For the past few years, much of the conversation about AI adoption has focused on getting people comfortable using it. We’ve encouraged employees to experiment, learn how to prompt, understand appropriate use and find practical ways AI can make their work easier.
That work still matters, but the technology is moving quickly and the relationship between people and AI is starting to change.
Until now, most workplace AI use has involved asking for something and receiving a response. Draft this email, summarise this document, analyse these comments or help me prepare for this meeting. The employee remains actively involved in the task and, importantly, usually sees the output before anything happens with it.
AI agents take us into different territory because we’re beginning to move from asking AI to help with a task to giving it responsibility for completing parts of the task.
An agent might be asked to research an issue across multiple sources, prepare an analysis, update information in another system and create the actions that follow. As these capabilities become embedded into workplace technology, people will increasingly be able to hand over pieces of work and move on to something else.
There is enormous potential in that. Done well, it could remove a significant amount of low-value administration and give people more capacity to focus on the work that genuinely needs their experience, relationships and judgement.
But it also changes the adoption challenge.
We’ve spent a lot of time teaching people how to use AI. We now need to help them understand how to delegate to it.
Think about what happens when someone joins your team. You don’t simply give them access to every system, hand them a vaguely worded task and assume everything will work itself out. You explain what you need, provide context, set boundaries, agree what they can decide themselves and make sure they know when they need to come back to you.
Over time, as their capability grows and trust develops, you might give them greater autonomy.
Working effectively with AI agents will require some of the same thinking.
People will need to understand what work is appropriate to hand over, what information the agent should have access to, what a good outcome looks like and how much checking is required. They also need to recognise when something requires human judgement and when the consequences of getting it wrong are too significant to simply delegate and move on.
This isn’t about creating fear around AI or putting so many controls around it that nobody wants to use it. It’s about building the judgement needed to use increasingly capable technology well.
That distinction matters because we have already seen what happens when organisations focus too heavily on AI usage as a measure of adoption. Encouraging people to use a tool more frequently doesn’t necessarily mean they understand how to use it effectively. The same will be true of agents, except the consequences become more significant when the technology can take action rather than simply provide an answer.
An employee asking AI to rewrite an internal email is one thing. An employee asking an agent to complete a process involving organisational data, customers, financial information or decisions affecting other people is something very different.
The question can’t simply be, “Can the technology do it?”
We also need to ask whether it should, what oversight is appropriate and who remains accountable for the outcome.
This is where organisations need to resist the temptation to treat agentic AI purely as another technology implementation. Giving people access to increasingly autonomous tools without building their understanding around them is unlikely to deliver the productivity gains organisations are hoping for.
Policies and governance will be important, but they won’t be enough on their own. People need practical experience. They need examples that relate to their roles, opportunities to experiment safely and clear guidance around where the boundaries sit. Managers will also need to be part of the conversation because expectations around delegation, checking and accountability will increasingly become part of how teams operate.
There will inevitably be mistakes along the way, and that isn’t necessarily a reason to slow everything down. It is a reason to create environments where people can learn before the stakes become too high.
The organisations that do this well won’t just have employees who know how to use AI. They’ll have employees who understand how to work with it, question it, direct it and recognise when their own judgement needs to take precedence.
Of course, the technology will continue to evolve and today’s agents will look relatively basic compared with what comes next. That makes building human capability now even more important, because the skills we’re developing aren’t really about learning one particular AI tool.
They’re about learning how to work effectively when some of the work can be delegated to technology.
We spent the first phase of workplace AI teaching people how to prompt. The next phase needs to teach them how to delegate.
And good delegation has never meant simply handing something over and hoping for the best.