A Deloitte survey published in August 2026 presented a set of numbers that are easy to misinterpret: 42% of companies are still piloting a limited number of AI Agent, 43% are expanding deployment across multiple departments, and only 15% have achieved scaled, coordinated multi-Agent operations. At first glance, it may seem that “everyone is doing it.” In reality, the vast majority of companies are stuck at the same point—they know how to use AI Agent, but cannot scale it across the organization.
For decision-makers in Taiwan’s traditional industries, the survey’s greatest value lies not in its growth projections, but in how it identifies a bottleneck that most companies do not even have a budget category for.
74% Versus One in Five: The Gap Is the Answer
Two figures in the survey become meaningful only when viewed together.
When asked about their expectations four years from now, 74% of companies said that at least half of their processes would be redesigned around AI Agent by then. Yet when asked about the present, only 16% said their business processes were ready, with another 5% saying they were highly ready. Even among companies that have already reached scale, only 46% believe their processes are prepared.
Within four years, companies expect to redesign half of their processes. Today, only about one in five believe those processes are ready to change. This is not a technology gap. It is a process gap.
The ranking of the three biggest obstacles points to the same conclusion: 72% are held back by the lack of a unified, accessible data foundation; 70% struggle to trust and govern Agent; and 67% face the cost and complexity of integration. None of these are model-capability issues. In Taiwan’s traditional industries, the first obstacle has a more familiar name: data silos. ERP data sits in one system, quality records are stored in Excel, and quotation histories remain buried in senior salespeople’s inboxes. In this environment, it does not matter which model a company adopts.
Layering Is a Bridge, Not the Destination
One of Deloitte’s recommendations deserves special attention: treat “layering” as a bridge rather than a destination.
Layering means adding AI Agent directly on top of existing processes without changing the underlying workflows. This approach delivers a fast return, often producing visible results within weeks. Most companies in the survey are currently at this stage.
Our view is that for traditional businesses with limited resources and no dedicated IT team, layering is not merely acceptable—it is the only sensible place to start. Frontline employees first need to see how AI Agent improves the work they do every day. That is how confidence is built, and only then will people be willing to support a broader process redesign.
But every bridge needs to lead somewhere. If there is no plan for what comes after layering, the company may end up three years from now with dozens of disconnected Agent, each serving a separate legacy process, unable to communicate with one another while the underlying processes remain completely unchanged. The distinction is simple: when an Agent goes live, does someone document which part of the process it is bypassing—the part that should ultimately be fixed? If so, it is a bridge. If not, it has already become the destination.
People Are Even Less Prepared Than the Technology
Among the executives surveyed, 43% expect “extensive” to “extreme” changes in work over the next 12 to 18 months. Over a two- to three-year horizon, that figure rises to 72%. Yet in the same survey, half of executives said their companies were underinvesting in workforce transformation.
Placed side by side, these figures reveal a contradiction: most leaders expect disruption, while half already know they are unprepared for it.
What are companies actually doing? Seventy-one percent are promoting basic AI Agent literacy, while 65% are retraining employees in roles likely to be affected. Both are steps in the right direction. But after the literacy course ends, how many people in the company can design their own prompts or recognize when an Agent should not be trusted?
Another figure in the survey helps answer that question: about one-third of executives believe human-AI collaboration delivers more value than pure automation, yet fewer than half of companies have defined the corresponding operating models. No tool can decide on a company’s behalf which steps should remain subject to human judgment and which should be executed by Agent.
What Decision-Makers Should Know About the Survey’s Limits
Before citing the survey, it is important to clarify what it actually measures.
The sample consisted of 501 senior managers and executives from U.S. companies, surveyed between April and June 2026, along with 20 in-depth interviews. Industries represented included consumer goods, energy and industrials, financial services, life sciences and healthcare, and technology, media, and telecommunications. Of the respondents, 55% were technology leaders, while 45% represented business and other functions.
One eligibility criterion is especially important: every company surveyed was already piloting at least one AI Agent.
This is therefore not a snapshot of “the business world.” It is a snapshot of the companies that have already started moving. The 15% scaling figure uses companies already engaged in AI Agent pilots as its denominator; interpreting it as the adoption rate across all companies would substantially overstate market penetration. Similarly, the 74% four-year expectation comes from organizations that have already invested, introducing a degree of self-fulfilling bias.
For decision-makers, the right way to use this survey is as a reference for how peers are approaching the journey—not as evidence of market pressure. The real pressure comes from your customers and competitors, not from percentages in a U.S. survey.
Four Questions to Ask at Your Next Meeting
Working backward from the survey, four questions can reveal where your company truly stands.
Which version of the process does our Agent currently support? If the answer is “the current version,” then you are in the layering stage. That is not a problem, but it should lead directly to the next question.
Which part of the process do we know needs to change, but are currently using Agent to bypass? If no one can answer, no one is tracking where the bridge is supposed to lead.
How many systems contain our data, and how many can the Agent access? The answer is usually more uncomfortable than expected—and it determines the upper limit of what the Agent can achieve.
If Agent handles half of our routine decisions next year, who will review them, and does that person currently know how to inspect prompts? This is a question about workforce investment—the same area in which half of companies admit they are underinvesting.
FAQ
This is a U.S. survey. Does it apply to Taiwan’s traditional industries?
The direction of the trends applies; the percentages do not. Process readiness lagging behind technology, fragmented data being the largest obstacle, and insufficient workforce investment are all evident among manufacturing and machine-tool customers in Taiwan. However, the adoption percentages come from a sample of large U.S. companies. Applying them directly to Taiwan’s small and midsize manufacturers would overstate adoption. Use the survey as a reference for the sequence of the transformation roadmap, not as a benchmark for progress.
If only 15% have achieved scale, does that mean there is no need to rush?
The denominator behind that 15% is “companies already running pilots,” not all companies. The real signals are two other figures: 43% are already expanding across departments, and 72% of executives expect extensive changes in work within two to three years. If you wait until adoption becomes widespread, what you will lack is not technology but the time needed to redesign processes and train people. Neither can be accelerated simply by spending more money.
Our data is highly fragmented. Should we organize it before adopting AI Agent?
We do not recommend waiting. If 72% of companies are being held back by their data foundation, requiring all data to be fully organized before starting would mean that most companies never begin. In practice, the sequence should be reversed: start with one process whose data is relatively complete, deliver tangible results, and give employees a reason to see why organizing data is worthwhile. Then expand outward. Data governance is an outcome of adoption, not a prerequisite.
Should we choose layering or process redesign?
Start with layering, but document from day one what the Agent is bypassing. Layering allows you to generate results and secure internal support within weeks; process redesign requires that support before it can move forward. The difference between the two is not the method itself, but whether you preserve a clear path to the next step.
What comes after the basic literacy course?
Literacy helps employees understand what AI Agent can do. The next step is developing people who can design and maintain them independently. Intellicon Solutions’ approach is to cultivate one or two prompt engineers in each department, guiding them from observation and hands-on practice to independent delivery. By the end of the program, they can modify their own department’s Agent themselves. Seventy-one percent of companies are providing literacy training, but literacy is only the starting point.
References
- The path to agentic transformation — Deloitte Insights, August 12, 2026, by China Widener, Laura Shact, David Jarvis, and Sayantani Mazumder. The survey covered 501 senior managers and executives at U.S. companies between April and June 2026 and included 20 in-depth executive interviews. All percentages cited in this article are drawn from the report.