A colleague shows you the new CRM: you can add customers, change statuses, and even view a management report. There was no development project, just a few conversations with an AI tool. The demo is convincing. However, the next question is not when we should start using it, but: what do we know about the system's security, ownership and long-term maintainability?
The enthusiasm surrounding vibe coding is understandable: ideas that were previously difficult to implement quickly become solutions you can try out. The danger begins when we confuse being able to try something out with its suitability for production use. In this installment of “The 10 biggest risks of vibe coding,” we examine that boundary, particularly from the perspective of shadow IT and the legal status of employee-built software.
1. Vibe coding explained: from idea to working code
The meaning of vibe coding in this article is AI-driven software creation in which users shape the solution through natural-language instructions and judge the outcome primarily by its visible behavior, rather than through substantive review of the underlying code. Instructions might include: “add a customer list,” “send a reminder,” or “fix this bug.”
This is not the same as all AI-assisted development. In controlled development, AI generates suggestions and code, but professional oversight governs requirements, architecture, permissions and releases. In risky vibe coding, users accept a solution without being able to meaningfully review the underlying code or its implications.
For a prototype, this can be a useful starting point. For a live business system, however, speed does not replace accountability. The dividing line is not whether AI is used, but whether oversight is in place.
2. What does a working demo not prove?
A demo generally shows that a selected sequence of actions can be performed with certain data. It does not prove that the system also works correctly with invalid inputs, during concurrent use, or when faced with an unauthorized access attempt.
Data security: can customer data truly be viewed only by authorized users?
Data consistency: what happens if two people edit a record at the same time?
Fault tolerance: is an operation lost or duplicated after a connection drops?
Recoverability: are backups merely available, or has restoration also been tested?
Maintainability: can someone else understand, test and update the solution?
In a CRM, for example, a working “delete customer” button is not enough on its own. You need to clarify what happens to related tasks, what retention obligations apply, and who may perform the deletion. These are business and regulatory requirements, not merely interface details.
3. Security: flaws hidden behind the screen
AI-generated code may be adequate, but it may also contain serious flaws. Apparent protection is particularly dangerous: the interface hides an administrator button, while the server accepts the same request without any checks. A login page does not prove that appropriate access controls are in place for individual records, either.
Other risks include API keys copied into source code, publicly accessible databases, vulnerable third-party packages, or error logs containing customer data. If the application also uses AI functionality, you must separately assess whether external content can manipulate the AI's behavior and what data or operations it can access.
Before going live, at least the following checks are required:
Test server-side permissions using multiple roles and unauthorized access attempts.
Manage secrets securely, replace leaked keys and restrict access.
Conduct security reviews of code and dependencies, with documented responsibility for remediation.
Establish backup, restoration, logging and incident management procedures.
Asking the AI that generated the code to “check security” can be a useful step, but it does not replace an independent professional review.
4. Shadow IT: when the company CRM lives in a personal account
Shadow IT encompasses solutions created or used outside the oversight of IT governance. A CRM built using an employee's personal subscription can be a typical example: the company uses it without knowing exactly where it runs, who can access it, or who can shut it down.
The problem often begins with good intentions. Administrative work is slow, official development is taking time, and someone steps in to solve the problem. Over time, however, the temporary tool becomes the primary repository of customer data. The organization becomes dependent on a system with no designated owner.
Where personal data is involved, GDPR requirements do not disappear simply because the solution is an internal tool. You must clarify the purpose and legal basis of processing, providers' roles, the necessary data processing agreements, any international data transfers, and deletion rules. Live customer data pasted into an AI tool during development requires a separate assessment.
A ban is not necessarily the answer. Provide an approved experimentation environment, synthetic test data, an application inventory, and clear rules on when IT, information security and legal specialists must be involved.
5. The employee who built the CRM resigns: who owns the software?
Suppose an administrative employee builds a CRM with AI, then resigns or has their employment terminated. Development was not part of their job description. That alone does not mean either that the software definitely belongs to them or that it automatically belongs to the company.
The job description matters, but it is not the only factor
Under Hungarian copyright law, unless otherwise agreed, an employer generally acquires the economic rights to a work created in fulfilment of an employee’s duties when the work is delivered. Determining whether this applies requires examining the employment contract, actual duties, employer instructions, agreements and circumstances of delivery. Using a company laptop or developing software during working hours does not, on its own, settle the issue.
