Executives have poured tens of billions into generative AI over the past two years, but the payoff has been shockingly disappointing. A new MIT report, The GenAI Divide: State of AI in Business 2025, finds that 95% of pilots are failing to deliver any financial return.
The authors Aditya Challapally, Chris Pease, Ramesh Raskar, and Pradyumna Chari describe a deep split: a handful of companies are capturing millions in measurable value. At the same time, the overwhelming majority remain stuck in costly experiments with no measurable profit and loss (P&L) impact.
Follow THE FUTURE on LinkedIn, Facebook, Instagram, X and Telegram
The report calls this the GenAI Divide. Most companies rely on generic tools that raise individual productivity but fail to shift financial performance. By contrast, the few achieving results are embedding “agentic AI,” systems that learn, remember, and adapt, directly into their workflows. These organizations are already seeing measurable benefits, including procurement savings, leaner back-office operations, and stronger customer retention.
So what’s the authors’ conclusion? Generative AI is not about buying the latest model. It is about the system’s capacity to learn within business workflows, adapting to feedback, remembering context, and improving over time. Systems that cannot “amplify insight” or “capture value” fade quickly, while those that integrate deeply and improve with use establish a durable advantage.
The MIT findings reflect a global pattern, but their implications are not limited to the U.S. or Western Europe. Cyprus’s economy, with players from both international enterprises and family-run SMEs, is just as exposed to the GenAI Divide. For Cyprus, the real question is which companies will turn AI pilots into profit-and-loss impact, and which will stay stuck experimenting while competitors move ahead.
Why 95% of AI Pilots Fail to Deliver
MIT’s The GenAI Divide: State of AI in Business 2025 opens with a headline figure that captures the scale of the problem: despite an estimated $30–40 billion invested by enterprises, 95% of organizations are seeing no measurable profit and loss (P&L) impact from their generative AI initiatives. The report describes how adoption is widespread, but most initiatives stall at the pilot stage rather than progressing to measurable impact on P&L. Tools like ChatGPT and Microsoft Copilot are widely deployed, often boosting individual productivity, yet they rarely deliver the kind of integrated results that show up in company accounts.
The report identifies four recurring patterns that define this GenAI Divide:
- Limited disruption: out of nine major sectors studied, only technology and media show signs of structural change.
- Enterprise paradox: large firms are running the highest volume of pilots, but are the least successful at scaling them.
- Investment bias: budgets flow disproportionately to visible, top-line functions such as sales and marketing, even though the back office often holds the higher return.
- Implementation advantage: Organizations that partner externally report double the success rate of those trying to build systems in-house.
The core barrier, the report concludes, is learning. Too many generative AI systems cannot retain feedback, adapt to context, or improve with use. This leaves enterprises with brittle workflows that stall at the pilot stage. In contrast, the minority of organizations moving ahead insist on process-specific customization and judge tools by financial outcomes rather than by model benchmarks. Vendors that meet those expectations are already securing multimillion-dollar deployments within months.
Even here, the report stresses that transformation does not necessarily mean mass layoffs. People are not losing their jobs overall, but selective impacts are appearing in functions such as customer support, administrative processing, and software engineering. The more significant savings come from cutting spending on business process outsourcing (BPO) and external agencies, alongside improvements in customer retention and sales conversion. In short, the findings show that when systems are designed to learn and integrate into workflows, they can deliver measurable business value without requiring wholesale restructuring.
Everyone’s Testing, Almost No One’s Winning
Generative AI may be everywhere, but business impact is almost nowhere to be found. Across industries, adoption rates are high: companies of all sizes are experimenting with chatbots, copilots, and automated content generation. These deployments, however, rarely move beyond the pilot stage. The report describes a familiar pattern: leaders approve pilots, vendors showcase proofs of concept, productivity bumps are noted, and then projects stall — stuck in what the report calls “costly experiments.” What is missing, the report argues, is the step from experiments to workflows where systems can actually “amplify insight” and “capture value.”
Large enterprises, paradoxically, are the most affected. They are running the highest number of pilots but struggling to scale them into production. They quote complex governance, legacy IT systems, and risk-averse compliance structures as the reasons their results often fail to show up in company accounts. By contrast, smaller and more agile firms sometimes move faster, but many still lack the capital and technical expertise to achieve a durable impact. The outcome is the same: high adoption, low transformation.
This limbo state is not because GenAI lacks potential. Instead, it reflects how organizations approach the technology. Too often, businesses start with generic tools designed for individual tasks, but fail to adapt them for context-specific processes. Without this tailoring, workflows remain brittle and unsustainable.
The report’s data shows just how uneven the divide has become. While a handful of firms are already deploying “agentic AI” that learns and adapts within core operations, most others remain stuck in what the authors call “costly experiments.” The lesson is that scale and adoption alone do not create advantage. Without learning systems that integrate into actual workflows, even the best-funded pilots risk producing little more than temporary respite.
