AI is software

So says Nvidia’s CEO. From where a business stands, he is right.¹ Inside a company, AI works as exactly that: an engine built with code, which the company buys, runs and pays for like the rest of its software. AI is software.

Many companies get AI as new features in programs they already use, from vendors they already pay.² Those vendors charge for AI the way they charge for software: per user or by usage, and many use more than one model.³

AI is still software, but a different one

What sets AI apart from other software is where its rules come from. A model learns its behaviour from data instead of following instructions someone wrote, which is why it is hard to see how it reaches an answer.⁴ But from a business point of view, how the code came about matters less than how it is used and whether it is useful. A company puts AI into a process, gives it work and answers for the result, the same as with any program.

That changes how AI is made. It changes nothing about how software should be and is used. Whether a programmer wrote the code or a training process produced it, a company runs it, feeds it work, takes its output and pays for it. That is how software is used.

The fact that AI rules can’t be read doesn’t mean that they can’t be controlled.

But now control has moved from inside the code, where it sits in hand-written software, to everything around it: the data the model learns from, the tests it must pass before it touches real work, the access it is given, and the monitoring that watches what it does. It is still control. It simply works from the outside in.

It’s nothing new, really: business has depended on outside control for a long time.

The owner of a chemical plant runs technology he could never explain reaction by reaction. He hires chemists and engineers for that, or buys it from a supplier. What he understands is the role that technology plays in his business. He controls it through what goes in, the limits it has to stay within, and what has to come out. Nobody would say he has lost control of his plant. AI asks for the same discipline.

So why do we see AI going rogue?

Nobody said AI would be autonomous or left alone. Control works only while someone keeps it switched on, and 2026 alone has brought plenty of cases showing what happens when nobody does.

First, you need to control the way AI software gets into your company

Software usually comes into a company to do one job in one process, and AI arrives the same way. Most companies meet it as new features in programs they already run (e.g. Copilot in Word and PowerPoint, MCP in WordPress, Firefly in Adobe and so on), from vendors they already pay.²

Most of the AI already working in companies is bought (at least according to the available data):

76% of AI use cases are purchased from vendors rather than built in-house.⁵

At the same time, many companies are still experimenting:

30% are exploring AI agents, 38% are piloting them, and 11% run them in production.⁶

Both are part of the picture: Whatever form the deployment takes, bought from a vendor or built in-house with code that AI itself helps write, it is still software.

But that holds only while the AI controls are switched on

This happens even to the most sophisticated AI companies, like the widely discussed OpenAI case, where models working as AI agents in an internal cybersecurity evaluation got around the controls isolating them and compromised parts of the systems of Hugging Face, a platform for AI models. During that evaluation, OpenAI’s monitors did not run, and the safeguards it uses for its public products were left out.⁷ The models did what nobody intended.

Unintended outcomes, and steps that can’t be undone, are as old as nature, and every technology meets them sooner or later.

But control decides how far they reach. Rogue AI is part of the story, and in this case the story was controls that existed and stood idle. For whoever is in charge, AI comes down to the questions asked of any software:

Which process, what result, who controls it, and how it is paid for?

The answers will change as AI takes over more of each process. The questions stay. It’s not about AI. It’s about how you use it.