Imagine this situation. A customer of an energy company sends a complaint. Today, someone reads it, creates a ticket, finds the contract in the CRM, verifies the payment history in the accounting system, prepares a response, and sends it for approval. The whole process takes hours, sometimes days. Now imagine that an AI agent can do all this in a few minutes. All by itself.
This is not science fiction. It is the reality of 2026.
From chatbot to digital colleague
Most companies already have some experience with artificial intelligence. Chatbots on websites, text generation, and data analysis. But these are all tools that wait for instructions. You ask, they answer. You give them a task, and they complete it. But don’t expect anything more from them.
AI agents work differently. You set a goal for them, and they plan how to achieve it themselves. They go through the company’s systems, pull out the necessary data, take action, and even verify that the result makes sense. When they encounter a problem, they try a different approach. This is the difference we refer to when we talk about agentic AI.
According to the analytical company Gartner, by the end of 2026, over 40% of enterprise applications will use AI agents focused on specific tasks. In 2025, this figure was less than 5%. And Deloitte’s Tech Trends 2026 study points out that companies that successfully deploy AI agents cannot simply add AI to existing processes. They have to rebuild those processes from the ground up.
What AI agents can actually do (and what they can’t yet do)
Let’s break it down with a specific example. Take an insurance company’s customer service. A client reports damage. Today, this involves a series of steps across several systems, most of which require human intervention.
An AI agent can handle this process as a whole. It accepts the report, identifies the client in the CRM, verifies the insurance policy’s validity, compares the damage description with the contract terms, requests any missing documents, and prepares a settlement proposal for review. Humans only get involved when a final decision is needed in more complex cases.
The situation is similar in banking. An agent can now process a loan application that used to take days in minutes. It verifies creditworthiness, checks registers, assesses risk, and prepares recommendations for the approver. It can handle most routine applications completely on its own.
Importantly, AI agents do not replace people; they change the nature of their work. Instead of filling out forms and clicking between systems, people can focus on exceptions, non-standard situations, and strategic decisions. On work where they add value.
What can’t the agent do yet? Situations requiring empathy, complex ethical judgments, or creative solutions without precedent. But this list is rapidly shrinking.
How agents talk to each other: the MCP protocol as a universal language for AI
One of the reasons AI agents finally work in practice is the Model Context Protocol (MCP) technology standard. It is called “USB-C for artificial intelligence.” Before USB-C, every device needed its own cable. Before MCP, every AI integration with a company’s system required its own custom connection. If you are familiar with the principles of APIs, imagine MCP as a universal API layer through which AI communicates with any corporate system uniformly.
MCP was created by Anthropic (the creator of the Claude AI model) as an open standard. And during 2025, it was also adopted by OpenAI, Google DeepMind, and Microsoft. Today, it is the de facto standard for connecting AI to enterprise systems.
In practice, this means that an AI agent can “see” into your CRM, ERP, accounting system, or database via MCP. You don’t need to build a special connector for each connection. One protocol, one method of communication, multiple systems. For companies, this means significantly faster and cheaper deployment of AI agents. Instead of months of integration work, we’re talking about weeks.
Why most companies have not yet succeeded with agents
According to McKinsey, only 23% of companies are actually scaling AI agents within their organization. Another 39% remain in the experimental phase. Pilot project, management presentation, applause, and then silence. Why?
Many companies make one mistake. They take an existing process, slap AI on it, and wait for miracles. But a process designed for humans doesn’t work well with AI. A typical example: a company wants an AI agent to approve invoices. But their approval process has nine steps, three signatures, and four exceptions that arose historically, and no one knows why.
The correct approach is the opposite. First, look at the process itself. Which of the nine steps actually add value? Which exceptions are unnecessary? Only then should you design how the agent will take over the simplified process.
The second common mistake: companies start with too broad a scope. They want to automate the entire customer service department or the entire onboarding process for new employees right away. Try a different approach. Choose one specific process where the agent can deliver quick, measurable results. For example, sorting and routing incoming requests or automatically preparing reports from data that someone currently compiles manually from five tables.
What companies that are successful are betting on
Companies that successfully deploy AI agents have a few things in common. First, clearly defined boundaries. The agent knows what it can do on its own, where it should escalate to a human, and where it should not act at all. This framework is called guardrails, and without it, deploying an AI agent is risky.
