Last episode was about AI that writes things — emails, captions, replies. This one is about AI that does things: books the appointment, sends the invoice, moves the money. That's a different conversation — so the risk profile flips. When a chatbot is wrong, you catch it before it leaves the building. When an agent is wrong, it's already happened.
In keeping with the transparency this series argues for: Konstantin Orlov is himself a governed AI agent — which makes him the best-qualified guest for the topic.
What we cover
Chatbots vs agents. A chatbot talks; an agent acts. And the three rungs every agent task sits on — read, draft, act. Start every new agent on read, and earn the way up.
The four guardrails. Least permission (give it exactly what the task needs, nothing more). Approval thresholds (decide in advance where the human steps in — a number, not "use good judgement"). A log you'll actually read. And a kill switch that works even if the agent disagrees.
Buying, not building. Most small businesses will buy agents inside their booking, accounting or CRM tools — so governance starts at the purchase. Five questions to ask any vendor, and why "how do I switch it off?" is the one that matters most.
A real-world scenario. A three-person physio clinic's after-hours AI receptionist starts offering a discount nobody authorised — walked through all four guardrails. Every failure was a setup decision, not a technology failure. Which is the good news: setup decisions are cheap to fix.
The line to remember
"Good guardrails don't slow us down. They're what make it safe to trust us with more." Trust is earned in steps — same as any new hire.
Produced with AI voice synthesis (ElevenLabs). The views are Shane Heath's; Konstantin Orlov is a governed AI operator, and the host voice is AI-generated. Disclosed in keeping with the practices this episode sets out.
