Agentic AI — Ongoing Coverage of Its Impact on Businesses
Agentic AI has left the labs and landed squarely at the center of corporate strategies. And that is not an exaggeration. According to Gartner, while less than 1% of enterprise applications featured this type of technology in 2024, that number is expected to jump to 33% by 2028, with 15% of everyday decisions being made autonomously, without any human intervention.
But what sets agentic AI apart from the AI everyone already knows? The answer is straightforward: while traditional automation follows fixed rules and pre-defined scripts, agentic AI learns from experience, adapts to new situations, and acts independently to achieve goals. It does not just answer questions — it makes decisions, executes tasks, and even communicates with other AI agents with very little human oversight.
This is a real paradigm shift, and the big tech companies are already racing to keep up. Oracle, Microsoft, Cisco, Google, Salesforce, IBM, Nvidia, Adobe, ServiceNow, and CrowdStrike have all made their moves. Meanwhile, real-world adoption cases like DeVry University and Stanford Health Care show that theory is turning into practice across sectors as diverse as education and healthcare.
But the road here has not been linear, and it is far from simple. McKinsey data shows that 39% of organizations are experimenting with AI agents, but only 23% have actually managed to scale the technology within a business function. In other words, going from pilot to production is still the biggest challenge right now. 🚧
In this article, we take a deep dive into what is working, what is still holding companies back, the security risks no one can afford to ignore, and what lies ahead in a race that has promised far more than it has delivered so far — but is also just getting started.
What Makes Agentic AI Different From Everything That Came Before
For years, when the topic was artificial intelligence in the enterprise, the conversation revolved around chatbots, virtual assistants, and data analytics tools. All of these solutions had one thing in common: they depended on a human to take the next step. The AI suggested, the human decided. The AI analyzed, the human executed. It was a support relationship, not one of autonomy.
Agentic AI breaks that model entirely because it does not wait. It identifies the problem, develops a plan, executes concrete actions, evaluates the results, and adjusts course — all within a continuous cycle that can happen without a single human needing to press a button.
This changes the logic of automation in a pretty profound way. We are no longer talking about scripts that follow a fixed flow and stop when they hit an exception. We are talking about systems that can handle ambiguity, make decisions in contexts that were never explicitly programmed, and in some cases, coordinate other AI agents to split complex tasks among themselves. This is what the industry calls multi-agent systems, and they are at the heart of the most advanced implementations that major players like Microsoft, with Copilot Studio and Agent 365, and Salesforce, with Agentforce 360, have already brought to market.
Agentic AI can also enable machines to interact with the physical world at an unprecedented level of intelligence, executing complex tasks in dynamic environments. This is especially relevant for industries facing labor shortages or hazardous working conditions. Microsoft, for example, is developing technologies like MindJourney, a framework that allows AI agents in video to explore three-dimensional spaces before making decisions, combining 3D environment analysis with contextual reasoning and motion prediction.
Real-World Cases Showing the Technology in Action
The DeVry University case is a great example of what this technology looks like in practice. The institution, which had already been using AI in its classrooms for ten years and began experimenting with NLP bots and generative AI use cases as soon as the technology became widely available, deployed its first AI agent in April 2025. The agents started supporting students throughout their academic journeys, from enrollment questions to personalized retention support. The result was a more agile operation with support available at scale, without needing to proportionally expand the human team.
Another notable case came from Stanford Health Care. Nigam Shah, the institution’s Chief Data Officer, demonstrated during Microsoft Build 2025 how agentic AI is helping redefine oncology care. Physicians are frequently overwhelmed by the administrative tasks of medicine, which leads to burnout. AI agents are being used to lighten that load, allowing healthcare professionals to focus on what truly matters: direct patient care.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Walmart is also going all in. According to Hari Vasudev, EVP and CTO for the company in the United States, the AI strategy of the 815-billion-dollar company will play a central role in maintaining its retail leadership. Agentic AI is being integrated into processes ranging from supply chain management to the consumer experience.
The Big Tech Race to Dominate the Agentic Ecosystem
If one thing has become clear in recent months, it is that every major tech company is betting big on Agentic AI — each with its own approach and competitive angle.
Microsoft and Controlling the Agent Ecosystem
Microsoft has been perhaps the most aggressive in this race. Beyond Copilot, which had its price reduced for small and medium businesses to 21 dollars per user per month, the company launched Agent 365 (A365), a control plane that lets customers deploy and govern AI agent usage as they connect to enterprise data. The concern is not just creating agents — it is avoiding what the company calls agent sprawl, an uncontrolled proliferation of bots that becomes impossible to manage.
The company also released the Microsoft Agent Framework, an open-source SDK and runtime for building, orchestrating, and deploying agents and multi-agent workflows, with full support for .NET and Python. And the new Researcher and Analyst agents in Copilot can now detail in real time how data is being analyzed to reach results, adding a layer of transparency that the market is increasingly demanding.
