From vibe coding prototype to production system: deployment in 7 controlled steps

A cutaway view beneath a miniature workstation reveals cabling, an enclosed data storage unit, a backup unit and a maintenance space.

A colleague builds a CRM with artificial intelligence over a few afternoons. It has a customer list, quote tracking and reminders, and the manager can see that it works. The critical questions often arise only later. Where is the data? Who can access it? Who fixes the system? And what happens if the colleague resigns tomorrow?

The deployment of a vibe coding application to production is not simply a matter of clicking a publish button. It also requires business, information security, data protection and legal acceptance. Within the topic of The 10 biggest risks of vibe coding, this is the point where rapid experimentation becomes a lasting corporate commitment. The following seven steps help make going live a controlled decision rather than a gamble.

1. Assign responsibilities and make shadow IT visible

Shadow IT is technology used for business purposes that operates outside official IT governance. An AI-built CRM running under a personal subscription is a typical example. A manager's awareness of it does not necessarily mean that its operation, access management or data processing arrangements are properly governed.

The first task is not to blame the initiator, but to register the system and separate responsibilities:

  • Business owner: defines what the application must support.
  • Technical owner: handles maintenance, backups and operations.
  • Security and data protection owner: assesses risks, involving a specialist where necessary.
  • Decision-maker: approves or rejects deployment to production, with a documented decision.

In a small company, one person may hold several roles, but none can remain unassigned. Having the application reviewed solely by its creator is not an adequate control either. Until the basic requirements are met, the prototype should not receive any additional live customer data.

A single metal chain link stands against a white background, with its red locking component not yet closing the gap.

2. Take over the system and clarify usage rights

Handover is not simply the delivery of a ZIP file containing source code. It requires the code repository and version history, deployment instructions, the data model, a list of dependencies, and access to hosting, the domain and external services. Secrets must be transferred securely, and keys tied to personal accounts should be replaced.

What happens if the secretary who built the CRM resigns?

The answer is neither that the software automatically belongs to the employee nor that it definitely belongs to the company. Under Hungarian copyright law, key factors include whether creating the work was a duty arising from the employment relationship and what the agreements stipulate. The job description is important evidence, but it does not necessarily settle the issue on its own: the specific assignment and the circumstances in which the work was performed also matter.

If development was not part of the employee's duties and no appropriate legal agreement exists, the employer cannot automatically assume it has acquired the necessary economic rights. A separate licence or remuneration arrangement may be required. For AI-generated code, the human creative contribution, the provider's terms and the licences of incorporated components must also be examined.

However, any rights to the software do not in themselves entitle a departing employee to take customer data, use trade secrets or disable the operational system. Before going live, it is advisable to have a lawyer clarify the terms of use, modification, further development and transfer. This is general information, not case-specific legal advice.

3. Test functionality, security and data processing

A successful demonstration only confirms that one selected path works. It does not prove that another user cannot see your customers or that a faulty request cannot delete the database. AI-generated code must be reviewed just like any other software development.

  • Permissions: does the server also check access, or does the interface merely hide buttons?
  • Data protection: does personal data end up in logs, AI prompts or external services without justification?
  • Technical security: are there hardcoded passwords, vulnerable packages or exploitable inputs?
  • Business functionality: what happens with duplicate customers, interrupted saves or simultaneous edits?

Document the purpose of processing, legal basis, retention period and handling of data subject requests. Check data processing agreements and any transfers to third countries. A GDPR data protection impact assessment is not mandatory for every application, but may be required for processing likely to pose a high risk. Wherever possible, use synthetic or properly anonymised data for testing.

Two colleagues at an office desk use documents and a laptop to jointly review the application's data handling terms.

4. Create and test a recovery plan

Rollback should not be planned after the first outage. Before the pilot, you should already be able to explain how to return to the previous way of working, who makes that decision and what happens to data recorded in the meantime.

Having a backup is not the same as being able to recover. A backup only provides a reliable fallback if restoration has been tested and the potential data loss or downtime is understood. Application and database versions must be managed together: after a database change, simply restoring the previous code may not be enough.

  • Have a tested backup and a documented recovery procedure.
  • Define acceptable downtime and data loss.
  • Document shutdown criteria and temporary working procedures.
  • Plan how to reconcile data generated during the pilot.

