Agentic AI — The autonomous revolution already transforming the corporate world
Agentic AI is no longer a future promise — it has become an urgent topic in the boardrooms of the biggest companies on the planet. What was once just a concept explored in research labs is now shaping strategic decisions, redesigning workflows, and fueling a tech race among giants like Oracle, Microsoft, Google, Salesforce, IBM, and Nvidia.
According to Gartner, by 2028, 33% of enterprise software will feature agentic AI — a staggering leap, considering that number was less than 1% in 2024. And it doesn’t stop there: the forecast is that 15% of everyday business decisions will be made autonomously by AI systems, with no direct human intervention at all.
So what exactly changes with this technology?
Unlike traditional AI, which follows predefined rules and responds when someone asks, agentic AI acts on its own, learns from situations, adapts to context, and can even communicate with other AI agents to carry out complex tasks. This is a real paradigm shift — and it’s already happening right now. 🚀
Companies like Oracle, Salesforce, Google, Microsoft, IBM, and Nvidia are already pouring billions in this direction, each placing their own bet to dominate the new era of autonomous agents in the enterprise landscape.
But along with all that excitement comes a series of challenges that still aren’t getting the attention they deserve — from fragile governance and legacy architectures that hold back scalability, to serious security concerns that set an autonomous agent apart from a simple chatbot.
In this article, you’ll find out what’s working, what’s still stuck, and which risks need to be on your radar right now. 👇
What makes agentic AI different from everything that came before
When most people think about artificial intelligence in the workplace, they still picture a virtual assistant answering questions or a system that crunches data and generates a nice-looking report. Agentic AI is something entirely different. It doesn’t wait to be triggered — it acts. It doesn’t follow a fixed script — it decides which path to take based on the context in front of it.
This ability to reason, plan, and execute tasks in sequence — often involving multiple systems and data sources at the same time — is what puts agentic AI in a category of its own within the world of enterprise automation. The core concept here is autonomy. A well-designed AI agent can receive a high-level objective — like reducing customer response time by 40% — and from there, break that objective into smaller tasks, prioritize what needs to happen first, trigger external tools, query databases, coordinate other specialized agents, and learn from the results along the way.
None of this requires a human overseeing every step. That’s the difference catching the attention of executives around the world, because it radically changes the productivity equation inside companies.
It’s also worth highlighting the ability for agents to collaborate with each other, something the technical literature calls multi-agent architecture. Imagine a scenario where one agent handles financial analysis, another monitors inventory in real time, a third manages communication with suppliers, and all of them share information in a coordinated way to make a smarter purchasing decision. This is already becoming a reality at companies that bet early on this technology, and the practical results are starting to show up in operational efficiency numbers across sectors like logistics, healthcare, and financial services.
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.
Beyond that, agentic AI promises to let machines interact with the physical world with unprecedented intelligence. Industries dealing with labor shortages or hazardous working conditions stand to benefit enormously from agents capable of handling complex tasks in dynamic environments. Think factories, construction sites, field operations — contexts where intelligent automation can make a real difference in both the safety and productivity of teams.
How the biggest companies are betting on this technology
The move by tech giants toward Agentic AI isn’t subtle — it’s an all-out declared race. And each company is choosing its strategic path based on its strengths and the audience it wants to win over.
Salesforce and the bet on autonomous sales and service
Salesforce launched the Agentforce platform, which is already on version Agentforce 360, focused on creating autonomous agents for sales, customer service, and marketing. The company also introduced the Trusted AI Foundation, which aims to turn the Salesforce platform into a true operating system for enterprise AI ecosystems. The goal is clear: agents that can close deals and solve problems with less and less human intervention.
Microsoft and the integrated agent ecosystem
Microsoft is embedding autonomous agents directly into the Microsoft 365 ecosystem, with Copilot evolving into a task orchestrator that operates within emails, spreadsheets, and meetings. The company launched Agent 365 as a control plane for managing and governing agent usage. On top of that, the Microsoft Agent Framework, an open-source SDK, enables building and orchestrating multi-agent workflows with support for .NET and Python. In March 2025, the new Researcher and Analyst agents began detailing in real time how they analyze data and reach conclusions — an important step toward transparency.
Google and open infrastructure for agents
Google is betting on Vertex AI and its model infrastructure to let companies build custom agents with a high degree of specialization. The company launched the open Agent2Agent (A2A) protocol, designed to connect agents across different ecosystems, along with the Agent Development Kit (ADK), which lets you create an AI agent in fewer than 100 lines of Python code. More recently, Gemini Enterprise replaced Agentspace as the gateway for AI agents in the workplace, serving as a unified platform for accessing and coordinating agents that automate corporate tasks.