The Right Side of the GenAI Divide: Where Systems Learn
On the other side of the divide, a small group of companies is demonstrating what works. These companies do not have bigger budgets or newer models, but they are a step ahead in their systems that learn. The researchers describe how these firms are embedding “agentic AI” into workflows, systems that remember context, adapt to feedback, and improve over time. This capacity to learn within business processes is what converts pilots into measurable business gains.
What sets these companies apart is how they measure success. Instead of focusing solely on benchmark scores, they judge outcomes by financial performance. Procurement platforms that learn from contract data deliver savings. Customer-service tools that adapt to user feedback improve retention. Sales engines that refine their pitch over time raise conversion rates. In each case, the system is woven into the workflow, creating measurable gains rather than short-lived productivity gains.
A second feature is design. These systems are built with a “human in the middle” approach, where managers and specialists remain active in the loop. They actively supervise outputs and provide corrections, allowing the system to learn and steadily increase in reliability and relevance. This human oversight not only prevents errors but ensures that the system gets better with use, steadily increasing its relevance and reliability.
The payoff is already visible. While 95% of firms remain stuck in experiments, those on the right side of the divide are reporting multimillion-dollar savings and stronger revenue streams. These are the companies that the researchers say have “crossed the GenAI Divide,” turning brittle pilots into adaptive systems that embed learning and deliver measurable returns.
Where Value Is Emerging
If most companies are stuck, where are the gains being made? The report highlights a set of use cases where agentic systems are already demonstrating measurable returns.
- Procurement and supply chain management: Agentic systems that learn from contract histories and supplier performance data are cutting costs in negotiations and flagging inefficiencies in sourcing. As these savings accumulate, procurement emerges as one of the first areas to show measurable value.
- Back-office operations: Firms are embedding learning systems into finance, HR, and IT workflows. Here, the payoff comes less from replacing people and more from automating repetitive steps, reducing error rates, and allowing staff to focus on higher-value work. The result is leaner operations without workforce displacement.
- Customer-facing processes: Adaptive AI in call centres, marketing, and sales is boosting retention and conversion. Systems that personalise responses and refine recommendations over time are strengthening customer relationships, transforming one-off contacts into repeat business.
- Innovation workflows: In pharmaceuticals, media, and the creative communication sectors, AI is speeding up iteration cycles by generating first drafts, learning from feedback, and improving with each round. This allows teams to test more ideas at lower cost, expanding creative possibilities and bringing products to market faster.
What connects these examples is not the specific sector but the design principle. This is where the 5% of successful firms are separating themselves from the rest, not by putting all their eggs in the generic productivity tools basket but by embedding adaptive systems that consistently deliver returns.
Misconceptions Holding Businesses Back
If the GenAI Divide is so wide, what keeps most companies on the wrong side? The researchers argue that the problem is not only technical but also conceptual. Businesses are holding onto misconceptions that flatten investment decisions and keep value out of reach:
- “It’s all about the model.” Many executives still believe that choosing the most powerful or most expensive model is the key to returns. The report disagrees: outcomes come from how systems are embedded into workflows, not from model selection alone.
- “AI works like plug-and-play.” Organizations often assume that adopting generic tools will naturally translate into business impact. This is not quite true. These tools may raise output for individuals but fail to improve financial performance. Without integration into processes where the system can learn and adapt, pilots remain marginal.
- “Cost savings mean job cuts.” The researchers are clear that the gains they observe do not come from reducing headcount. They come from error reduction, process efficiency, and higher-quality work. Framing AI adoption as a threat to employment can therefore blind firms to its real value.
- “Benchmark scores equal business results.” Some leaders overestimate the value of models that perform well on academic tests or vendor benchmarks. Benchmarks rarely capture the messy, feedback nature of real workflows. As the authors put it, “performance in the lab is not the same as performance in the P&L.”
Crossing the GenAI Divide in Cyprus
Generative AI is not failing everywhere, but it is failing almost everywhere. Ninety-five percent of pilots bring no return, while a small minority convert agentic systems into measurable business impact. The difference lies not in budgets or access to cutting-edge models but in how deeply systems are embedded into workflows.
For Cyprus, this divide is no less real. The island’s economy is unusually diverse, with players from international enterprises in shipping, finance, and professional services alongside family-owned SMEs in tourism, retail, and real estate. Multinationals operating in Cyprus must decide whether their local operations will follow global best practices or remain testing grounds. Smaller firms face a parallel challenge: whether to keep experimenting with generic tools or to invest in embedding AI directly into their processes, even at a modest scale.
Success will not come from chasing the newest model or deploying tools in isolation. It will come from building systems that learn within workflows. The choice is whether to remain among the 95% stuck in pilots, or to cross the GenAI Divide and join the few already capturing value.
In a competitive region where margins are tight and talent is mobile, the companies that take that step first will set the pace for Cyprus’s business landscape.