Google and Interoperability Between Agents
Google came in strong with Agent2Agent (A2A), an open protocol introduced at Cloud Next that aims to solve one of the market’s biggest problems: agents built in different vendor ecosystems cannot communicate with each other. A2A seeks to create an interoperability standard that allows agents from different platforms to work together in a coordinated way.
Additionally, Gemini Enterprise replaced Agentspace as the gateway to agentic AI in the workplace, offering enterprise search capabilities to access data from different applications. Google also launched the Agent Development Kit (ADK), an open-source framework that lets you build an AI agent in fewer than 100 lines of Python code, and revealed the Agent Payments Protocol (AP2), developed with more than 60 payments and technology companies to support secure agent-led transactions.
Oracle, Salesforce, IBM, and the Push Into Enterprise Applications
Oracle revamped its Fusion Cloud Applications with the launch of Fusion Agentic Applications, embedding AI agents directly into transactional workflows. The company also added pre-built agents to its Fusion Cloud CX to automate sales, service, and marketing processes.
Salesforce updated its offering with Agentforce 360 and the Trusted AI Foundation, evolving from an application platform into an operating system for enterprise AI ecosystems. IBM, meanwhile, launched the Enterprise Advantage consulting service to help CIOs take agentic AI applications from experimentation to production at scale, while also adding AgentOps capabilities to watsonx Orchestrate with real-time monitoring and policy-based controls.
Why Scaling Is Still the Biggest Bottleneck for Companies
If there is one number that sums up the current state of Agentic AI in the enterprise, it came from McKinsey: 39% of organizations are testing AI agents at some level, but only 23% have managed to scale to an actual business function. This gap between experimentation and production is no accident. It reflects a series of technical, cultural, and operational challenges that companies are still learning to navigate.
Technical Integration Challenges
From a technical standpoint, integrating AI agents with legacy systems is often the first hurdle. A large share of companies operate on infrastructure built decades ago, with databases, ERPs, and CRM platforms that were never designed to communicate with autonomous systems. Getting an AI agent to access, interpret, and act on data distributed across multiple systems requires an integration effort that goes well beyond deploying the model itself. It involves architecture, data governance, API orchestration, and in many cases, a deep overhaul of how the company organizes its information.
This is precisely the problem that an analysis published by CIO World called the single-model trap — the idea that all you need is a good LLM for everything to work in production. In reality, AI agents fail because the operational environment is messy: requests change format, latency budgets conflict, tools break, costs spike, policy constraints shift, and failure modes pile up. It is a far more complex scenario than any lab demo can simulate.
Databricks acknowledged this challenge when it announced the acquisition of Neon, an open-source serverless Postgres company, specifically because agentic AI demands a new type of architecture. Traditional workflows create bottlenecks that hurt speed and performance, and without the right data infrastructure, agents simply cannot operate reliably at scale.
Cultural Resistance and the Human Factor
From a cultural standpoint, the challenge is equally significant. Delegating decision-making to an autonomous system requires a mindset shift that does not happen overnight, especially in regulated sectors like healthcare, finance, and legal. Teams accustomed to approving every step of a process need to learn to trust systems that make decisions without asking permission — and that generates legitimate resistance.
A survey cited by CIO World revealed an interesting split: while most CIOs and CTOs are optimistic about agentic AI, believing it will soon become essential, operational-level IT professionals — the very people who will be responsible for implementing the agents — have serious doubts. This disconnect between leadership and execution is a factor that could significantly delay large-scale adoption. 🤔
Security research has also added weight to a truth that infosec professionals had already noticed: AI agents are still not very smart in certain contexts and can be easily tricked into doing the wrong or dangerous thing. Agents still need humans to teach them — specific procedural knowledge is essential for them to perform their roles well, and they cannot teach themselves, as recent research suggests.
Security, Governance, and the Risks That Cannot Be Ignored
When an AI system acts autonomously, the risk surface grows proportionally to its ability to take action. And this is a point companies need to take very seriously before expanding any Agentic AI implementation. Agents that have access to sensitive data, that can execute transactions, send communications, or modify system configurations represent a far more attractive attack vector than a simple chatbot. If an agent is compromised, manipulated through prompt injection, or simply misconfigured, the consequences can go well beyond a wrong answer in a conversation.
The Role of MCP and Its Risks
The Model Context Protocol (MCP) has become the plug-and-play standard for agentic AI applications to access real-time data from multiple sources. Thousands of MCP servers are already available from a wide range of vendors, allowing AI assistants to connect to enterprise data and services. However, like any widely adopted protocol, MCP also introduces significant security risks. The way it is deployed can create vulnerabilities that malicious actors are more than willing to exploit.