If development was not part of the employee’s duties and there is no appropriate assignment of rights or licence agreement, the company’s rights may become uncertain. This may give rise to a need for a separate agreement and remuneration, but it cannot be assumed in advance that any claim by the employee is well-founded.
Code, data and access are separate issues
AI involvement calls for further examination: which parts reflect human creativity eligible for copyright protection, what terms does the AI provider apply, and what licences govern the incorporated components? Copyright protection for purely machine-generated output cannot automatically be taken for granted.
Rights to the code do not, however, confer unrestricted control over customer data or trade secrets. An employee does not acquire rights to these simply by building the system. Hosting it under a personal account may give them technical control, but that is not the same as having legal authority over it.
As a preventive measure, rights, the handover of documentation and source code, the use of company accounts, and the revocation of access when an employee leaves should all be agreed in writing. In a specific dispute, employment law and copyright specialists should be consulted; this overview is not case-specific legal advice.
6. xFLOWer: we use AI to design processes, not develop standalone software
In the xFLOWer approach, the starting point is not a standalone AI-generated application, but a controlled business process. Who initiates the case? What data is required? Who approves it? What happens when an exception occurs? AI’s role is to support process design, not to substitute for an uncontrolled software development lifecycle.
With xFLOWer.ai, the first question to clarify is therefore: do you really need a new application, or do you need to organise the tasks, approvals and responsibilities within your existing customer management operations? If the latter is the goal, it is worth exploring a process-centric approach: We use AI to design processes, not develop standalone software. The deciding factor should be whether business requirements can be met within a verifiable framework.
Using workflow frameworks such as xFLOWer can therefore be a practical way to address the risks outlined above. From a security perspective, the aim is to model business processes within an audited, tested framework that is supervised independently of the employee designing the process, rather than rebuilding basic safeguards every time. With central governance, custom development requirements may be limited to an interface connecting to existing systems. This interface also requires approval and review, but it does not require running an entire application that is invisible to IT.
Data processing can also become more traceable: appropriate permissions, logging and retention rules make it possible to verify who deleted, transferred or exported data, and when data that is no longer needed is due for deletion. This supports GDPR compliance but does not, in itself, guarantee it. The legal starting point is also clearer: the employee is customising software used under a contract, rather than creating a standalone CRM. Even so, the platform’s usage rights, the legal status of customisations and integrations, and control over and access to data must still be settled in advance.
An advantage for cost planning is that the entire software framework is not rebuilt through repeated rounds of AI generation: licensing, configuration, integration and operating costs can be estimated in advance. Token costs for any AI usage must also be budgeted separately; a framework does not automatically mean a lower total cost.
The boundary between a prototype and production operation can also be governed. Processes within the framework can have separate demonstration, testing, training and production lifecycle stages, with appropriate data and approval checkpoints. This allows even the prototype to be developed under controlled conditions, while deployment to production remains subject to a separate decision and review.
Platform-based operations do not eliminate the need for oversight. Process ownership, permissions, data processing, integrations and change approvals still need to be addressed. The value lies in a shared, verifiable operational framework, not in the disappearance of accountability.
7. Summary: the demo is where assessment begins
Vibe coding can make an idea quick to test, but deploying it to production requires a separate decision. At the next demo, start with four questions: Who reviewed security? Are usage rights settled? Who is responsible for operations and recovery? Will company access remain assured after the creator leaves? Until these questions have clear answers, the CRM should remain in an isolated test environment using fictitious data.
A workflow framework such as xFLOWer does not remove oversight; instead, it allows us to bring the process, data processing and deployment requirements into a shared framework from the prototype stage onwards.
Looking to accelerate customer management or internal administration with AI? Talk to the Click On team about xFLOWer-based process design: let’s clarify operations, responsibilities and control points first, rather than focusing solely on the demo screens.
This article is part of our comprehensive guide: The Top 10 Risks of Vibe Coding: A Comprehensive Enterprise Guide from Security to Copyright
xFLOWer.ai
Design the process behind this article – for free!
AI reads the processes hidden in this article; open any of them in xFLOWer.ai with one click.
