With vibe coding, we describe our application idea in natural language, and AI generates the code for it. AI agents, by contrast, perform tasks and can decide on the next step as they go. The two capabilities can be combined, but their use in an enterprise raises different questions.
What capabilities should we put in the hands of business professionals, and what IT controls are needed to turn an idea into a solution that works over the long term?
What if a company’s digital processes could be shaped by the people who know them best?
Will Everyone Really Be Vibe Coding?
Author: Zsolt Szederkényi, founder of Click On / xFLOWer
With AI, almost anyone can now build applications, create AI agents and automate tasks. I increasingly hear that this will soon be as commonplace a skill as using Excel.
But is that really what an HR manager, a procurement professional or a business process owner needs?
And how does an application built in a few hours become a secure, sustainable system that also works reliably in an enterprise environment?
What I Have Been Hearing Lately
Over the past few months, hardly an IT or digital transformation conference has gone by without the topic of vibe coding or AI agents coming up.
The pace of technological progress is incredible, and the demonstrations are impressive. Meanwhile, I am hearing increasingly interesting claims about how all this will work in companies.
Two have particularly stayed with me.
1. “You do not need to tell AI how to solve a task. Just define the goal, and the agents will figure out the solution.”
For example: increase our revenue!
This is no joke. It is a genuine and exciting direction in technology.
Agents can try different solutions, evaluate the results, adjust their strategies and try again.
The question is how much time and how many resources they need to find the right solution.
It is a bit like playing the lottery. In theory, after enough attempts, we might even find the right numbers. We just need the patience and funds to keep trying.
Of course, AI uses far more intelligent search and learning methods than that. The analogy simply highlights that knowing the goal does not, in itself, mean knowing the shortest path to the solution.
In robotics, for example, learning methods based on trial and feedback are indeed used.
Imagine telling a humanoid robot nothing more than this:
“Make sure the plates are clean!”
But we do not show it what counts as a plate, where to find them, how to use the dishwasher or what tools are available.
In the right learning environment, with plenty of attempts and feedback, it can arrive at a solution.
However, showing it how we usually perform this task can significantly accelerate learning.
That is why, alongside reinforcement learning, learning from human demonstrations also plays an important role in robotics research.
And I believe that automating enterprise back-office processes raises a very similar question.
Why should an AI agent have to work out how procurement approval works through lengthy trial and error when the procurement manager can explain it precisely?
Why should it have to rediscover the steps of an HR process when they are already established within the company and have been working for years?
AI offers enormous potential for autonomous problem-solving. But it does not have to start from scratch every time.
Companies already have a wealth of knowledge that these solutions can build on.
The other claim I often hear is:
2. “Soon, everyone will be creating agents. It will be as fundamental a skill as using Excel.”
The use of AI agents may indeed become that widespread.
Basic digital skills naturally become part of everyday work. Today, being able to send an email, use the SUM function in a spreadsheet or share an online document is no particular achievement.
We once considered these IT skills. Today, we would hardly put having a Gmail account on our CV.
But I am not convinced that this means everyone needs to develop applications and agents.
Perhaps the opposite will happen.
Technology will become so simple that people will need to concern themselves less and less with how it works technically.
And that has direct implications for enterprise digital transformation.
The first working application is only the beginning
One of the biggest attractions of vibe coding is how quickly it can take us from an idea to the first working prototype.
For example, suppose we want a procurement application.
We tell the AI to manage requests, approvals, and order statuses.
Before long, we have an application we can click through and try out.
The demo is a success. Everyone is enthusiastic.
Then we show it to the head of procurement, finance, IT, and a few actual users.
And the real work begins.
Figure 1 – The lifecycle of a hypothetical vibe coding project
Elapsed time | Milestone | Typical reaction |
|---|---|---|
Day 0 | The idea takes shape | “We need a procurement application!” |
1 business day | First working prototype | “Incredible, it already works!” |
Day 2 | Initial fixes | “Let's tweak this a little more.” |
Week 1 | First reality check | “There is still so much missing!” |
Month 1 | Expanding requirements | “Finance, procurement, and IT all want a few more things.” |
Month 2 | Go-live considerations | IT security, testing, integrations, operations |
From month 3 onward | Sustainability | “Who will maintain and further develop this?” |
This timeline is an illustrative scenario, not a research-backed industry average. The actual time required depends on the project's complexity and enterprise requirements.
