AI agents can automate a lot of manual work - reading, drafting, triaging, moving information between systems. That's the appeal, and it's real. But there's a quieter shift underneath it: the more we delegate the doing, the more our value lies in the deciding. Two principles keep coming up whenever I build or teach this, and both put the human firmly back in the loop.
The more we delegate the doing, the more our value lies in the deciding.
1. Don't outsource judgment to AI
AI can save us enormous time on manual work. But understanding and evaluating its output should still sit with us. The point of saving time is not to think less - it's to spend that time asking better questions.
When an agent hands you a result, that's the start of the work, not the end of it. The review process is what separates a professional who uses AI well from one who just blindly accepts its output:
- What assumptions is this output making?
- What gaps are there?
- Are there logic flaws or contradictions?
- What hasn't been considered?
- Does this actually make sense in the real context?
A lot has been said about AI eroding our ability to think critically. I'd push back on that - the choice is ours. The difference is in how you use it:
AI makes us passive. We accept whatever it produces and move on, spending the time saved on more consumption rather than better thinking.
AI gives us more space to question, evaluate and decide - because the mechanical work is no longer eating our day.
AI can even help here - stress-testing its own outputs, surfacing blind spots, arguing the other side. But those critiques are inputs to your thinking, not replacements for it. The final judgment still has to come from the human.
2. Design guardrails before capabilities
Guardrails for AI agents are too often an afterthought - bolted on only after something goes wrong. We should flip that order entirely:
Starting with "what shouldn't it do?" sounds restrictive, but it's the opposite. When you know the boundaries up front - which actions need human approval, what data the agent can touch, where it must stop and ask - you can let it move quickly and confidently inside them.
For AI agents, constraints aren't just safety features tacked on at the end. They're part of the design itself. An agent with well-chosen constraints is one you can actually trust to run on real work.
An agent without guardrails isn't more capable; it's just more likely to do the wrong thing at scale.
Judgment and guardrails are skills you build
Both of these are learnable, and you learn them fastest by building. When you wire up an agent yourself - deciding where to keep a human in the loop, what to forbid, how to test that it behaves - these principles stop being abstract advice and become muscle memory. You develop an instinct for when to trust an output and when to verify it, and for where a guardrail belongs before it's needed.
Learn to build agents with judgment and guardrails
If building AI agents with judgment and guardrails like these is something you'd rather do hands-on, that's exactly what we built our Applied AI Agents workshop for. In a single practical session you'll build working AI agents from scratch - no prior coding required - and leave knowing where the human belongs in the loop and how to design the constraints that make an agent trustworthy.
June slots are nearly full and July slots are now open. The takeaway either way is the same: don't just hand work to AI - learn to direct it, question it and bound it.