An employee uses an AI development tool to build a simple CRM over a few afternoons. To test it, they paste in a real customer list, upload some correspondence and then connect the application to the company's systems. The solution works, and the manager is pleased. Yet no one has clarified where the data has gone, who can access it or who holds the rights to use the software.
The GDPR question surrounding vibe coding is therefore not simply about whether personal data may be entered into a prompt. AI-assisted development based on natural-language instructions can bring together data processing, information security and employment law issues. In this instalment of “The 10 biggest risks of vibe coding”, we examine when a quick fix becomes an uncontrolled business risk.
1. Customer data in a prompt: this already constitutes data processing
Personal data may be entered into an AI development tool, but not automatically or without oversight. Uploading, storage by the provider and analysis can all constitute data processing operations. A finished application is not required: even a customer complaint pasted in for debugging can involve the AI provider in processing the data.
Besides names and email addresses, personal data can include order histories, case IDs, contact notes or log entries that can be linked to a customer. A complaint may contain health information, which is special category data. Database exports and screenshots also often contain more data than is necessary to complete the task.
The assessment must be based on purpose limitation, an appropriate lawful basis and data minimisation. The original lawful collection of customer data does not provide unrestricted authorisation to use any AI tool. However, not every such operation requires fresh consent: the decision must be based on the specific purpose, the existing lawful basis and the circumstances of the processing.
Use synthetic test data as the first choice for prototypes.
Do not assume that removing a name automatically anonymises the data: the remaining fields may still identify individuals.
Process live data only with a documented justification, in an approved tool and limited to the minimum necessary.
2. Who processes the data, where is it stored and how long is it retained?
The AI tool's contractual terms and technical operation must be reviewed together. A generic claim of “GDPR compliance” is not enough. A business is typically the controller if it determines the purpose of processing customer data. The provider may act as a processor for certain operations, while processing data for its own purposes in others. These roles must be determined based on how the service actually operates.
Data processing and model development
A controller–processor relationship requires contractual terms that comply with Article 28 of the GDPR. Sub-processors, adherence to instructions, security obligations and deletion rules must be examined. A separate question is whether the provider may use prompts or uploaded files for model development. Disabling training does not, by itself, prove that there is no storage, logging or human access.
Data transfers and retention
Processing or accessing data outside the EEA may raise international data transfer issues. An adequacy decision or other appropriate safeguards may be required; transfers based on standard contractual terms may also warrant an assessment of the circumstances and supplementary measures. Hosting in Europe does not, in itself, rule out access from third countries.
When reviewing retention, check conversation histories, logs, backups and uploaded files separately. Clarify what can be deleted, within what timeframe and how data subject requests can be fulfilled. Privacy information must also be aligned with actual processing practices.
A data protection impact assessment is not mandatory for every use of AI. Under Article 35 of the GDPR, it is required where processing is likely to result in a high risk. Large volumes of special category data, profiling or systematic monitoring, for example, may make a detailed assessment particularly warranted.
3. Shadow IT: when an experiment quietly becomes a production system
Shadow IT is the use of tools and systems outside an organisation's IT governance. Vibe coding has a low barrier to entry: an employee develops using a personal subscription, deploys the application in their own cloud account and then invites colleagues. By the time IT learns about it, real customer data is already being stored there.
A working interface does not prove that an application operates securely. An AI-generated application may contain faulty access controls, a publicly accessible database, a secret key exposed on the client side or a vulnerable dependency. These flaws are not exclusive to AI; the danger increases when they reach production without expert review.
Access: can each user see only the data needed for their tasks?
Secrets management: where are API keys, passwords and integration credentials stored?
Operations: who handles updates, monitoring and troubleshooting?
Recovery: are there verified backups and tested restore procedures?
Responsibility: who are the application's business and technical owners?
Any suspected unauthorised disclosure or access must be investigated as a potential personal data breach. The need to notify the supervisory authority must be assessed based on risk; where notification is required, the general deadline is 72 hours after becoming aware of the breach.