Prototype development time can be reduced to a fraction of what it was, while many of the requirements for enterprise operations remain unchanged.
In fact, rapid prototypes can also generate additional requirements.
When users see how easily new features can be created, they naturally keep coming up with more ideas.
That is a good thing in itself.
The problem arises when the application's technological foundations were originally designed only for a simple prototype.
Rapid application generation does not necessarily mean faster enterprise deployment.
What enterprise IT professionals are concerned about
In an enterprise environment, an application must do far more than pass a successful trial with a handful of users.
Figure 2 – Typical enterprise risks of applications built through vibe coding
Area | The real question | Potential consequence |
|---|---|---|
IT security | Are the generated code and data handling secure? | Vulnerabilities, data leaks |
Access management | Who can view, modify or approve which data? | Unauthorized access |
Extensibility | What happens if fundamental changes to functionality are needed? | Architecture that is difficult to modify |
Maintainability | Who understands the system that has been created, and who can fix its bugs? | Growing technical debt |
Operational reliability | What happens in the event of an overload or system failure? | Downtime, data loss |
Integrations | How does it connect to ERP, HR or financial systems? | Incorrect or inconsistent data |
Auditability | Can you trace who made which decision and when? | Operations that cannot be audited |
Testability | Does it handle exceptional situations properly as well? | Hidden business logic errors |
Change management | How can a new version be deployed safely? | Regressions, operational issues |
These are, of course, not problems unique to vibe coding. They must be addressed in every enterprise software solution.
The difference is that a prototype built extremely quickly can easily create the impression that a significant part of the development work has already been completed.
Research also supports a cautious approach.
Veracode's 2025 study found known security vulnerabilities in 45% of the AI code generation tasks examined.
In Stack Overflow's 2025 Developer Survey, 66% of respondents reported being frustrated by AI solutions that are almost correct but still need fixing.
These findings do not mean that AI-built applications are necessarily worse. They do, however, highlight that rapid code generation alone does not guarantee security or correct functionality.
The question, then, is not simply how quickly we can generate an application.
Rather, it is how quickly we can turn it into a properly validated, sustainable and secure enterprise solution.
And how much does all this cost?
Producing the first working application can, in many cases, be surprisingly inexpensive.
An AI service subscription, some token usage, a few hours of work, and the prototype is ready.
From an enterprise perspective, however, it is the total lifecycle cost that matters.
Figure 3 – The visible and hidden costs of an AI-built application
Cost component | What resources might it require? |
|---|---|
AI models and tokens | Generation, modification, repeated attempts |
Internal staff | Refining requirements, testing, reporting issues |
Development experts | Architecture, code review, bug fixes |
IT security experts | Vulnerability assessments, access reviews |
Integrations | Connections to ERP, HR, finance and other systems |
Operations | Infrastructure, monitoring, backups, availability |
Further development | Changing business needs, version control, maintenance |
AI usage costs are visible and relatively easy to measure. The time spent by internal experts, security reviews and subsequent maintenance costs are much less so.
Agentic operation also raises the question of how many model calls, attempts, checks and corrections are needed.
For a well-defined business task, it is therefore worth considering whether autonomous problem-solving is truly necessary, or whether a predefined, reliable workflow offers a more efficient solution.
What really matters is not the price of the first working version, but the cost of operating the application at enterprise level and maintaining it over the long term.
Why reinvent what business professionals already know?
We have been digitizing and automating business processes for more than twenty years.
During that time, we have sat down with HR managers, procurement specialists, finance executives and process owners on countless occasions.
At the start of each project, we typically ask the same questions.
Who can initiate the process? What data needs to be recorded? Who can approve it? When are additional checks required? What should happen if someone is late or a condition is not met?
These professionals generally know the answers precisely.
They know how the company works.
In the traditional development model, this knowledge is turned into a specification, and the IT team then builds the application.
