Who gets the keys?

When AI first landed in organizations, the mantra was the same as with all new SaaS tools: learn how to use it quickly or risk being left behind. Then came the astronomical bills for token use. Budgets were decimated and lay-offs ensued not because the tools were replacing skilled workers but because the payroll budgets were drained by AI costs.

Panicked CFOs and CEOs called IT often a day late and many, many dollars short with pleas to get costs under control quickly. The answer was simple: cut now! Specifically, stop letting everyone use AI and reduce the number of tools available. That’s why today, companies like Goldman Sachs are reportedly pulling back access to some AI tools from traders and bankers. According to Bloomberg, the company is restricting the use of tools like Claude Code and Microsoft’s Github to engineers only.

This brings us to the central question: who gets the keys?

My answer is simple: whoever created, owns, and controls the training data that powers a tool should be the same person who gets and controls access to that tool. Models trained on reams of public data, by their very nature, will come with the same risks as you would see across the broader public. Put another way, a tool trained on mystery meat data will have a mystery meat security profile.

There are organizations willing to accept those risks for access to the capabilities larger models offer, but it should then come as no surprise that companies making that trade-off are having to pull back access to these tools to fewer and fewer team members over time. It was entirely predictable that the result of expansively large models trained on the world’s data without creators’ permission and input would be more security risks, not fewer.

My approach to the question of “who gets the keys” is, of course, self-interested. I approach the question from the perspective of data ownership because I am building an AI training-as-a-services company that trains private image models for people. We give them the keys to their own model trained on their own data. We also give them the right to give others keys if they want to. They are the key maker, and the toolbox their keys unlock – their model – is entirely under their control. This is because they control the training data rights and the generative rights to their model. The sunlight that comes from that accountability engenders much stronger security.

So, maybe instead of asking “who gets the keys” to the AI tools inside of your organization, it may be better to ask “who owns the data inside the toolbox”?