DAILY ISSUE · 2026-07-05 · AGENTIC AI · ENTERPRISE

When 40% of Enterprise Apps Include AI Agents — Three Barriers That Decide Whether Gartner’s Forecast Becomes Operational Reality

Gartner’s forecast sent a strong signal to enterprise IT: by the end of 2026, 40% of enterprise applications would feature task-specific AI agents, up from less than 5% in 2025. The market direction is now visible. But the hard part is not adding an “AI agent” label to software. The hard part is making agents work inside legacy systems, organizational accountability structures, and budgets that must prove return on investment.

The forecast is becoming an implementation problem

Roughly one year after Gartner’s 40% forecast, enterprise software vendors are moving in the predicted direction. Salesforce has embedded agent capabilities through Agentforce. ServiceNow is bringing AI agents into ITSM workflows. SAP is positioning Joule inside ERP environments. With services such as AWS WorkSpaces for AI Agents extending automation toward legacy desktop environments, the range of systems that can be touched by agent workflows is widening.

The important shift is that AI agents are no longer being treated only as a feature enhancement. They are becoming part of business process redesign. Early experiments often focused on repetitive tasks, but enterprise use cases are now moving closer to actual operational flows: order handling, customer response, internal approvals, data analysis, and report generation.

That does not mean the forecast has already been fully realized. AgentMarketCap’s analysis points to a more complicated pattern: a sizable share of companies that adopt AI agents need meaningful adjustment within the first six months. In other words, “40% adoption” and “40% requiring correction” can exist at the same time. The number may show momentum, but implementation quality decides whether the momentum turns into value.

Barrier 1: Legacy integration

The most immediate barrier is integration with existing systems. Global technology firms often run on modern SaaS stacks with API-based architectures, making agent adoption comparatively straightforward. Many Korean enterprises — especially in finance, manufacturing, and the public sector — still rely on systems built over decades: mainframes, older ERP systems, internal tools, and workflows that were never designed for autonomous software agents.

In those environments, the agent’s first problem is not intelligence. It is access. If there is no stable API, the agent cannot simply call a clean function to retrieve data, update status, or complete a step. Desktop-control approaches can help by letting agents operate user interfaces, but they are slower and more fragile than API-based automation. If a screen layout changes, the agent workflow may need to be revised.

This creates a strategic choice for companies. A UI-operating agent can be a temporary bridge, especially when replacing a legacy system is too expensive or too slow. But it should not be confused with true modernization. The deeper solution is still to expose business capabilities through reliable interfaces, permission structures, and logs. Enterprises evaluating agents need to distinguish between “a workaround that lets us start” and “an architecture we can safely scale.”

Barrier 2: Governance, trust, and accountability

As AI agents begin to participate in real work, the question shifts from “Can it do the task?” to “Who is responsible when it does the task?” This is the governance problem at the center of enterprise agent adoption. Anthropic’s guidance on hiring and deploying AI agents emphasizes that success depends less on the model alone and more on organizational fundamentals: information transparency, role clarity, shared goals, and incremental trust.

For Korean companies, accountability culture adds another layer of complexity. If an agent recommends a decision, triggers a workflow, or prepares a response that later causes a problem, responsibility must be defined before deployment. Organizations that cannot answer that question clearly will struggle to move agents beyond low-risk experimentation.

The practical governance split is usually simple to describe but difficult to execute: define what the agent may do automatically, what it may only recommend, and what requires human approval. Status lookup may be safe to automate. Cancellation, payment, refund, access changes, or customer-facing commitments may require a human checkpoint. Without this separation, companies either over-restrict the agent until it becomes useless or over-trust it until failures become inevitable.

Samsung’s AX transition offers a useful lens in the Korean context. The key lesson is not merely that a large company is adopting AI agents. It is that agent adoption has to be paired with organizational redesign: which decisions can be delegated, where humans retain final approval, and how those rules are communicated across the company. Trust is not something a model generates by itself. It is created by governance that people can understand and audit.

