Alfonso Medela Book a call

What an AI project costs when Claude Code writes the code

Claude now writes more than 80% of Anthropic's code. What still costs money in an AI project, and when you do not need to hire anyone.

By Alfonso Medela
· 6 min read

According to Anthropic, in May 2026, more than 80% of the code merged to its codebase had been written by Claude. At first, it was hard to admit, but today it surprises no one that a data scientist uses Claude Code; if anything, the opposite would be surprising. I use it, too, especially to read repositories, write code, run tests, and fix whatever goes wrong.

Many clients, especially the more tech-savvy ones, ask me what’s the point of hiring someone like me to develop AI if they’re already capable of doing “anything” with Claude or ChatGPT. To be honest, over the past few months I’ve asked myself the same question. However, I don’t think AI is going to significantly diminish the value of professionals with extensive experience. And I’ll explain why.

I’m firmly convinced that the quality of what a person using AI delivers is directly proportional to their intelligence. And not only that: their domain knowledge also factors into that equation. Therefore, we could say that the result is a combination of the quality of the AI (architecture, model, etc.), the user’s intelligence, and their knowledge.

Claude Code at work

On the same page where it cites that 80% figure, Anthropic clarifies two points. First, its productivity figure, eight times more code per engineer, almost certainly overstates the actual improvement. And second, every change goes through an automated reviewer before being merged. Not even Anthropic lets its own code go through without scrutiny. As you’d expect.

Writing code was the visible part of the job. The rest always took more effort: deciding what to build, obtaining the right data, proving that it works, and responding when it fails. Claude Code greatly speeds up and improves code writing. The rest still depends on someone to make decisions and respond. And to make quality decisions and respond effectively, we need a human who knows what they’re doing: the human component of the equation.

Correct, but flawed

In Veracode’s March 2026 security tests, more than 95% of the code generated by the most widely used models was correct. However, only 55% was secure.

The code works in the demo. Someone with no programming or even domain knowledge is able to create something that looks functional, attractive, and high-quality. But then, in production, it leaves a door open that no one has reviewed. In addition to security, I’ve seen another major problem: scalability. Not just in terms of the number of users, but also in the code itself, which, if it grows without control or oversight, ends up becoming a Frankenstein’s monster that no one knows how to maintain.

The same thing happens with an AI model, but it’s more subtle. It can perform very well with your test data and fail on the very first day with real data. At Legit.Health, I learned this lesson with skin photos. The training images don’t resemble the ones coming from a patient’s phone, with different lighting, a different camera, and different skin. And, above all, they’re of much poorer quality than a photo taken by a healthcare professional. If you don’t evaluate the model with data like the real-world data, you don’t know what you’re dealing with. The tool can’t warn you, because it doesn’t know what data will arrive tomorrow.

An agent isn’t accountable for anything

In July 2025, Replit’s programming agent deleted the production database of a project that Jason Lemkin, founder of SaaStr, was testing. It contained records for more than 1,200 executives and 1,190 companies. The agent did this in the middle of a code freeze, despite being instructed not to touch anything. Lemkin recovered the data manually.

In a real company, that’s a lost week or a data breach that must be reported. It isn’t fixed with a better model. It’s fixed by someone who separates the environments, decides what the agent is allowed to touch, and reviews what goes into production.

If your AI helps diagnose, someone signs

Software with a medical purpose that assists in making a diagnosis is a medical device under the EU MDR and is, at a minimum, a Class IIa device. It requires clinical evaluation, post-market surveillance, and a manufacturer that is responsible for it. The EU AI Act adds its own requirements for high-risk systems. Following the extension approved in July 2026 (Regulation 2026/1744), the requirements take effect on December 2, 2027, for Annex III systems, and on August 2, 2028, for AI within regulated products, such as medical devices.

Claude Code can draft the technical documentation for you in an afternoon, but it cannot sign it. At Legit.Health, I served as the Person Responsible for Regulatory Compliance (PRRC) for a Class IIb medical device. I know what an auditor asks, and why “the AI did it” is not an acceptable answer. In this context, it is no longer enough to understand why the AI did what it did: the results must be audited, validated through clinical studies, and demonstrated to be in full compliance with regulations.

When you do not need to hire anyone

If you want an internal tool, a dashboard, or a prototype for your team, and a mistake won’t cost money or anyone’s health, then do it yourself. A good developer using Claude Code, or even someone with basic knowledge, can get it done in a matter of days. In that case, paying someone like me doesn’t make sense.

You need someone when any of the following applies:

What it costs

I have not found a reliable study of what an AI project costs in Spain or in Europe. The figures that circulate come from vendors quoting each other. So here are mine, which are public:

Why don't they drop, if Claude Code makes me faster? Because the price is fixed per phase. If it saves me weeks, you get more scope for the same price, not a smaller bill for half the work. I still write the model, the data pipeline and the evaluation myself.

If any of these applies to you, book a 30-minute call. On that call I will tell you whether you need me or whether Claude Code is enough.

AI writes the code. Someone has to sign it

Sources