A recent report from MIT Nanda seems to confirm what many business leaders already feel. Despite large investments in GenAI, most organizations are achieving very little measurable ROI. The report, “The GenAI Divide – State of AI in Business 2025”, concludes that 95% of enterprise AI investments are not delivering expected returns. While around 80% of enterprises have explored or piloted general‑purpose tools like ChatGPT and Microsoft Copilot, only about 5% of custom enterprise AI solutions reach production deployment.
With all the promises of GenAI and agentic AI, where are the productivity gains?
This is not a new phenomenon. We have seen this before. Productivity gains from powerful technologies often arrive with a lag. How big that lag is depends less on the technology itself and more on whether organizations are willing to redesign how they work. To see why, it is useful to start with a historical example.
The Productivity Lag
With the introduction of electricity, the potential applications were obvious. Factories could be illuminated to enable longer shifts. Machines could, in principle, be powered more reliably. Constraints due to coal‑fired steam engines could be removed. Yet by 1900, steam power still dominated. Less than 5% of American factories used electric motors. The technology was available, but the productivity gains were largely absent. Why?
To understand this, we need to understand the structures of steam‑powered factories. In steam‑driven plants, one large steam engine powered the entire factory through a central steel shaft. This shaft connected to smaller shafts, belts, and gears that delivered power to individual machines like hammers and presses. Belts transferred power upward through ceiling openings to upper floors, with “belt towers” built around the belts to protect them. The steam engine ran almost constantly.
When electricity became available, many factories chose a straightforward technology substitution strategy. They simply replaced the central steam engine with a large electric motor and left the underlying mechanical architecture unchanged. The new motor drove the old system. Compared to the investment needed, the efficiency gains were disappointing. So much so that up to around 1910, many argued that steam was preferable. The lack of efficiency gains was not a failure of electricity; it was a failure to redesign the system of production to utilize the capabilities of the new technology.
The real opportunity lay in something different. Instead of a single massive engine and a central drive shaft, factories could be equipped with multiple smaller motors, each powering a local shaft. Each workbench could have its own machine tool driven by its own electric motor. Perhaps most importantly, electricity enabled factories and workers to be organized according to the logic of a production process, rather than the constraints of a central engine and line shafts.
The substantial efficiency gains were realized only when factory architecture and layout were changed, production processes were redesigned, and workers were retrained. In other words, productivity came when factories, processes, and roles were redesigned to use the capabilities of the new technology, instead of simply substituting an old technology with a new one.
By the 1920s, productivity in manufacturing increased at an unprecedented rate. The economic historian Paul David believes this is mostly because manufacturers had learned how to use the technology. We have noted the same pattern with computers and the web. This phenomenon is sometimes referred to as the “productivity paradox.” The conclusion is clear: the lag between visible technological diffusion and measurable productivity gains is correlated with the time required for organizations to redesign their structures, processes, and practices around the new technology.
The Pattern Is Clear
Across multiple technology waves, we see the same pattern:
- A powerful technology is introduced and becomes visible and widely discussed.
- Adoption spreads, but mostly as a substitution for earlier tools within existing structures.
- Productivity gains are limited, and a paradox emerges: evident technological change without commensurate performance improvement.
- Over time, some organizations begin to redesign processes, structures, and roles to fully exploit the technology.
- As these new designs diffuse, aggregate productivity gains finally appear, often with a lag.
The critical inflection point is when business management and leaders stop asking, “Where can I swap in this new tool?” and start asking, “If we built this process from scratch with this technology, what would it look like?”
History Repeats: GenAI and Agentic AI Today
Today, we see the same pattern playing out with GenAI and agentic AI. In many organizations, Generative AI currently plays a role similar to that of early spreadsheets or word processors. It is used as a tool for drafting, summarizing, or accelerating discrete tasks. Its presence is highly visible, but it is often embedded within pre‑existing workflows and structures that remain largely unchanged.
Learning from history, the companies most likely to gain a significant and durable advantage in their markets are those that are willing to rethink:
- How work is decomposed, coordinated, and sequenced, and which steps are automated or reconfigured (process redesign).
- How roles and responsibilities are defined, including which are human, which are digital, and which are hybrid (work structures).