Security Frameworks and the Zero Trust Approach
Cisco, which launched its AI Defense platform specifically to address these issues, identified that most companies lack adequate visibility into what their AI agents are doing in real time. The company is implementing identity and access management capabilities, a toolkit for customers to embed security controls into AI agents, and automation features that let security operations teams see and respond quickly to problems.
CrowdStrike also made a massive bet in this space, launching its Agentic Security Platform and the Agentic Security Workforce after acquiring Onum for 290 million dollars. The goal is to outpace adversaries who are also using AI with real-time intelligence, automation, and a common language of defense.
The lessons from the OWASP Top 10 for agentic AI risk management are also relevant here. LLM-powered chatbots already carry risks that make headlines almost daily, but chatbots are limited to answering questions. AI agents, however, access data and tools and execute tasks, making them infinitely more capable — and more dangerous for businesses.
What the most mature companies in this space have been doing is adopting a zero trust approach applied specifically to AI agents. This means each agent operates with the minimum permissions necessary for its function, all actions are logged and auditable, and there are human review mechanisms for decisions that exceed certain impact thresholds. IBM and Google have been pushing exactly in this direction with their respective platforms, and the market for AI observability tools is growing rapidly precisely because security and transparency are not optional. They are prerequisites.
The Transformation of Entire Industries — From Retail to Cybersecurity
Agentic AI is not just affecting IT departments. It is reshaping entire sectors of the economy, and some of the most interesting examples come from places not always associated with cutting-edge technology.
Retail and Data Cleanup
In retail, so many challenges stem from unreliable product data. AI agents are being evaluated for their ability to clean and organize that data enough to make a real difference in operations. If it works for retail, the same approach can be applied to other business verticals suffering from the same data quality problem.
Cybersecurity on the Front Lines
In cybersecurity, agentic AI has already moved from lab demos to real deployments in Security Operations Centers (SOCs). Unlike traditional automation scripts, software agents are designed to act on signals and execute security workflows intelligently — correlating logs, enriching alerts, and even taking first-line containment actions. At the same time, security professionals warn that agentic AI is both an ally and a threat, since adversaries are also using these same capabilities to make their attacks more sophisticated.
The Future of RPA and SaaS Applications
Another significant transformation involves the future of RPA (Robotic Process Automation). The sector is accelerating toward a crossroads, with IT leaders and experts debating whether more powerful AI agents will replace this two-decade-old technology or whether agents and RPA will work side by side. There are strong arguments on both sides.
Similarly, an emerging theory suggests that AI agents could kill the SaaS business model as we know it. The claim is not new, but it is resurfacing with figures like Microsoft CEO Satya Nadella voicing this position. Nadella went so far as to say that agents will replace all software. Nvidia CEO Jensen Huang predicts that we will soon see hundreds of millions of digital agents operating within enterprises. These are bold statements, but expert consensus is still divided on how much of this vision will materialize — and on what timeline.
What Comes Next in This Race
With Gartner projecting that 33% of enterprise applications will incorporate Agentic AI by 2028, and that 40% of enterprise software will have task-specific AI agents by the end of 2026, it is clear that the question is no longer whether this technology will go mainstream — but rather how fast and at what level of maturity it will happen.
The race among big tech companies is already set in terms of who the main players are, but the competitive edge over the next few years will depend far less on who launched the most powerful agent and far more on who can build the most trustworthy, secure, and easy-to-integrate ecosystems that match the operational reality of businesses.
Nvidia, which might look like a hardware company at first glance, is increasingly positioned as critical infrastructure for agentic AI, with its GPUs serving as the foundation on which most major models and agent platforms run. The launch of the AgentIQ toolkit for connecting distinct agents and frameworks, along with the Nemotron 3 family of open models, reinforces this position. The company is betting that AI agents need to cooperate, coordinate, and execute across large contexts and extended time periods, which requires a new kind of infrastructure — one that is open.
Meanwhile, Oracle is betting on deep integration with its own ERP and CRM systems, creating a compelling argument for companies that already live within the Oracle ecosystem and want to reduce implementation complexity. AWS created a division dedicated exclusively to promoting agentic AI on its platform. And Red Hat made agentic AI the dominant theme of its Summit, announcing improvements to enterprise Linux with enhanced support for containers, edge devices, and of course, AI agents.
An EY executive warned that companies struggling to keep up with the arrival of AI agents should brace themselves: even more complex agentic AI technologies are arriving fast, including physical AI with robots and quantum computing. The message is clear — anyone who thinks the current moment is already challenging has not seen anything yet.
The 2028 horizon might seem far off, but at the pace this technology is evolving, three years is enough time for companies that started experimenting now to reach robust and scalable implementations, while those that sat on the sidelines will face a gap that is tough to close. Intelligent automation, autonomous decision-making, and security by design are no longer future concepts. They are being built right now, in production, with real results. And the companies that understand this sooner will have an advantage that goes well beyond operational efficiency. 🚀
