, shadow AI, composite AI, and the role of professions.

Figures that dampen enthusiasm

Three years after the hype began, the results are far more nuanced than the initial promises. Two major studies allow us to measure the gap.

  • According to MIT (NANDA Initiative, 2025) :
    • Only 5% of AI projects generate real revenue acceleration.
    • 95% stagnate or fail, with no tangible impact on the income statement.
    • Success stories are concentrated in young, agile startups that choose a specific problem, execute quickly, and rely on strategic partnerships.
  • According to Gartner :
    • Generative AI has moved from the peak of inflated expectations to the pit of disillusionment of its hype cycle.
    • The main obstacles : hallucinations, unreliability, and exorbitant energy costs.
    • Gartner estimates that it will take two to five years to reach the “slope of enlightenment” and create value at scale. 

Massive obstacles : exorbitant costs and lack of ROI

The major obstacle to AI adoption today is economic. The costs associated with AI are colossal, and the returns on investment are still very uncertain.

  • Infrastructure costs : Training and running advanced models requires enormous energy and computing resources. Some companies report multi-million dollar bills for applications that are still experimental.
  • Human costs : Building an in-house team of data scientists, AI engineers, and specialized architects has become a luxury reserved for giants.
  • Lack of ROI : Most pilot projects fail to prove their financial value, creating growing frustration among management committees.

The example of OpenAI illustrates this reality : despite global adoption and global recognition, the company is still recording abysmal financial losses. If the iconic industry leader is struggling to achieve profitability, how can we expect its corporate clients to quickly generate a measurable return? This equation calls into question not only the viability of providers’ business models, but also the sustainability of the technology itself in its mass adoption.

Another obstacle : reliability and organizational integration

Beyond costs, another major obstacle lies in the reliability of the models and their integration into the company.

  • Hallucinations and inconsistent results : Generative AI remains capricious and cannot yet be deployed in environments where precision is critical. Luc Julia, Chief AI Officer at Renault and co-creator of Siri while working at Apple, provides a striking demonstration: when a model is asked the binary question of Victor Hugo’s date of birth, it receives different answers depending on the day. This simple example raises profound questions about the degree of veracity and reliability that can be expected from these systems.
  • Shadow AI : Uncontrolled use of tools like ChatGPT by employees exposes organizations to compliance, privacy, and cybersecurity risks.
  • Workflow rigidity : While ChatGPT is effective for individual use, its integration into complex business processes faces structural limitations..
  • Poor budget targeting : More than half of AI budgets are still directed toward sales and marketing, while MIT observes that the real ROI lies in back-office automation and streamlining operations.

Homegrown AI versus market solutions : two approaches, two results

The MIT survey highlights a stark contrast. Companies that have attempted to develop their own “homegrown” AI (often an internal ChatGPT trained on their data) have, by and large, failed. Infrastructure, labor, and R&D costs have proven prohibitive, and results have rarely been forthcoming.

Conversely, those that have adopted off-the-shelf solutions are achieving far greater success. This choice sometimes means giving up some data sovereignty, but it quickly yields tangible gains.

To summarize : CIOs that succeed in adopting AI will be, above all, very good buyers (capable of identifying the right suppliers and negotiating the right terms) rather than very good builders building sophisticated tools from scratch.

In this context, a rule of caution is essential: avoid at this stage conducting AI projects on highly sensitive data, which expose the company’s sovereignty and security. Until the organization has built solid experience with less critical use cases, priority should be given to areas where the risk is manageable and the benefits quickly measurable.

What attitude should you adopt ?

To succeed in this context, CIOs must adopt a lucid, methodical and strategic stance.

  • Prioritize pragmatic use cases : back-office automation, reducing the need for outsourcing and streamlining administrative processes remain the most likely areas for rapidly generating value.
  • Exploring “AI composite” : It is not a question of betting everything on a single model, but of orchestrating several technological building blocks (LLM (language models), traditional machine learning, computer vision, autonomous agents) in order to compensate for the weaknesses of each isolated approach.
    Concretely, composite AI is based on a simple idea: no AI technology is perfect in and of itself. LLMs excel at text generation but suffer from hallucinations. Machine learning is robust at prediction but limited in contextual understanding. Computer vision is very effective at image recognition, but lacks reasoning capabilities. Finally, agents provide autonomy but still lack reliability. By combining these techniques in a composite architecture, companies are able to leverage the strengths of each while minimizing their weaknesses.
    Many companies are already switching to this multi-technology approach. In retail, an internal search engine can combine an LLM (to understand the query) with an ML algorithm (to recommend products) and a vision module (to identify an item in a photo). In the supply chain, warehouse optimization can rely on forecasting models, logistics algorithms, and a chatbot to assist operators.
    Relying on a single building block is a strategic dead end. Composite AI provides robustness, adaptability, cost control, and scalability. This is the key to moving beyond the “permanent POC” and industrializing AI.
  • Building through strategic partnerships : choosing a key vendor (whether Mistral AI, Anthropic, Google, or other strategic players) cannot be treated as a simple IT decision. It is a corporate governance issue. The selection of an AI partner must be negotiated with senior management, integrated into the overall strategy, and validated by the legal department to anticipate the implications of sovereignty, confidentiality, and technological dependency.
  • Decentralize adoption : AI should not remain confined to a central “AI Lab.” It must be driven by the businesses themselves. Field managers know their processes and their challenges better than anyone : giving them control over experimentation is the best way to prevent AI from remaining an off-the-ground innovation.
  • Managing Shadow AI intelligently : It would be unrealistic (and counterproductive) to try to completely ban employees from using tools like ChatGPT or Claude. Shadow AI is often the first driver of adoption: it reveals curiosity, creativity, and the real needs of businesses. The role of the CIO is not to block these uses, but to channel them, establish safeguards, and create a secure framework that allows these individual experiments to be transformed into collective practices.
  • Strengthen governance and cost control : AI must be framed as a strategic resource. Who can use it, on what data, with what limits, and at what cost? These decisions require governance committees bringing together IT, legal, business, and senior management to steer the transformation in a responsible and sustainable manner.

Untouched potential, but not at the cost of illusion

Both MIT and Gartner emphasize that despite the current disillusionment, the potential of AI remains immense. With more reliable models, better prepared data, and more rigorous evaluation frameworks, generative AI will permanently transform productivity and innovation.

But we must put an end to the illusion of immediate profitability. AI is an expensive, demanding, and uncertain technology. For CIOs, the real challenge is not to “do like everyone else” but to transform an exciting but capricious promise into a sustainable performance driver. Those who can orchestrate AI with discernment, patience, and method will be the real winners in the coming years.

About the author

Ismail has 15 years of experience in IT and digital consulting. He spent nearly 7 years at Gartner. He has supported innovative startups in their growth strategy and worked with CIOs of large groups on their digital transformation. In 2021, Ismail founded Hubadviser to help CIOs challenge their vision with top-level experts.