AI agents are the most talked-about topic on the enterprise technology agenda. Yet the gap between interest and results is still wide. In a June 2025 forecast, Gartner stated that more than 40% of agentic AI projects would be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls.
A year later, the picture shows expectations growing rapidly. According to Gartner's 2026 Hype Cycle for Agentic AI, agents sit at the Peak of Inflated Expectations. Only 17% of organizations have deployed agents so far, while more than 60% expect to deploy them within two years. Gartner describes this as the most aggressive adoption curve among the emerging technologies it measures.
In this article, we cover, with sources, the tasks in which agents create value today, the points at which projects stall, and the questions to ask before an investment decision.
What does an AI agent do?
The word "agent" is used for systems with very different capabilities. The classification Gartner proposed in May 2026 offers a useful framework for clarifying the distinction. In this framework, agents fall into four levels of autonomy:
- Observe: Read-only access to defined data sources. Document summarization and information access are typical examples.
- Recommend: Produces drafts, suggestions or recommended actions; a human reviews every output and carries out the action. Email drafts and decision support belong at this level.
- Act with approval: Can perform actions such as writing data or sending communications, but every action requires explicit human approval.
- Act autonomously: Operates independently within defined boundaries; humans monitor exceptions and aggregate outcomes rather than individual decisions.
This distinction matters because a project's risk, cost and required level of oversight depend largely on the level at which the agent operates.
Where do AI agents create value?
According to Gartner's Hype Cycle assessment, organizations are trying out agents mainly in areas such as software development, customer support and operations, to automate individual tasks. The same assessment stresses that most deployments remain narrow in scope and that fully autonomous agents are not ready for most enterprise use cases.
McKinsey's global survey, published in August 2026, paints a similar picture. About two in ten respondents report that their organizations have moved to scaling AI agents across the enterprise. Among large organizations with annual revenue above $1 billion, the share scaling agents in at least one function rose from 27% to 40% in a year; in smaller organizations it stayed flat at 22%.
One of the most visible areas of value is software development. According to McKinsey, 32% of respondents say their organizations have decided against buying at least one software product or feature because they were able to build it in-house with agent-based coding tools.
Customer service, meanwhile, is where agents are moving from "answering" to "completing tasks." In a Gartner survey of 3,566 customers conducted in February–March 2026, 58% of customers who use generative AI have had these tools complete a task on their behalf; in business (B2B) settings, that share rises to 74%. However, 87% of customers say companies using these tools must offer the option to reach a human representative. Gartner recommends that AI not be a mandatory first step for every issue; instead, it should understand intent, gather information and attempt a resolution only when confidence is high.
The common denominator is clear: value comes from tasks with a well-defined scope, measurable outcomes and defined human oversight.
Why are projects canceled?
1. Building an agent for work that does not need one. According to Gartner, many use cases positioned as agents today do not actually require an agent-based application. The firm notes that most early-stage projects are experiments or proofs of concept launched on the back of hype, and recommends pursuing the agent approach only where it delivers clear value or return on investment.
2. Overlooked costs. Gartner stresses that integrating agents into legacy systems can be technically complex, can disrupt workflows and can require costly changes. In the McKinsey survey, roughly one in five respondents say AI operating costs have started to constrain usage, and roughly one in ten say agent usage is limited by cost. According to McKinsey, even as unit costs fall, the volume of processing consumed is growing faster.
3. Governance gaps. Gartner predicts that by 2027, 40% of organizations will reduce the authority of or decommission their autonomous agents because of governance gaps that are noticed only after incidents in production. According to Gartner, the fundamental mistake is applying the same rules to all agents. When simple agents are restricted more than necessary, delivery slows down and uncontrolled development increases; when autonomous agents are not supervised sufficiently, operational, security and compliance risk grows. An approval mechanism is not a safeguard on its own either: under time pressure and approval fatigue, human approval can lose its meaning.
4. Externally dependent implementation. Gartner predicts that by 2028, 70% of organizations will abandon agent solutions that vendor engineers develop directly inside the organization, because of rising costs and an inability to develop such solutions on their own. Gartner recommends clarifying knowledge transfer, intellectual property and exit responsibilities before the contract is signed, and measuring success by whether the organization can manage the solution independently.
Capability, not a label
There is one more warning for the selection process. Gartner notes that some vendors take existing products such as AI assistants, RPA and chatbots and rebrand them as agents without substantial agentic capability, and calls this "agent washing." For this reason, an evaluation should ask not what the product is called but which of the autonomy levels above it genuinely operates at.
Pre-investment checklist
The steps below are a framework derived from the findings in the sources; each organization needs to adapt it to its own structure.
- Define the task first. Does the task require decisions, or is it a repetitive flow with well-defined rules? In Gartner's words, agents should be used where decisions are needed, automation for routine workflows, and assistants for simple information retrieval.
- Set the autonomy level up front. Put in writing what the agent will be allowed to read, recommend or execute, along with the oversight requirements at each level.
- Calculate the total cost. Measure integration, workflow change and operating costs during the pilot phase; do not look at license cost alone.
- Preserve access to humans. In customer-facing scenarios, keep the path to a human representative open.
- Plan for in-house capability. In projects that start with external support, write knowledge transfer and an exit plan into the contract.
Conclusion
AI agents offer a genuine leap in capability, but value comes less from the technology itself than from choosing the right tasks and from governance. The three sources of cancellation risk — unclear value, overlooked costs and inadequate controls — can largely be managed through decisions made before a project starts.
From Orient Global's perspective, a sound first step before starting an agent investment is to take a process inventory and answer the question "assistant, automation or agent?" for each task. Starting the first chosen agent on a narrow, measurable, human-approved task, and increasing autonomy gradually as results become visible, markedly reduces risk.
This perspective is informed by operations experience: Orient Global's operations and supply-chain consultant has more than 30 years in logistics and supply chain, including managing nationwide warehouse and transport operations and large-scale distribution networks, and leading teams of 100+ people.
#interimmanagement #forwarddeployedengineering