AI in development

AI writes your code faster. But a person stands behind the result, not the model

We use AI every day and see no reason to hide it. This page is about where it helps us, where we do not trust it, and what must happen before your code reaches people.

How we create code
Our values

Eight rules that apply here

Anyone can vibe-code a website over a weekend today. The difference is what happens next – and who stands behind it.

A person is responsible for the result, not the model.

AI generates and hunts bugs, but the final word and signature always belong to a senior developer.

Security is not an optional feature.

Every AI-generated piece of code goes through the same security audit as code written by a person.

The code belongs to you, not the tool.

No lock-in to a prompt-based platform. The code and documentation are 100 % yours.

Speed is a side effect, not the goal.

We deliver fast development with a guarantee that what we build will also last.

Architecture cannot be generated by a prompt.

We design it upfront and integrate AI into it with guidance, so it does not fall apart at the first deviation.

Decisions with real impact belong to a person.

Finances, access, anything irreversible. AI proposes, a person decides.

We do not leave large codebases to AI without supervision.

On extensive and taken-over projects we guide it hand in hand, with clear standards and senior oversight.

AI cannot learn company standards on its own.

We have processes to teach it. And what does not pass, we rewrite.

AI does not write code instead of people here. It writes it faster – but a senior developer stands behind it.

Jan Tölg

Jan Tölg
CEO and founder of Involve

How we work with AI

How code review actually works here

Quality is not held by the tool, but by people and steps that cannot be skipped.

  1. Requirements and context come from a person, not the model

    Before we generate anything, it is clear which architecture it belongs to and which project standards apply. AI does not get free rein — it gets a framework.

  2. AI writes the first version, a developer reads it line by line

    Not diagonally. We check logic, edge cases, and places where the model likes to invent its own solutions.

  3. Security and data handling separately

    Inputs, permissions, database queries. AI-generated code breaks most often here, so it gets its own review round.

  4. Tests and runs on real data

    Automated tests plus manual walkthrough of what changed. If something behaves differently than before, we fix it before deployment.

  5. Signed off by a named senior

    Every merge has a name. A specific person who stands behind it even six months later.

Transparency

Where your data goes

With us, you know exactly.

QuestionIn shortIn detail

Do you send our code to the model?

Only the part being worked on, with the context it needs.

We never send customer production data, credentials, or keys in prompts.

Which models do you use and where do they run?

Anthropic (Claude) and OpenAI, in data centres in the EU or USA.

We work via enterprise API and tools such as Cursor or Claude Code. Where a project requires it, we deploy models locally.

Does anyone train on our code?

No.

We only use enterprise modes with training disabled and minimal data retention. We do not use free tool versions that keep your data.

Who can access our code?

Only the team working on the project. We sign NDAs.

We review access regularly and revoke it when a project ends or someone leaves, including tokens for AI tools.

Do not want AI on your project at all? That is possible too – see the questions below.

Where we do not let AI in

Speed is a side effect. These are four places where we deliberately sacrifice it.

AI does not write tests for its own code

The model that wrote a function will also write a test that fits it perfectly. Tests for AI code are therefore written or at least reviewed by a person. Otherwise the code approves itself.

Code no one on the team understands does not go to production

"It works and we do not know why" is the most common and most expensive state with AI code. When no one can explain what a part does, it gets rewritten.

Irreversible changes to production data

Database migrations, bulk edits, deletions. AI prepares the procedure, a person runs it, always with a rollback path. We do not rely on the model not missing on live data.

Dependencies no one verified

AI will happily suggest a library that either does not exist or has not been maintained for years. Every new package in a project goes through a human.

Numbers from practice

What AI actually saves us

The numbers come from specific projects of our team over the past twelve months.

TaskWithout AIWith AI
Finding a bug in the database / SQL1–2 hoursapprox. 10 minutes
Payment gateway implementation16 hours8 hours
Development task (regular agenda)9 hours4 hours
Converting design from Figma to componentshoursapprox. 15 minutes
Clickable UX/UI prototypeexternal agency, days1–2 hours in-house
Functional demo on the client’s real websitepractically impossibleminutes

What did not get faster

Review. On the first pass only part of the AI code passes, the rest is adjusted or rewritten from scratch – and review today takes more time than before. Review is the only part of the process we do not want to speed up.

The time saved does not go into our margin, but into your offer. A payment gateway in eight hours instead of sixteen counts the same for us as for you. And above all: bug hunting shortened from hours to minutes means a senior developer is not fixing typos, but solving architecture and decisions AI cannot handle.

Ondřej Tölg

Ondřej Tölg
Project manager

Honest answers to questions you might be asking

  • To a person. Full stop. AI is a turbo boost for our efficiency, not a filter between us and you.

  • You can, and we respect that. We put it in the contract and development runs purely by humans. Be prepared though that for tasks AI normally shortens, it will take longer and cost more. We will tell you upfront by how much.

  • We are. Specifically the senior developer who approved the code, and Involve as the vendor. From the contract and warranty perspective it does not matter whether the first version was written by a person or a model – responsibility is not shared with the tool.

  • From the trail review leaves behind. In project history you see specific names, pull request comments, and changes that came out of review. We can give you repository access and walk through it with you. When someone just ran code through a model, it shows in history – those questions and edits are missing.

Have questions?
Let's talk

This page is not an offer, but a description of how we work. If you want to know where specific code stands – whether AI, an agency, or someone who left long ago wrote it – we will take a look. See our AI code audit, or write directly to Honza.

Jan Tölg

Jan Tölg

Founder & CEO, Involve

+420 774 000 025
jan.tolg@involve.cz
LinkedIn