Barrier 3: Cost efficiency and ROI proof

The third barrier is cost. AI agents often cost more to run than simple chatbot interactions because they perform multi-step reasoning, tool calls, context handling, and repeated inference. As research and operating experience quantify the energy and compute burden of agent workflows, finance teams are asking a sharper question: how much business value does this agent actually create?

Even if agent operation costs have declined compared with 2025, reliable ROI measurement remains difficult. Some reports highlight large productivity gains, but those gains are usually tied to specific workflows and cannot be applied blindly across an entire enterprise. Korean market analysis has also emphasized that companies often need months of operation before they can see whether agent adoption is producing measurable return.

That uncertainty makes small, measurable deployment the safer path. Replacing an entire process at once is risky. A better approach is to start with work units where input, output, time saved, error rate, and escalation rate can be measured. Examples include triaging inquiries, preparing draft responses, generating recurring reports, or checking status across systems. Once the organization can measure value and failure modes in a contained workflow, it can expand with more confidence.

ZHS’s own operating lesson is consistent with that approach: start small, measure, and expand only after the workflow is observable. A vague promise that “agents will improve productivity” is not enough. The useful question is whether a specific agent workflow reduces processing time, lowers error rates, improves response quality, or frees human operators from repetitive work in a way that can be tracked.

Korean enterprise reality and the second-half outlook

Large Korean companies such as Samsung, Hyundai Motor, SK, and LG are actively exploring or deploying AI agent capabilities as part of broader AI transformation. Samsung has moved around AX initiatives. Hyundai Motor is applying AI to manufacturing and logistics contexts. SK has coordinated AI initiatives through group-level structures. These examples show that the enterprise interest is not abstract.

The situation is different for mid-sized and smaller firms. They face higher relative barriers: infrastructure cost, shortage of specialized staff, unclear ROI models, and IT systems that are often less modernized than those of large enterprises. For these companies, agent adoption will depend heavily on managed services, open-source tools, and practical implementation playbooks that reduce the cost of experimentation.

Several changes could narrow the gap in the second half of 2026. Managed environments such as AWS WorkSpaces for AI Agents can reduce infrastructure friction. Open-source agents such as Hermes Agent and OpenClaw can let smaller teams experiment without committing to a full enterprise platform. Efficiency improvements — including inference caching, lightweight model routing, and better workflow design — can reduce operating costs over time.

But the core point remains: enterprise agent adoption is not a race to attach agents everywhere. It is a discipline of choosing the right workflow, setting the right permissions, measuring the right outcome, and improving the organization’s ability to use automation safely.

ZHS operator angle — what running Hermes Agent makes visible

ZHS has been operating Hermes Agent since May 2026, starting with content automation and expanding into a broader workflow that connects more than 34 tools. The agent now supports work such as SEO report generation, content distribution, social media operations, and site monitoring. That experience makes one lesson very clear: agents do not magically fix organizational problems. They expose them faster.

If information is poorly organized, the agent will produce poor results more quickly. If decision rules are unclear, there is no reliable way to judge the agent’s output. If responsibility is undefined, every meaningful automation becomes a risk. The first step in agent adoption is therefore not only technical setup. It is cleaning up information flow, decision criteria, permissions, and escalation paths.

This is why Gartner’s 40% forecast should be read as a pressure test for enterprise operating systems. The model layer is improving quickly. The harder question is whether companies have the data structures, governance habits, and measurement discipline to make agents useful in production.

Gartner’s 40% forecast is becoming plausible, but the number itself is not the goal. The real question is whether enterprises can overcome three barriers: legacy integration, organizational governance, and cost-effective ROI proof. The technology is arriving faster than the operating discipline around it.

For Korean enterprises, the practical path is to begin with bounded workflows, define human approval points, measure results, and expand only after trust is earned. AI agent adoption will not be won by the company that announces the most agents. It will be won by the company that turns agents into controlled, measurable, and accountable operating assets.