Part I The decision: does it make sense?
Chapter 3: Can a colleague with AI build our software?
Today, a capable colleague with a tool like Claude Code can build a working app in a few days. This is called vibe coding: you describe what you want to the AI in plain language, and it writes the code. But writing the app is the smaller part of the work. The harder part is ensuring security, connecting to production data, and maintaining it for years to come, even after the author leaves. The best use is a clickable prototype that becomes the brief for professional development.
What AI really speeds up, and what it does not
Study results vary with the type of task and the developers’ experience. In a lab experiment with a single simple task, developers using GitHub Copilot finished 55.8% faster Peng et al., 2023. In three experiments in real workplaces, 4,867 developers using AI completed 26% more tasks. Less experienced developers saw a bigger benefit Cui et al., 2025.
A 2025 randomized controlled trial by METR, however, showed the opposite for experienced developers working on open-source projects they knew well: with AI they were 19% slower, while they themselves believed they had been 20% faster METR, 2025. The feeling of speed may not match reality.
DORA describes AI as an amplifier: it strengthens both the strengths and the weaknesses of an organization DORA, 2025. Without tests, reviews, and rules, it can also speed up the creation of bugs.
The capabilities of models and AI agents are growing fast. We are writing this section in October 2026, and in a few months it may look different. One thing still holds, though: as soon as an app works with sensitive data or signs users in, an expert should oversee it.
Writing the code is the smaller part
A significant share of the costs over a software system’s whole lifecycle is maintenance: fixes, updates, and changes. A classic reference work puts it at 40–80% of all costs Glass, 2002 (chapter 4). An app built “for free over a weekend” has years of work ahead of it that nobody is talking about yet.
On top of that, a layperson cannot spot a bug in AI-generated code. Even professionals do not trust the output blindly: 46% of developers distrust its accuracy, up from 31% a year earlier Stack Overflow, 2025.
Shadow IT and security holes
Employees build tools without IT knowing. The result is untracked apps containing personal data, with nobody to stand in for the author and with unclear responsibility for GDPR and cybersecurity. The organization bears the responsibility, not the colleague.
In 2025, IBM found that 97% of organizations that had an AI-related security incident lacked proper AI access controls. One in five organizations had an incident caused by an AI tool IT did not know about, and the damage was then up to USD 670,000 higher, roughly CZK 14 million at CZK 21/USD IBM, 2025.
Quickly written apps typically lack access management, reliable sign-in, and event logs. Apps with AI add prompt injection, excessive AI agent permissions, and leaks of sensitive data OWASP, 2025. Even a single shared account with access to a production database can cause a serious problem.
An AI model on your own hardware
A language model does not have to run at a cloud provider. An organization can run an open-weight model (one with published weights) on its own server or in a private cloud. Documents then never leave the organization, and it has full control over the data and the model. This, however, requires powerful hardware and experts EDPB SPE, 2025.
Where it fits:
- anonymizing or pseudonymizing a document before you send it to a commercial AI: the local model replaces names, addresses, or contract numbers with placeholders,
- summarizing and sorting sensitive documents, such as contracts or personnel files,
- search and answers over internal documents.
Advantages:
- the data stays inside the organization, making it easier to comply with GDPR and NDAs,
- you do not depend on the provider’s terms and prices,
- with a large volume of work, costs are more predictable.
Disadvantages:
- even the best open models lag behind top commercial ones by several months Epoch AI, 2026, and your own server will usually run a smaller, weaker model,
- buying a server with graphics cards (GPUs) and running it,
- updates, security, and monitoring are up to you, not the provider,
- anonymization by a model is not fully reliable, so a person checks the result before it is sent.
And one more trap: according to the Municipal Court in Prague, AI output is not a copyrighted work; only a human can be an author Prague Municipal Court, 10 C 13/2023. The court dealt with an image, not code, but purely generated code may not be protected. How to handle this with a vendor is covered in chapter 11.
Green: go ahead yourself
A personal or one-off tool, no sensitive data, nothing depends on it.
- a one-off analysis of a spreadsheet without personal data
- file conversion, personal automation
- a simple app for a repetitive task that runs only on your computer and does not work with sensitive data
- a clickable prototype to refine the brief
- screen mockups and a demo for management
Amber: only with IT and rules
A tool for a team or connected to company data, but one that key operations do not depend on.
- team record keeping or approvals for a few colleagues
- reports that only read from company systems through their own account
- low-code apps in a managed environment
- conditions: registration, owner and deputy, backups, code review, a date to decide “scrap it or rebuild it properly”
Red: professional development
Operations depend on it, it contains personal data, or people outside the organization use it.
Who will fix it when your colleague leaves?
Every amber app needs a deputy for the author, a one-page description (what it does, where it gets its data, how to run it), and its code in the organization’s repository, not on a laptop. Otherwise, within a year a useful tool can turn into an app that nobody understands and nobody dares to change.
The best use: a prototype as the brief
A product owner or analyst builds a clickable prototype in a few days, tries it out with colleagues, and hands it to the vendor. The prototype narrows the cone of uncertainty even before the RFP (chapter 1). But it is a brief, not the foundation of the production system. Add a short description: who uses the app, what they do in it, what data it works with, and what the prototype deliberately leaves out. The vendor then does not have to guess, and you get a more accurate proposal (chapter 6).
Do not ban AI across the organization. A ban can lead employees to use it in secret. The AI literacy obligation under the AI Act has applied since February 2, 2025; from July 27, 2026, Regulation (EU) 2026/1744 relaxed it Regulation (EU) 2026/1744. One page of rules is enough:
- approved tools and tools you must not use,
- what does not go into AI: personal data, trade secrets, passwords, and access keys,
- traffic lights: green, do it yourself; amber, report it to IT; red, a proper project,
- where apps are registered and who their owner and deputy are,
- training and contacts for IT and security.
| How it goes wrong | How to do it right |
|---|---|
| How it goes wrongA colleague builds an app over a weekend, and a month later invoicing depends on it. | How to do it rightAnything that operations will rely on goes through IT and professional development. The colleague’s version serves as a prototype. |
| How it goes wrongNobody knows how many such tools are running in the organization. | How to do it rightA list of apps, each with an owner and a deputy. |
| How it goes wrongA customer export is pasted into a public AI tool. | How to do it rightPersonal and confidential data only goes into approved tools that do not train on it. |
| How it goes wrongThe app accesses the production database under a single shared account. | How to do it rightRead-only, a separate account, role-based permissions, access logging. |
| How it goes wrong“It works for me” as the only test. | How to do it rightCode review by an experienced developer and a test on anonymized data. |
| How it goes wrongAI is banned in the organization. | How to do it rightAI is allowed, with rules and training. |
- Public authorities often fall under the Cybersecurity Act (Act No. 264/2025 Coll.). An untracked app working with the authority’s data then also becomes a problem for meeting the requirements on asset, access, and supplier management.
- Do not put citizens’ personal data into public AI tools. Assess where the provider processes the data and whether it trains on it.
- Under the AI Act, AI that decides on access to public services and benefits is a high-risk AI system. The obligations for such systems apply from December 2, 2027 Regulation (EU) 2026/1744.
- A public authority will find it hard to keep its own developer. Only 6% of ICT specialists work in public administration, education, and healthcare, and the average salary in the public pay sector is CZK 58,000 compared with CZK 102,000 in the private wage sector ČSÚ, 2026. “The colleague who programmed it” is an even bigger risk here.
- A prototype from an in-house workshop is a good basis for a preliminary market consultation under Section 33 of the Public Procurement Act (ZZVZ).