IBM, Oracle, and Nvidia in the infrastructure trenches
IBM is positioning itself with a focus on the most heavily regulated industries, like banking and insurance, where automation needs to happen within very well-defined boundaries of compliance and auditability. watsonx Orchestrate offers over 500 customizable tools and agents, with AgentOps capabilities for real-time monitoring. The company also launched the Enterprise Advantage consulting service to help CIOs move AI applications from the experimental stage to production at scale.
Nvidia is in the game on the infrastructure side, providing the hardware and frameworks that make it viable to run multi-agent systems at scale. The company launched the AgentIQ toolkit for connecting different agents and frameworks, along with the Nemotron 3 family of open models, designed specifically for the agentic era. And Oracle integrated agentic capabilities directly into its ERP and CX products with Fusion Agentic Applications, meaning enterprise customers can activate these features without having to overhaul their systems from scratch.
What’s interesting to watch in this race is that it’s not being fought on the technical front alone. The battle is also happening at the trust layer — whoever can convince IT, legal, and compliance teams that their agents are auditable, explainable, and controllable is going to come out ahead. Because at the end of the day, no CFO is going to sign off on an autonomous system making million-dollar decisions if there isn’t a clear trail of how that decision was made and how it can be reversed if something goes wrong.
The real challenges nobody likes to admit
As exciting as the landscape may be, there are a number of concrete obstacles slowing down the large-scale adoption of Agentic AI within enterprises.
Legacy infrastructure that won’t cooperate
The vast majority of organizations still run on systems that were built decades ago, with architectures that were never designed to communicate with autonomous agents in real time. Layering agentic AI on top of these systems is technically possible, but it tends to be slow, expensive, and full of failure points that only surface in production — when it’s already too late for a quick fix. It’s no surprise that one of the most recent Computerworld reports highlights how the complex enterprise architectures of large organizations frequently struggle to meet the demands of this technology.
Governance still under construction
Who’s responsible when an autonomous agent makes a bad call? How do you ensure it will act within the organization’s ethical and legal boundaries? How do you audit a sequence of actions that happened in milliseconds, involved dozens of systems, and was executed with no human intervention? These questions still don’t have standardized answers, and that creates a massive zone of discomfort for risk and compliance teams.
A McKinsey survey reveals that while 39% of organizations are experimenting with agents, only 23% have started scaling AI agents within a business function. Gartner predicts that 40% of enterprise software will have AI agents for specific tasks by the end of 2026, but the lack of mature governance frameworks is perhaps the most powerful brake on corporate adoption today.
Projects that don’t scale
One of the most commonly observed patterns in the market is pilot projects that never make it past the proof-of-concept stage. Rising costs, unrealistic expectations, and fragile governance are the main culprits. Gartner senior analyst Anushree Verma notes that most agentic AI projects today are early-stage experiments, driven mostly by hype and frequently misapplied. Separating the initiatives that survive from those that are quietly shelved requires rigorous planning and disciplined execution.
Hallucination in an agentic context
Then there’s the problem of hallucination in an agentic context. When a language model hallucinates during a chatbot conversation, the result is embarrassing. When an autonomous agent hallucinates in the middle of a decision chain involving financial data, contracts, or critical processes, the damage can be far greater. Agentic systems amplify both the capabilities and the errors of the underlying AI, and that puts enormous pressure on engineering teams to build verification, redundancy, and rollback mechanisms that can contain failures before they propagate through the automation chain.
Recent security researchers reinforce a truth that industry professionals had already understood: AI agents aren’t that smart and can be easily tricked into doing dangerous things. That doesn’t invalidate the technology, but it demands a much more cautious approach when granting autonomy to these systems.
Security: the issue that can’t be put off
Security in Agentic AI systems is a field still being built in real time, and that should concern any company that’s accelerating adoption without giving this aspect the attention it deserves.
An autonomous agent, by definition, has permissions to access systems, move data, execute actions, and in some cases, authorize transactions. This creates an entirely new attack surface that goes well beyond traditional cybersecurity vectors. If an agent is compromised — whether through malicious prompt injection, an authentication flaw, or a vulnerability in the tool chain it accesses — the impact can spread much faster and more broadly than any conventional attack.
The threat of indirect prompt injection
One of the most discussed attack vectors in the security community today is called indirect prompt injection, where an AI agent processes a document or webpage that contains malicious instructions disguised as legitimate content. The agent, without realizing it, executes those instructions as if they were part of its original task. This is particularly dangerous in scenarios where the agent has access to emails, external documents, or any type of input that comes from outside the company’s controlled environment.
OWASP has already published a Top 10 list of risks specific to LLM-powered chatbots, and those risks escalate dramatically when we’re talking about agents that don’t just answer questions but access data, tools, and execute tasks. That difference between a limited chatbot and an autonomous agent with the power to act is exactly what makes agentic security such a critical field.