More than twenty years ago, the core idea behind the development of xFLOWer was already to enable business users to build working workflows and applications themselves using visual tools, without IT expertise.
AI has now taken this concept to the next level.
Today, users do not necessarily even need to learn how to use the process editor. Simply describe how you want the process to work.
And at this point, it is worth returning to agents.
Two approaches: devise the process or use existing knowledge?
Imagine two different tasks.
In the first scenario, an AI agent is given the goal of making the company's procurement more efficient.
It can explore options, collect data, develop recommendations and test different solutions.
This is a valuable approach, particularly when the solution is not yet known.
In the second scenario, the procurement manager explains exactly how they want operations to be structured.
Which approvals are required, which value thresholds apply, when financial sign-off is needed, who can stand in for the decision-maker and which data must be transferred to the ERP system.
In this case, the AI does not need to devise the process independently.
It needs to turn existing professional expertise into digital operations.
Figure 4 – Two paths to the same business goal
Goal-driven AI agent | AI workflow built on business expertise | |
|---|---|---|
Starting point | “Let's make procurement more efficient!” | “This is how procurement should work!” |
The AI's role | Finding, designing and evaluating potential solutions | Modeling and creating the specified process |
Operation | Dynamic decisions and tool use | Defined steps, rules and permissions |
Key benefit | Flexibility, independent problem-solving | Predictability, auditability |
Key challenge | Maintaining control over decision quality, costs and outcomes | Accurately specifying and maintaining the process |
Typical task | Research, analysis, problems in changing environments | HR, procurement, finance, approvals, administration |
The two approaches are not mutually exclusive.
AI agents can also work within a workflow. They can process documents, interpret incoming requests, develop recommendations or help handle exceptional situations.
The key is to choose the right solution for the right task.
Anthropic's professional guidance on designing agentic systems also emphasizes this distinction: workflows are more predictable for well-defined tasks, while agents can be beneficial for tasks requiring greater flexibility, often at higher cost and with longer execution times.
The value of intelligent automation does not necessarily come from AI devising every step itself. Rather, it comes from being able to automate the right steps in the right places.
When business expertise becomes an application directly
Consider an HR manager who wants to digitize new employee onboarding.
They explain that HR initiates the process, the manager approves the necessary equipment, IT prepares access rights and personnel administration prepares the documents.
Certain tasks can run in parallel, while others can only start after approval. If any task is delayed, the system sends an alert.
They can even describe this into a microphone.
The xFLOWer.ai can create a complete, executable workflow from a natural-language description, including its steps, decision points, forms and business rules.
The business manager can review the result, try it out, refine the rules and modify the process.
This requires no programming knowledge, nor does building the application independently require the involvement of xFLOWer developers.
The system runs on an established workflow platform that has been used in enterprise environments for many years.
This matters because a new HR, procurement or case management application does not require rebuilding the workflow engine, task management, access management and essential enterprise services every time.
Of course, the business logic, access permissions and integrations of the completed application still need to be reviewed and tested.
The company's security, access control and operational policies determine how the application is deployed to production.
However, business professionals can now create applications independently, without having to build new IT infrastructure for each application.
Three Paths to an Enterprise Application
It is worth comparing the three approaches side by side.
Figure 5 – Traditional development, vibe coding and xFLOWer.ai
Aspect | Traditional development | Vibe coding | |
|---|---|---|---|
Starting point | Business specification | Prompt, application idea | Natural-language process description |
Implementation | Development team | AI-generated code | AI-generated workflow, forms and rules |
Technology foundation | Custom or selected architecture | Varies by project | Established enterprise workflow platform |
Role of the business professional | Defining requirements, testing | Prompting, validation | Independent application building |
Extensibility | Depends on the architecture implemented | Depends on the quality of the generated system | Unified framework for workflow and rule management |
Access management | Custom-built or selected solution | Must be developed or integrated separately | Can build on existing platform capabilities |
Deployment | With enterprise controls | With established controls | Under corporate policies, on an existing platform |
Typical use | A broad range of custom systems | Various applications and prototypes | Back-office and business processes |
This comparison illustrates typical approaches. In every case, the actual outcome depends on the tools selected, the architecture and the quality of implementation.