For a CRM, controlled use of the previous system could provide such a temporary solution. However, entering data in parallel introduces another source of errors unless specific reconciliation rules are in place.

5. Launch a limited, measurable pilot

A pilot is not a quiet, full-scale production rollout. It should have a defined user group, a limited process scope, a start date and exit criteria. Real personal data may only be used here if appropriate data processing and security arrangements are in place: “we are just trying it out” is no exemption.

Decide in advance what to assess. For example: can a lead be managed from start to finish without data loss? Does delegation work correctly? Can a change be traced? Are error messages understandable? Alongside user feedback, event logs and support tickets should also provide evidence.

Set clear stopping criteria. The pilot must be suspended in the event of unauthorized data access, data loss or data transfers whose circumstances are unclear. After fixing the issues, repeat the affected tests. Urgent business needs are no substitute for verification.

Disorganized paper checklists on the left; an orderly digital workflow with a red stop gate and a fix-and-retest loop on the right.

6. Obtain documented go-live approval

Approval records that the responsible parties consider the launch acceptable based on the available evidence. It is not merely an administrative formality, nor is it the same as saying “the manager liked the demo.”

The following should be available for the go-live decision:

  • acceptance documentation and confirmed usage rights;
  • test results, the pilot evaluation and a list of unresolved issues;
  • access rules, data processing documentation and a rollback plan;
  • the designated operations owner, their deputy and support arrangements.

Identify the remaining risks explicitly. An awkward filter and a missing authorization check carry very different levels of risk. Assign an owner and a deadline to each accepted shortcoming; fixing serious security issues should be a prerequisite for launch. Significant subsequent changes require a new review, because approval is not unlimited authorization.

7. Monitor the system and reduce dependency on individuals

In production, monitor errors, failed login attempts, unusual data exports, backup results and external service outages. Every important alert should have a designated recipient and a response procedure. Avoid collecting sensitive data unnecessarily when logging events.

The offboarding process should include revoking access, rotating shared secrets and verifying knowledge transfer. A useful test is whether another specialist can deploy, restore and maintain the application using the documentation. If not, the key-person risk remains.

xFLOWer: implementing processes on an existing framework

A key advantage of the xFLOWer approach is that process implementation can build on an existing framework, rather than requiring every operational foundation to be created in a custom prototype. This allows attention to focus on process steps, responsibilities, permissions and exceptions.

Using a workflow framework such as xFLOWer can therefore help manage the transition from prototype to production. A business process mapped in the framework can have its own lifecycle: demonstration prototype, testing, training use and then approved production operation. Separate environments, data and access permissions must be established for these stages. This allows even the prototype to be developed under controlled conditions, rather than having to retrofit operational foundations into an application created without oversight.

This does not automatically guarantee security or regulatory compliance: configurations, integrations and data processing still need to be checked. However, it can reduce the scope of custom development and the concentration of knowledge in a single employee. During implementation, assess what the framework provides and where a genuinely custom solution is needed.

An unlabeled donut chart, bars and flow bands provide an overview of the shared system foundation and individual connections.

The security benefit depends on verifying that the underlying framework has been audited and tested, and ensuring that its operation is not overseen solely by the process creator. Custom development may be limited to interfaces with existing systems, while the process remains under central governance. This reduces shadow IT risk, but new interfaces must also be registered, tested and approved: connecting them to the framework does not, in itself, make them secure.

In summary: deploying a vibe coding application is controlled when responsibilities, rights, testing, rollback, the pilot, approval and monitoring are all properly addressed. A working screen is only the beginning; the goal is sustainable business operations that can be handed over. In a properly governed workflow framework, all of this can be built into the prototype lifecycle, rather than being left to a final pre-launch review.

Do you already have an in-house prototype? Talk to the Click On team about which processes would benefit from an xFLOWer foundation and which checkpoints are needed for a secure implementation.

This article is part of our comprehensive guide: The top 10 risks of vibe coding: a complete enterprise guide from security to copyright

A semicircular miniature inspection facility stands against a dark background, with a red rail connecting its storage, maintenance and recovery areas.

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.

Contact us

Let's talk about your processes

Tell us briefly which process you'd like to automate – we'll get back to you within 1 business day.