A report from Deloitte Insights (Tech Trends 2026) similarly suggests that companies face the challenge of reimagining their operational processes while simultaneously rethinking their infrastructure strategies. Agentic AI has captured enterprise attention with promises of autonomous operation and intelligent execution, much like the potential promises of electricity at the end of the 19th century. Yet, despite this enthusiasm, a gap remains between pilot projects and scaled production deployment.
According to Deloitte's 2025 Emerging Technology Trends in the Enterprise study, around 30% of surveyed organizations are exploring agentic options and 38% are piloting solutions. However, only 14% have solutions ready for deployment and a mere 11% are actively using these systems in production. More strikingly, 42% of organizations report that they are still developing their agentic strategy roadmap, with 35% having no formal strategy at all.
This reality gap stems from repeating the mistakes of the past. Organizations are attempting to automate existing processes without reimagining how the work should actually be done. Leading enterprises do not simply layer agents onto existing workflows. Instead, they redesign processes from the ground up to leverage the unique strengths of agents. This requires examining end‑to‑end processes rather than searching for isolated automation opportunities within current operations.
In short: don’t automate, rethink.
Agents as a Digital Workforce
Perhaps the most significant rethinking involves recognizing that agents represent a new form of labor. Some organizations are beginning to think beyond using agents as simple automation tools and are exploring ways to integrate them with their human workforce. One useful way to think about agentic AI is as a new kind of “digital workforce.”
These agents can perform many of the tasks humans handle today such as taking inputs, analyzing documents, synthesizing information from multiple sources, drafting reports, and filling out forms or registrations. Yet there is an important distinction. We can imagine these AI resources as a very large pool of workers who are effectively free to use, but each worker can perform (reliably) only a single, well‑defined task.
This framing raises two important questions. First, if we could design an end‑to‑end process with an unlimited number of such specialized agents, each responsible for exactly one task, how would we design that process? We would need to break work down into clear, discrete steps, define interfaces between those steps, and orchestrate handoffs so that the output of one agent becomes the input to the next.
To make this more concrete, it might be useful to think in terms of three levels of agent autonomy:
- Augmentation: Agents enhance human capabilities, as many do today. They draft, summarize, and recommend, but humans retain decision rights. This is often the fastest and safest starting point.
- Automation: Agents execute well‑defined tasks within human‑designed processes. For instance, by validating data, generating routine correspondence, or updating systems of record. Humans handle exceptions and oversee outcomes.
- Autonomy: More advanced systems are entrusted with broader goals and limited decision authority. They can coordinate other agents and, in some cases, humans, operating with minimal oversight within defined guardrails. Humans set objectives and boundaries rather than prescribing every step.
Second, if we deploy an “armada” of agentic AI systems, each with its own specialty, how do we manage them over time? This is where the analogy to an HR‑like function becomes helpful.
An AI HR or Agent Operations capability might encompass:
- Onboarding: Training agents on enterprise‑specific data, systems, and constraints, while ensuring human supervisors understand where and how to direct them.
- Performance management: Monitoring accuracy, latency, incident rates, and escalation volumes.
- Life cycle management: Providing ongoing updates as products, regulations, and processes change. Redeploying agents to higher‑priority areas and retiring obsolete ones.
- Cost and risk management: Tracking the full cost of running and maintaining agents, including hidden costs from poorly configured interactions, and putting in place financial and risk frameworks to control and optimize these costs.
Different organizations will progress through these levels at different speeds. The key is that leaders explicitly decide which level of autonomy is appropriate for which parts of the business, rather than allowing this to happen implicitly and unevenly.
Conclusion: The Lag Is a Choice
The organizations that successfully navigate the dual transformation of reimagining work for agentic AI while simultaneously rethinking infrastructure strategies will gain competitive advantages in AI deployment and operation. Those that fail to adapt will face escalating costs, performance limitations, and strategic vulnerabilities as AI becomes increasingly central to business operations.
The question for every leader is not whether AI will transform their organization, but how quickly they can harness its full potential through thoughtful process redesign and strategic infrastructure decisions. If your AI roadmap is primarily a list of tools to deploy into current workflows, you are repeating the steam‑engine mistake. In other words, you are substituting technology while preserving outdated structures.
The productivity lag is, once again, up to each organization.
Don’t automate, rethink.