The three models have different strengths.
Vibe coding is particularly exciting when we want to test custom solutions quickly or create new types of applications.
Autonomous agents can be particularly useful for tasks where the next step cannot always be defined in advance and the system needs to make decisions based on results.
Many corporate back-office processes, however, are different in nature.
When approving an invoice, we generally know what data is needed, who is authorized to decide, which rules apply and what needs to be logged.
It often makes sense to define these operational rules explicitly and execute them in a verifiable way.
AI and agents can, of course, take over many tasks here too: interpreting documents, preparing recommendations, validating data and handling exceptions.
But all of this requires a well-defined operating environment.
Not every application needs the same tool
Our xFLOWer.ai was not designed to create every conceivable type of software.
Our focus is on digitizing and automating companies’ internal, administrative and business processes.
HR, procurement, finance, contract management, customer complaints, approvals, document management, internal administration and related system integrations.
This is an exceptionally broad field in itself. Almost every company has processes that still rely on emails, Excel spreadsheets, manual data entry or coordination that is difficult to track.
The business expertise needed to digitize these processes is often available, but development time and IT resources are limited.
With xFLOWer.ai, we aim to change that.
xFLOWer is backed by more than twenty years of enterprise experience and a continuously evolving workflow platform used in production environments. Even before AI, we enabled business users to build working applications without IT expertise.
The arrival of AI has made this capability even more natural and accessible.
Much of the technological foundation is already in place. Users can focus on what they know best: how their own area should operate.
Will everyone really need to build agents?
I think the more interesting question is what they want to use them for.
A procurement manager’s goal is not necessarily to create new AI agents. They want to handle procurement faster, with greater transparency and less administration.
An HR manager does not necessarily need to become an application developer either. They want onboarding to be handled properly, on time and in a traceable way.
And a finance manager would probably welcome approvals no longer getting lost among emails and spreadsheets.
These professionals already have the most important expertise.
They know how the company operates, its rules, its problems and, often, the possible solutions.
Giving them the ability to build applications directly could significantly transform how companies approach digitalization.
And this matters not only when creating new applications.
A company's operations are constantly evolving. New organizational rules, approval procedures and workflows emerge. Digital applications must keep pace with these changes.
If process owners can directly modify and develop their own applications, the gap between business requirements and their implementation can be significantly reduced.
Of course, the expertise of IT professionals remains indispensable. Designing enterprise architecture, integrations, security and operations professionally remains an important task.
The goal is not to replace IT, but to enable IT and business teams to collaborate in new ways.
Business expertise can form the foundation of a new application development capability
Today's discussions about enterprise AI devote enormous attention to models, agents, prompts and generated code.
These are important technologies, and they are advancing rapidly.
But companies already have one of their greatest assets at hand: the expertise of professionals who understand their operations.
People who know exactly how an onboarding, procurement, complaint handling or logistics process should work.
For more than twenty years, we have been working to make it possible to turn this expertise into functioning digital processes without requiring IT knowledge.
AI is now bringing this possibility even closer to everyday business users.
Perhaps in a few years, using AI agents will indeed be second nature to everyone.
But I believe the next major step in enterprise digitalization will not be defined by how many people can generate agents or applications.
Instead, it will be defined by how many companies can quickly turn their own professionals' expertise into secure, sustainable digital processes.
With xFLOWer.ai, we aim to provide a tool to make this possible.
Because the people who understand a company's operations best should play a leading role in shaping its digital operations.
Discover how a business idea can become a working workflow:
Professional sources referenced
Veracode – 2025 GenAI Code Security Report
https://www.veracode.com/blog/genai-code-security-report/
An examination of the security risks of AI-generated code, including vulnerabilities identified in 45% of the tasks tested.
Stack Overflow – 2025 Developer Survey, AI
https://survey.stackoverflow.co/2025/ai
The use of AI tools, experiences with the quality of generated code and developer trust.
Anthropic – Building Effective Agents
https://www.anthropic.com/research/building-effective-agents
The differences between workflow-based and autonomous agent systems, with considerations around complexity, costs and predictability.
