When you bring on a new hire, it takes weeks, or sometimes months, before they get approval authority. They may have spending limits, or route calls through a manager. Leaders may take time to see how they handle the small stuff before trusting them with larger or more complex decisions. An AI agent should be treated no differently.
But even with a slow rollout, the question remains: how much authority should an AI agent have. And how much should you let it decide on your behalf?
We all know – and expect – that humans will make mistakes. You might even make a few yourself. Most are easily fixed and everyone moves on. Occasionally something really bad happens, we deal with it, and people are held accountable. But most people give others the grace we hope to have given to us.
AI agents aren’t like that.
Without proper oversight – something you’d likely always have with a human in one form or another – an AI agent could make the type of mistake carries a large blast radius, something not so easily undone or harmless.
For instance, if an agent messes up the formatting of a slide deck, the flagging of duplicate records, or pulls the wrong data for a review, it doesn’t carry a large consequence. You shrug and redo them (though, you’d likely be annoyed).
Then, there’s the other instance, the one where records are deleted or permissions are allowed where they shouldn’t be. These high-blast radius mistakes cost money, reputation, and, potentially, your job.
These are the areas where you have to draw a line. Think about how long it would take to notice a problem. And how hard would it be to undo? If you think through the question of “how much will this cost if the decision is wrong?” you’ll know the answer to “how much should I let the AI agent decide?”
At this point, it should be obvious that you can’t let AI agents run unimpeded. You’ve got to have guardrails in place for critically important decisions. But even then, the pace of work will increase. That’s the whole point, right?
So what are you going to do when you become the bottleneck? When decisions are being made by the agent faster than you can keep up?
If you start skimming or approving things without full review, you’ve just abandoned any safeguards that are in place. Remember: a real person is responsible for an agent’s decisions. If you’re that person, you have to treat each approval like it’s your own.
It’s worth thinking about the ways you’ll handle a capacity problem. You could spread authority across the team. You could set a higher standard for the work that requires your sign-off, especially as agents get more and more right decisions under their belt. In short, make sure the decisions that need your attention are worthy of it. As it earns more trust – like a human – trust it more. But don’t ignore it.
You don’t have to figure this all out on day one. As with any quickly developing technology with growing adoption rates, most are trial-and-erroring their way through to some extent. You have to decide what works best for you, try it, and adjust. Even your best guest can go wrong. But with the right evidence and patience, you will get it right.
Just like that new hire, watch the AI agent closely. Work closely with it and see where it goes off track. Which categories of decisions does it do consistently well, and where does it struggle? Where did you get an odd result? Where were you pleasantly surprised?
Build the track record you need to put trust in the agent. Once you’ve seen the evidence that it can handle things in a way you’re comfortable with, widen the lane. Let the agent decide a little more. Pull back where you need to. Think of it as a calibration more than a right/wrong scenario. Tune the agent to your liking.
But be conservative in the beginning. If the stakes are high, keep it for yourself (or your team). Yes, things are moving fast. And you might be feeling the need to keep up. But not at the expense of the company or your career. You’ve got time.
TL;DR: The most workable approach to agentic decision-making is a slow one. Consider the cost of it making the wrong decision, and work backwards. Start with small, easy-to-fix decisions, then grow from there.
Wouldn’t it be safer to just keep human approvals as the standard?
If you’re reading this post, you know the answer to this. Sure, it’s safer. But is doing things the same as you’ve always done them the point?
How do I know when to give the agent more freedom?
Give it time. As it continues to get decisions right, that evidence you can open it up a bit more. But go slowly. No need to rush.
What if we get the boundaries wrong?
You probably will. And that’s okay. Keep calibrating, keep testing, keep watching. Rome wasn’t built in a day.
