Every line of AI code still costs $1

Generating a line of code is becoming almost free. I think that is true, and I also think it is the least interesting fact about building software right now.

My thesis for this week is that every line of code in an enterprise AI application still costs you about a dollar. The token price fell. The cost of an enterprise-ready application barely moved, because almost none of that cost was ever in typing the line.

Working code and a solved problem are different things

Vibe coding did not create this gap, but it has made it much more visible. You can now get to the 80% point extremely quickly, and most people understand that part.

What working code proves is that something can happen. It does not tell you the problem has been solved safely, reliably and repeatably. Almost all of that difference is invisible in your first iteration, and the work to close it can be a big multiple of the original proof of concept.

When you start looking at real enterprise usage, you have to understand all the edge cases. What happens when someone does not save something properly. What happens when a system drops, or access to a gateway fails. None of that shows up in the demo.

What sits outside the code

Security and identity. Permissions. Auditability. Those are the obvious ones, and they are already more work than people budget for.

Then there is the architecture. Asynchronous processes where two things are happening at once and you have to deal with both. Dealing with an AI through a gateway, which is asynchronous by nature, so you are always waiting on a response. Deployment and recovery. What your application does when it loses connectivity to the AI entirely. These are genuinely complex problems, and in the end they still require a high quality understanding of how to architect software properly.

I read an article recently where people were declaring that databases are no longer needed, because everything can go in flat files. Statements like that are useful, in a way. They show precisely which understanding is missing, and that understanding is exactly where the cost has moved to.

Lines of code was always a bad metric, and it still tells you something

Engineers have argued for decades that lines of code is a poor measure, and they are right. AI can generate enormous amounts of rubbish, so volume is certainly not the outcome.

The reason I still use it is that meaningful outcomes need a meaningful amount of code, no matter how it gets written. We have millions of lines in 1KE. What I am really pointing at with the dollar is that the cost of a line has to encompass everything, not just the tokens.

At the moment the focus is almost entirely on token cost. That does not recognise the cost of understanding how the workflows should operate. Or the cost of understanding the real business operation you are building an analogue for. Or the effort around security and resiliency in the application code. The effort of writing the line is the tip of the iceberg. That line had to have a lot of thought, a lot of process and a lot of deep iterations behind it to be right.

We learned that the hard way. You can prototype quickly and get to a sense of achievement quickly, and AI is brilliant for proving something is possible. Taking that prototype to a production outcome is a different exercise, and it took us more than a year to properly understand the size of the difference.

Could it fall to cents?

I get asked whether the dollar eventually becomes a few cents. My sense is no.

Generation of the line will become almost free, and that is not the issue. The metric should be the human cost: understanding the problem, managing your data, getting a handle on your workflows, building the architectural resilience so applications can recover into different regions. All of that still needs thought and human intervention.

The core of your application, the problem domain, is never going to reside within the AI. It is always going to be owned by us. There is a real cost attached to that, and you cannot mitigate it with cheaper tokens.

What I would take away

AI has made the code cheap, line by line. It has not taken away the complexity of understanding the problem, or of knowing whether what you built is the right thing.

A chief executive's job is not to do the work, it is to make the right decisions about what work needs to be done. If you are using AI to build enterprise code, that is fine, and I would encourage it. You still have to make sure you are building the right code, and building it the right way.

Watch the full conversation with Sidney on YouTube: https://youtu.be/yBbIfYbhwD4

Subscribe to The AI Take for one AI lesson a week, and what it changes for your business.

About 1KE:

1KE is the presales autopilot for technical sales teams. It turns discovery calls, emails, and technical documents into structured requirements and customer-ready solutions in minutes, so teams can respond while the deal is still warm.

Learn more → https://1ke.ai

Michael → https://linkedin.com/in/michaelchanter/ · https://x.com/1keceo

Sidney → https://linkedin.com/in/sidneyminassian/

Follow 1KE → https://linkedin.com/company/1ke-ai/

‍