As AI agents take on more responsibility, organizations need to think carefully about access, permissions and accountability.
More organizations are giving AI agents administrator access to their websites, allowing them to make changes directly to live systems. Sometimes without much thought about what that access actually means.
We use AI extensively in our work, from software development and research to publishing and data management. We also build automated systems that take on increasingly complex tasks. But there’s a difference between putting AI to work and handing over control of the infrastructure.
We don’t think giving an AI system the keys to a website is a particularly good way to get there.
A lot of our work involves preparing information with AI, having someone review it, and then importing it into the appropriate database or publishing system. We use AI to help write code, but we test it before it goes into production. We build automated workflows, too, including processes that can run independently once they’re properly configured.
The amount of human involvement depends on the task. Some things are perfectly suited to automation. Others need someone to check the information, approve a change or make a decision. We don’t see a reason to treat every task the same way.
Permissions are part of that. If an AI tool needs to read information or prepare an update, it doesn’t necessarily need permission to install plugins, create administrator accounts or change the entire website. We prefer to give tools access to what they actually need.
We also want a record of what they’ve done. Changes should be traceable, backups should be available, and there should be a way to undo something when necessary. Organizations should know what tools are operating on their websites, what they can access and who is responsible when something goes wrong.
We’re not interested in keeping AI away from our infrastructure. Quite the opposite. We want to build better tools, automate repetitive work and give these systems more useful responsibilities.
But there’s a difference between building an AI system that can do something and giving it permission to do absolutely everything.
We use AI to get more done. We still take responsibility for what happens to the systems, information and projects we look after.