4. The employee who developed the CRM resigns: who owns the software?
Operations built around personal accounts and individual development efforts raise more than data protection concerns: the rights to use the software must also be clarified. This is a separate copyright issue. Consider an administrative employee who builds a CRM on their own initiative, even though development is not part of their job description. The company starts using the system, but the employee later leaves. Can they prohibit further use or demand payment for it?
The absence of development duties from the job description does not, by itself, determine who holds the rights. Under Hungarian copyright law, a key consideration is whether creating the work was part of the employee’s duties. This may be determined jointly by the employment contract, job description, lawful employer instructions and the actual circumstances. For software created as part of an employee’s duties, the general rule is that the employer acquires the economic rights upon delivery, unless otherwise agreed.
However, if the development genuinely fell outside the employee’s duties, the employer’s acquisition of rights cannot be considered automatic. The legal basis, scope and remuneration for use must be agreed separately. Using a company computer or creating the solution during working hours does not, on its own, replace this assessment.
With AI-generated code, a further question is which parts contain the individual, original human creative contribution required for copyright protection. The licences of incorporated components and the AI provider’s terms also matter. It cannot automatically be asserted that the entire program belongs exclusively to the employee who entered the prompts.
Taking customer data, using trade secrets or disabling the system are separate legal issues: a potential copyright claim does not provide blanket authorisation for these actions. The keys to prevention are written agreements on rights, company-managed accounts, documentation and mandatory handover procedures. Specific disputes require an individual legal assessment.
5. xFLOWer: controlled processes instead of individual improvisation
The risks discussed so far share a common lesson: data processing, access rights and responsibilities must be clarified before going live. The xFLOWer approach follows this process-centred philosophy: it starts with the business process, not the first working screen. What data is generated, who can work with it, what approvals are required, and what happens in exceptional situations?
When designing an xFLOWer-based solution, it is worth defining the following in advance:
the purposes of data processing and the necessary data fields;
roles and access rules;
review, approval and logging requirements;
retention, deletion and incident management steps;
the process owner, cover arrangements and change management procedures.
A workflow framework such as xFLOWer can offer a practical solution to uncontrolled data processing: business processes are mapped within a shared set of rules rather than in standalone, unchecked applications. This also requires auditing, testing and professional oversight of the framework and its specific settings. Data transfers, exports and deletion can be made subject to permissions and, where necessary, approval and appropriate logging; deletion tasks and responsible owners can be assigned to retention deadlines. This can reduce the risk of someone transferring data unnoticed or of customer data being retained for decades after it is no longer needed.
New interfaces connecting to existing systems must undergo the same checks: a shared framework does not automatically make external data connections secure. In this scenario, the employee primarily configures a process within a given software platform rather than redeveloping the entire CRM. This may reduce the legal uncertainty associated with individual development, but rights to custom extensions and customisation still need to be addressed.
This makes operations less dependent on a single employee’s personal account or memory. However, no platform on its own automatically guarantees GDPR compliance. Appropriate configuration, contractual arrangements and consistently applied internal policies are also required. The same data-flow-based assessment must be performed when evaluating xFLOWer.ai or any AI integration.
6. Summary: rapid development still requires corporate controls
So, can customer data be entered into AI development tools? Yes, but first you must clarify the purpose and legal basis of processing, the scope of the data required, and the conditions governing provider access, storage and deletion. Use synthetic test data as the first choice for prototypes. Before going live, security checks and agreements on usage rights are also essential: a working prototype is not yet an approved corporate system.
Start by creating an inventory of the AI tools and in-house applications already in use. Assign an owner to each system, review data flows, and put access rights and legal arrangements in order.
In a properly supervised xFLOWer-based workflow, data processing controls can become part of day-to-day operations: a controlled process can govern not only data use, but also its transfer and timely deletion.
Would you like to build on transparent business processes rather than individual improvisation? Speak with the Click On team about designing more controlled operations based on xFLOWer, and where AI can play a safely defined role.
This article is part of our comprehensive guide: The top 10 risks of vibe coding: a complete business 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.