Second, they build on an architecture that accounts for multiple agents. One agent analyzes, another executes, and a third checks. This approach (multi-agent orchestration) is more robust than relying on a single “super-intelligent” agent that does everything.
Third, they measure results from day one. How much time has been saved? How many requests did the agent handle without human intervention? What is the error rate? Without numbers, you have no argument for expansion. To run a company, you need concrete numbers, not general fairy tales about the potential of AI.
This approach has proven effective in practice. When companies design AI solutions, the best results come from conducting a process audit at the outset, selecting a single pilot scenario, and setting clear metrics for success.
Security and compliance
Whenever we talk about AI agents that autonomously access corporate systems, security is on the table. Rightly so. An agent that has access to CRM and accounting systems must be subject to the same security rules as an employee. Perhaps even stricter ones.
In practice, this means role-based access control, logging of all actions, encryption of communications, and regular audits. Companies in regulated industries (finance, insurance, energy) must also ensure compliance with legislation. The European AI Act is being introduced gradually: rules for general AI models will apply from August 2025, and obligations for high-risk AI systems will be added from August 2026. Anyone who underestimates these risks not only faces a security incident but also regulatory sanctions.
The good news is that modern platforms and the MCP standard take security into account from the outset. Access rights, audit trails, encryption—all of this is part of the design, not an afterthought.
How to get started?
You may be wondering where to start. Here is a procedure that can serve as a guide:
Map the processes that cost you the most time. Not the most technologically interesting ones, but the most expensive ones. Processing customer requests, internal reports, document approval, and sorting incoming communications.
Select one process for a pilot project. Ideally, one where the rules are clear, the data is digital, and the results can be easily measured. Don’t complicate the beginning.
Redesign the AI process; don’t just patch over the old one. This is the kind of work where it pays to have an experienced partner on hand. It’s not just about deploying technology, but about rethinking the entire workflow.
Set guardrails and metrics. Where does the agent make decisions on its own? Where does it escalate? How do you measure success? Without this, there’s no point in starting.
Expand gradually. Is the pilot working? Great. Add another process. And another. This is how you build AI infrastructure that lasts.
2026 won’t wait
According to Gartner’s forecast, global spending on AI will reach $2.5 trillion this year, up over 40% year over year. Most of this money will go towards integrating AI into everyday business systems and processes. Companies are moving from experimentation to full-scale deployment.
Flexible companies have an interesting opportunity at this point. The market is mature enough for the technology to work and young enough for early adopters to gain a competitive advantage. But those who wait for everyone else to deploy AI agents will only be playing catch-up later.
Not every process can be solved by an agent, though. For complex, regulated environments, specific business logic, or deep integration with legacy systems, custom software development remains the best approach. The ideal approach often combines both: AI agents take over routine processes and, when required, run under the hood of a tailor-made solution.
At eMan, we have been dedicated to digital transformation for over 15 years. Today, AI is a natural part of the projects we do. If you are interested in AI agents, please contact us.
Frequently asked questions (FAQ)
What is an AI agent, and how does it differ from a chatbot?
An AI agent is an autonomous system that plans its own steps based on a specified goal, accesses company systems, performs actions, and verifies results. A chatbot answers questions, while an agent solves problems from start to finish.
How much does it cost to deploy an AI agent in a company?
A pilot project usually takes 6 to 8 weeks. The price depends on the complexity of the process and the number of systems the agent works with. We recommend starting with a smaller pilot and expanding based on measurable results.
Is deploying AI agents safe for companies in regulated industries?
Yes, if the solution is designed with security in mind from the outset. This includes access control, encryption, audit logs, and compliance with the European AI Act. The MCP standard natively supports these requirements.
What is the Model Context Protocol (MCP)?
MCP is an open standard for connecting AI to enterprise and other systems. It was created by Anthropic and adopted by OpenAI, Google, and Microsoft. In practice, it allows an AI agent to access Google Drive, Slack, email, CRM, ERP, or a company database through a unified interface without having to build each connection separately.
In which industries do AI agents have the greatest impact?
They deliver quick results in insurance (claims settlement), banking (application processing), energy (customer request management), logistics, and other areas where repetitive processes span multiple systems.
How long does it take for an AI agent to start delivering measurable results?
With a properly selected pilot process, the first measurable results can be seen within 8 to 12 weeks of the project launch, including analysis, design, and deployment.