The MCP protocol and its risks
The Model Context Protocol (MCP) has quickly become the plug-and-play standard for agentic AI applications to pull real-time data from multiple sources. With thousands of MCP servers already available from various vendors, the protocol is accelerating adoption — but it’s also opening new doors for malicious actors looking to exploit weaknesses in how MCP is deployed. It’s a classic example of how innovation and risk advance side by side.
Emerging best practices
Best practices emerging in this space include:
- Principle of least privilege — an agent should have access only to what is strictly necessary to perform its task, with no extra permissions that could be exploited if it’s compromised.
- Detailed logs of all actions — full traceability of every decision and execution by the agent.
- Human approval mechanisms — for decisions above a certain level of financial or operational impact.
- Regular adversarial prompting tests — to identify vulnerabilities before real attackers find them.
- Agentic red teaming — traditional penetration testing techniques updated for the AI world.
Cisco is investing heavily in identity and access management for AI agents, with toolkits that embed security controls directly into the agents. CrowdStrike, following its acquisition of Onum for 290 million dollars, launched its Agentic Security Platform and the Agentic Security Workforce, aiming to outpace adversaries who are also using AI. Security in agentic AI isn’t a one-time project — it’s an ongoing practice that needs to be embedded into the development culture from day one.
Real-world cases showing where this is headed
While many companies are still in the experimentation phase, some are already seeing concrete results with agentic AI in production.
DeVry University deployed its first AI agent in April 2025 to assist current and prospective students, after a decade of using AI technology in its classrooms. Stanford Health Care is using agentic AI to ease the burden on oncology professionals by automating the administrative tasks that lead to physician burnout. Walmart, the world’s largest retailer with 815 billion dollars in revenue, has stated that its artificial intelligence strategy — with a strong agentic component — will be a key piece in maintaining its retail leadership.
In the retail sector more broadly, AI agents are being tested to solve one of the industry’s oldest and most persistent problems: product data quality. Unreliable data is the root cause of countless issues in logistics, pricing, and customer experience, and agentic AI promises to clean up and organize that data in ways that would be unfeasible with traditional methods.
In the networking and infrastructure space, Forward Networks launched an agentic AI system built on top of a network digital twin, allowing network teams to ask complex questions, understand network behavior, validate results, and automate workflows securely. Riverbed also updated its AIOps platform with predictive and agentic capabilities, helping IT organizations shift from reactive to predictive operations.
The relationship between AI agents and the future of SaaS and RPA
One of the most provocative discussions in the market is whether AI agents will kill the SaaS business model as we know it. Microsoft CEO Satya Nadella himself has said that agents will replace all software. That’s a bold statement, and experts are divided on it. But what seems clear is that the way companies consume software is changing — from applications you need to learn how to use to agents that do the work for you.
The relationship with RPA (Robotic Process Automation) is also evolving. Some IT leaders believe that more powerful and autonomous AI agents will replace RPA, a technology that’s already two decades old. Others predict that AI agents and RPA will work side by side, with each complementing the other’s capabilities. The most likely scenario is a middle ground, where RPA handles well-defined repetitive tasks and AI agents take on processes that require reasoning, adaptation, and decision-making in ambiguous contexts.
What to expect going forward
The trajectory of Agentic AI in the enterprise will depend heavily on how the industry resolves the tensions between speed of adoption and governance maturity. The companies moving fastest are those that have managed to build internal structures for safe experimentation — small teams with the freedom to test agents in controlled settings, learn from mistakes, and gradually scale what works. This progressive adoption model is proving more effective than the massive all-in transformation bets many organizations attempted in the early years of the AI era.
In the near term, one of the most relevant trends is the standardization of communication protocols between agents. Today, each platform has its own way of making agents talk to each other, which creates silos and makes integration difficult. Initiatives like the MCP protocol, Google’s Agent2Agent (A2A), and other interoperability efforts are trying to create a common language so agents from different vendors can collaborate without friction. Once that level of standardization arrives, multi-agent automation will become far more accessible to companies of all sizes.
Deloitte predicts that in 2025, 25% of companies using generative AI will launch agentic AI pilots or proofs of concept, growing to 50% by 2027. Some applications, in certain industries and for specific use cases, could see real adoption in existing workflows as early as this year. What’s clear is that Agentic AI is no longer a question of whether companies will adopt it, but how and with what level of preparedness they’ll do it.
Organizations that are thinking today about infrastructure, governance, and security will have a real competitive edge when this technology reaches operational maturity. And those that ignore these fundamentals will learn the hard way that autonomy without control is just another word for risk. 🎯
