Vibe coding means describing what you want to an AI and letting it write the code, without reading it line by line. In business, it works very well for prototypes and small internal tools, and it most often breaks on security, real data and monitoring when the tool goes into production. This guide is for managers whose teams are already building applications this way.
We are not against it: at Codito, we use these tools every day, and we train managers to use them. The useful question is when a tool built in a few days deserves the controls of enterprise software. For the governance of Claude Code in particular, see our dedicated guide.
What works with vibe coding?
Vibe coding works when the stakes are low, the scope is narrow and the user is close to the person building. You then get in a few days what would otherwise have waited months in a development queue.
- The decision prototype. Showing a clickable version before paying for development changes the quality of the specification.
- The low-stakes internal tool. A tracking table, a calculator, a document generator used by 5 people.
- Personal automation. A script that cleans a file and that its author knows how to rerun.
- Business exploration. The manager understands what is feasible and talks better to service providers.
At Gustave Rénovation, a general building contractor that we support, the manager was trained in Claude Code, including key and environment variable management, Git and going live. He is developing the application that handles his quotes, invoices, schedule and subcontractors himself. This case shows what becomes possible with a real framework and real learning, and it does not mean an identical result elsewhere.
What breaks in production?
In production, vibe coding breaks in 5 places: security, data, load, dependence on one person and monitoring. The demo works, which is reassuring, yet these 5 points are not visible on screen.
| What breaks | What it looks like | When it is discovered |
|---|---|---|
| Access and security | A page or a database readable without permission, an API key in code visible from the browser | During a review, or an incident |
| Real data | No tested backup, a migration that overwrites rows, the same database for testing and production | The day of the loss |
| Load and edge cases | Everything is fine with 3 users, and it locks up at 50 or on an unusual file | On rollout to the whole team |
| Dependence on one person | Nobody else knows how the tool is built, or where the accounts are | When that person leaves or is absent |
| Monitoring | Nobody is alerted to an outage, an error or a rise in costs | When a customer calls |
One fix often breaks another, because the AI modifies the code without keeping an overall view of it unless it is made to. Without automated tests, every request becomes a bet on what was working the day before.
What do the figures and incidents say?
The security risk is documented: in its 2025 report, the security vendor Veracode tested 80 coding tasks on more than 100 large models and found that AI introduces flaws in 45% of cases. This figure comes from a player that sells security and dates from 2025: models have evolved since. It remains of interest for a simple reason: code that works can still be vulnerable.
The most widely discussed incident dates from July 2025. The founder of SaaStr publicly recounted that Replit's agent had deleted his project's database during a code freeze. According to The Stack, Replit's chief executive called the incident unacceptable and announced the automatic separation of development and production databases. The exact sequence of events is still debated, and this is only one case. The lesson remains useful: an agent must not be able to touch production on its own.
Anthropic, for its part, writes in its documentation that the user remains responsible for reviewing the proposed code and commands before approving them.
When does a tool change category?
A tool changes category as soon as it processes personal data, touches customers or money, or the business would stop without it. This change is what should trigger the checks, whatever the number of lines of code. Sort each tool into 3 levels:
| Level | Example | Expected control |
|---|---|---|
| 1. Sandbox | Prototype, fictitious data, 1 user | Professional accounts, no real secret |
| 2. Internal tool | Non-sensitive internal data, small team | Git, backups, named access, owner, security review |
| 3. Critical system | Personal data, customers, payments, invoicing | The 10 checks below, independent technical review, monitoring and a commitment to take incidents in charge |
A tool often slides from level 1 to level 3 without anyone deciding: the prototype is liked, real customers are put into it, and 6 months later invoicing depends on it. Take an inventory every quarter.
What checks should you put in place?
10 checks are enough to cover the essentials before an AI-built application goes into production, and a technical person should verify them. They do not require rewriting everything.
- Inventory. Which tools exist, who built them, where they run, what they touch.
- Git repository and review. All code is versioned; sensitive changes are reviewed by a human.
- Separate environments. Distinct test and production, with fictitious data in test, and no direct access by the agent to production.
- Access. Proper authentication, permissions checked on the server side, database access rules tested with an unauthorised account.
- Secrets. No key in the code or in the browser, storage in a vault, rotation after any possible leak.
- Tested backups. A backup that has never been restored does not count.
- Tests. Essential scenarios automated, and rerun before every release.
- Dependencies and vulnerabilities. An automatic scan, and a security review for the sensitive parts.
- Monitoring. Availability, errors and API costs, with an alert that reaches a named person.
- Documentation and handover. Where the accounts are, how to deploy, how to roll back, and who takes over.
Add a compliance check: if the tool processes personal data, the GDPR applies, and your data protection officer or your lawyer must know about it. In a company with at least 50 employees, a tool deployed to teams may also fall under consultation of the CSE (the works council) on new technologies. Our AI charter template includes a clause and a register for these tools.
Who carries the responsibility when it goes down at 9 a.m. on a Monday?
Your company carries the responsibility towards its customers and employees: what remains is to decide which person receives the alert at 9 a.m. and what they are allowed to do. Ask yourself 4 questions, without waiting for the incident.
- If the application stops at 9 a.m. on Monday, who sees it first, and how?
- Does that person know how to go back to the previous version, alone, in under an hour?
- If they are absent, who takes over and where are the access details?
- Can you prove to a large account how the application is secured?
If you can answer for everything, keep the system in-house: that is very good news, and a part-time AI lead may be enough to keep your bearings. If one of the answers is "nobody" or "I don't know", the system is running without a safety net.
When should you turn to the Enterprise Partnership?
The Codito Ergo Sum Enterprise Partnership suits organisations from around €1 million in turnover whose business relies on systems built quickly, without governance or monitoring. On quotation, from €2,000 excl. VAT per month, with no commitment, it covers monitoring and maintenance, security and compliance, documentation and training, and a development allowance to evolve what already exists.
The first month is used for setup and is invoiced as normal: inventory, taking over the access details, putting systems under monitoring, and a first Health Score out of 10. The commitment covers taking incidents in charge during business hours, not a resolution time. Reversibility is included, with an exit plan. A fuller analysis of maintenance is available on our blog.
It is poorly suited if nothing is running in production yet: the Premium Partnership, at €799 excl. VAT per month, with no commitment and cancel at any time, trains your teams to build properly, with 4 hours a month split as you wish. For a new project with a clear specification, look at Custom Development, on quotation from €3,000 excl. VAT. If you no longer know what is running in your company, an AI Audit maps it out. To discuss it, see the offer or let's talk about your ecosystem.
Sources (checked on 4 November 2026)
- Veracode, 2025 GenAI Code Security Report: official press release
- The Stack, Replit and SaaStr incident: article
- Anthropic: Claude Code security documentation
- French Labour Code, Article L2312-8


