Over the past two years, generative AI has been heralded as a once‑in‑a‑generation technological breakthrough. Since the launch of ChatGPT in late 2022, executives, investors, and policymakers have rushed to proclaim the dawn of a new productivity era. AI, we were told, would rewrite how we work, automate knowledge tasks at scale, and deliver rapid gains in efficiency and growth.
Yet, as the initial excitement settles, a more sobering reality is emerging. Despite massive investments, generative AI has so far struggled to deliver the transformative business value many expected.
A cross-country survey of nearly 6,000 senior executives in the US, UK, Germany, and Australia found that most firms report no productivity impact from AI so far. 89% report no impact on labour productivity (sales per employee) over the past three years, and executives’ own usage averages about 1.5 hours per week. Likewise, the “State of AI in Business 2025” report from MIT Project NANDA paints a similar picture from a different angle: it argues that 95% of organisations are getting “zero return,” and that while many companies investigate or pilot GenAI tools, only a small share of task-specific tools reach production at scale.
The question, then, is not whether AI matters, but why its impact has so far fallen short—and how organisations can move beyond the hype.
Even Microsoft CEO Satya Nadella recently acknowledged that AI has yet to deliver meaningful macroeconomic productivity gains. Instead of a revolution, many organisations are seeing incremental improvements, pilot projects that fail to scale, and mounting uncertainty about where real returns will come from.
This pattern is familiar. Technological hype cycles often promise disruption long before institutions, processes, and complementary capabilities are ready. In Europe, commentators have already begun to warn of an AI bubble, drawing uncomfortable parallels to past episodes of technological over‑exuberance.
The question, then, is not whether AI matters, but why its impact has so far fallen short—and how organisations can move beyond the hype.
Why generative AI alone is not enough
One reason for this growing disappointment lies in how generative AI is commonly framed. Much of the public narrative has been shaped by an ideology of replacement: the idea that large language models (LLMs) will substitute human labour across a wide range of tasks. From writing reports to coding software, generative AI has been portrayed as a digital worker waiting to take over.
Technically, however, generative AI is far more limited than this narrative suggests by three key constraints:
- GenAI is a reactive technology: Most GenAI systems respond to prompts; they do not continuously monitor context, set goals, initiate action, or run workflows end-to-end. This makes them powerful as drafting, summarising, or ideation tools, but limits their ability to drive sustained process change without additional orchestration. As a result, their value remains tightly tied to human input and supervision
- GenAI optimises for plausibility, not guaranteed correctness: LLMs can be impressive at producing fluent text, but they can be unreliable in novel or high-stakes contexts because their outputs are based on probabilistic pattern matching rather than semantic or contextual comprehension. In other words, LLMs are trained and rewarded to predict likely sequences, not to provide truthful information. As a result, an increasing number of critics point out that LLMs do not truly understand what they are talking about. In practice, firms must add retrieval, validation, human review, and monitoring to use GenAI safely in core workflows.
- The biggest gains require process redesign, not better prompts: When GenAI is bolted onto a workflow without redesign, there are no clear handoffs, no quality gates, no measurement, and no data integration; the value stays local and informal. Many of these stalled pilots follow a broader pattern. In 2024, Gartner predicted that 30% of generative AI projects would be abandoned after the proof-of-concept stage by the end of 2025, citing weak data foundations, rising costs, and unclear business value. Recent reporting in the Wall Street Journal suggests that this correction is indeed underway, as companies tighten budgets and demand measurable returns before scaling AI initiatives.
In practice, this means that many generative AI applications become productivity add‑ons rather than productivity engines. They help individuals draft faster, summarise better, or brainstorm more efficiently, but they rarely transform how work is organised or how value is created. Moreover, their reliance on statistical regularities reaches clear limits when applied to edge cases or novel situations that are poorly represented in training data. When organisations expect replacement and get assistance instead, disappointment is almost inevitable.
The real challenge, therefore, is not to push generative AI harder, but to rethink how it is embedded into workflows. This requires abandoning the idea that AI must replace people to be valuable. Instead, AI should be understood as a tool that complements human judgment, scales coordination, and reduces friction in complex systems.
Lessons from the “ChatGPT moment” in manufacturing
This shift becomes particularly visible in manufacturing, which is experiencing its own “ChatGPT moment.” Importantly, this moment is not about substituting factory workers or engineers with algorithms.
Instead, AI is being layered onto existing production systems as an additional form of automation. In industrial settings, generative models are used to interpret machine data, support maintenance decisions, assist in design iteration, or translate between different technical systems. The value does not come from autonomy, but from augmentation, from enabling humans and machines to work together more effectively.
This is a critical insight. Manufacturing shows that AI delivers value when embedded in socio‑technical systems, not when treated as a standalone substitute for labour. By serving as a connective layer that links data, processes, and human expertise, AI can unlock efficiencies that neither humans nor traditional automation could achieve on their own.
Seen this way, the limited impact of generative AI so far is not a failure of the technology, but a mismatch between expectations and capabilities. The real opportunity lies one step further.
Agentic AI can unlock value, but only when paired with scoping, governance and automating the right workflow.
Enter agentic AI: from tools to collaborators
If generative AI represents intelligent assistance, agentic AI represents intelligent action (within defined boundaries). Agentic AI systems are designed not just to generate content, but to pursue goals, make decisions, and execute multi‑step tasks autonomously within defined boundaries.
Agentic AI represents the next frontier of enterprise systems. Unlike generative AI, which waits for prompts, agentic systems can initiate actions based on context, coordinate workflows across systems, and adapt over time. At minimum, they combine: a goal (and subgoals), planning over multiple steps, tool access (APIs, databases, enterprise apps), memory/state tracking, monitoring and escalation, and governance controls (permissions, approvals, audit logs).
This distinction matters for value creation. Many of the productivity gains envisioned in the early days of generative AI, such as end‑to‑end automation, faster decision cycles, and reduced coordination costs, require precisely these agentic capabilities. Without agency, AI remains confined to isolated tasks. With agency, it can manage processes.
Agentic systems are most convincing in workflows that are high-volume, repetitive but not identical, governed by clear policies, and measurable end-to-end. Examples include:
- Customer operations: solve issues, retrieve entitlements, trigger approved backend actions (e.g., reset credentials, update accounts), and escalate exceptions.
- IT service management: collect logs, run diagnostics, propose remediations, open/route tickets, and verify resolution.
- Procurement: monitor spend thresholds, draft purchase requests, route approvals, and ensure compliance checks are executed and logged.
This is where the long‑promised AI revolution becomes plausible.
From hype to impact
Agentic AI is also the newest magnet for hype.
Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027 due to escalating costs and unclear business value, and warned of widespread “agent washing,” where ordinary assistants are rebranded as agents without real capabilities.
The implication is clear: Agentic AI can unlock value, but only when paired with scoping, governance and automating the right workflow.
GenAI has already changed how people write, search, summarise, and code, but task-level acceleration is not the same as enterprise transformation. The missing ingredient in many AI revolution expectations is agentic. Or as Satya Nadella has argued, the true promise of AI lies in systems that can reason, plan, and act.
Agentic AI makes that future more credible. But it also raises the bar. The winners will not be the firms that chase the newest model. They will be the firms that redesign workflows, define bounded autonomy, and build governance that allows AI to act safely at scale.
This article is based on the ESCP Business School Impact paper Agentic AI: The Next Frontier in Autonomous Enterprise Systems.
Daniel Völker is the Head of the AI Solutions Department at valantic.
Marc Oberhauser is an Associate Professor of International Business at ESCP Business School.