Artificial intelligence

Writing code is easy now. Answering for it is not.

Anyone can generate an application in an afternoon today. What no tool hands out is the judgement of more than eleven years of the founders keeping critical software in production. Here AI sits on top of experience, never in its place.

AI does not level teams out. It pulls them apart.

A multiplier amplifies what it finds. On a team that knows the domain it multiplies the speed of the work. On someone who cannot yet tell a reversible decision from one that is not, it multiplies the number of wrong decisions and shortens the time in which they pile up.

The cost of those decisions does not show up in the demo. It shows up at your first traffic peak, in the sixth month when a requirement changes, and at your first audit. That is where more than eleven years of maintaining what we built ourselves turn into an advantage no tool hands you.

The same assignment, two routes

Moment	Without experience, with AI	With experience, with AI
The first prototype	An afternoon. It works and it looks good.	An afternoon. There is no difference here, and saying so is honest.
The second screen over the same data	The pattern gets copied and the business rule ends up written twice.	Where the rule lives is decided before it is written once.
The first traffic peak	The query with no index is discovered in production.	The index was there from the data model.
A requirement change in month six	It gets rewritten, because nobody knows what else depends on it.	It changes in one place, with the tests that prove it.
The first audit	There is no access log and no documented legal basis.	It was in the design: logging, retention and a data owner.
The third year	Nobody understands the system, not even the AI that generated it.	It is still in production, it gets updated, and it takes whatever team you put inside it.

What AI still does not decide

What gets stored and who can see it: The data model and the permission model are the hardest decision in your project to reverse: it is taken in the first month and you pay for it for five years.
Where a transaction begins and ends: A half-taken payment, stock deducted twice. The errors that do not show up in the demo come out of the transaction boundary.
Which migration can be done without cutting the service: Changing the schema with clients inside is a speciality of its own: order of steps, backward compatibility and a rollback ready.
Which dependency is accepted: Every suggested library is a maintenance commitment, a licence and an attack surface. Knowing which ones are still maintained five years from now cannot be deduced from the code.
What not to build: AI builds what it is asked for, even when the right answer is to withdraw the screen and save the whole module.
Who answers when it fails at 3am: That part is not delegated. It is what turns a deliverable into a service, and the reason we sign a contract with you.

AI is above all an amplifier: it magnifies the strengths of high-performing organisations and the dysfunctions of struggling ones. It increases delivery instability unless the surrounding system and culture change — and what goes with it is strong version control practices, working in small batches and a quality internal platform.

DORA, State of AI-assisted Software Development 2025, from nearly 5,000 responses

Where it enters our work

Reading legacy systems: Understanding a hundred thousand undocumented lines of your system stopped costing weeks. The dependency map is validated by a person who has maintained that kind of code, before it determines any architecture decision.
Tests on code that had none: Covering what already works before touching it. Every test is reviewed: a badly framed test gives false coverage, and that is worse than having none.
Repetitive, bounded work: Mechanical migrations, adapters, forms, the first version of the documentation. Defined task, verifiable result.
A second reading, never the only one: It finds the forgotten edge case before the person who signs the change reads it.

Where it stops

Architecture decisions: A model proposes them; the team discusses them and a person signs them.
Your clients' data: It does not enter third-party services. If a project needs AI in production, where it runs and what leaves the perimeter are agreed in writing.
Unread code: Nothing reaches production without someone on the team having validated it and being able to explain it without the tool open.
The invoice: The productivity the team gains turns into scope and schedule, not into billing the same work more times.

Model policy

Commercial providers: Your code does not train their models and your data is not retained beyond what is needed to process the task. Both excluded by contract.
Open models: Where your sector requires it or technical judgement justifies it, your code goes to no third party: they run under our own control.
Explicit authorisation: You decide use by use. Accepting AI for the analysis of your system does not authorise its use in, say, generating content from your data.

Regulatory framework

The EU AI Act: Regulation (EU) 2024/1689 applies in stages from 2 August 2026; the high-risk obligations were deferred by Regulation (EU) 2026/1744 to December 2027 and August 2028. This describes how we comply: it is not legal advice.
The US and Canada: No federal law of general scope: sectoral and state rules. Canada never passed its Artificial Intelligence and Data Act. We apply the same criteria as in Europe.
Latin America: No country yet has a law of equivalent scope to the European Regulation in force. What does apply is data protection law: Law 1581 in Colombia, LGPD in Brazil.
ISO/IEC 42001: Artificial intelligence management systems. Certification in progress with external audit, expected in 2027.

Where do we start?

Two ways to begin, depending on what you have in front of you today. Answered within one working day.

Talk to us