AI agents: where autonomy breaks past five steps

September 4, 2026 · 5 min · Free, with full transcript

Every vendor now ships an "AI agent", so what is actually different from a chatbot? The answer here is mechanical: an agent wraps a language model in a think, act, observe loop and lets it call tools such as search, email, code execution or a database query. The position taken is that autonomy multiplies errors rather than averaging them out, and that 90 percent reliability across ten chained steps leaves about a 35 percent chance of finishing correctly. You walk away with a test for any pitch: count the steps, ask what happens when step four is wrong and who catches it. Examples include GitHub Copilot agent mode, where tests verify output, and Klarna's customer service assistant, plus agent platforms from Salesforce, Microsoft and Google.

Listen to the podcast

4 min

Chapters

Key takeaways

  • Ask any vendor what happens when step four is wrong and who finds out before it reaches a customer.
  • Limit agents to three to five steps in a domain with clear right answers where mistakes are cheap or reversible.
  • Treat chained reliability as multiplication: 0.9 across ten steps is roughly a 35 percent success rate.
  • Do not pay autonomous prices for work a human still reviews, which is an assistant rather than an agent.
  • Start with narrow, repetitive, high volume tasks that have a verification step, then extend the leash slowly.
Read the full transcript

Host:Welcome to the Leaders Insights Podcast. Today's episode, what an AI agent actually is and where the hype ends. Everyone's selling me an AI agent this year. My CRM has one, my email has one, my coffee machine practically has one. What actually is it?

Expert:An agent is software that takes a goal, breaks it into steps, and acts on those steps using tools without you approving each move. The word people abuse is acts. A chatbot answers,

Host:an agent does. That sounds like a distinction cooked up to justify a price increase. Sometimes

Expert:it is, but there's a real mechanical difference. A plain language model, the thing behind chat GPT, predicts text. You ask, it replies, done. An agent wraps that model in a loop. Think, act, observe the result, think again. That loop is the whole trick. Walk me through the

Host:loop slowly like I've never seen a computer. Say the goal is book me a flight to Berlin

Expert:under 400 euros. The model reasons, I need to search flights. It calls a tool, a flight search interface, meaning a connection to an actual booking system. It reads the results, nothing under 400. It reasons again, try nearby dates, searches again, finds one. Then

Host:it books. Each step feeds the next. So the intelligence is really just the model deciding

Expert:which button to press? Exactly right. And that's the part vendors hide. The model isn't magically capable. It's a decision maker that's been handed a toolbox. Search, send email, run code, query a database, easy day, and it picks tools in sequence. Take the tools away and your agent is a very expensive parrot. Where does this fall over? Because I've

Host:seen the demos and demos always work. Demos are the natural habitat of the agent that

Expert:never fails. Real failure comes from the loop compounding errors. Step one is 90% reliable. Sounds great. But chain 10 steps and you're at 0.9 to the 10th power. About 35% chance the whole thing works end to end. Autonomy multiplies mistakes instead of averaging

Host:them out. So the longer the task, the worse it gets. That's the opposite of what I'm being sold. It's the dirty secret of the field in 2026. Salesforce, Microsoft, Google.

Expert:They all launched agent platforms and their own reliability numbers. When you find them buried in the docks, drop sharply past a handful of steps. The good deployments keep the lease

Host:short. Define short leash. I want something I can actually use Monday. Three to five steps

Expert:in a domain with clear right answers where a mistake is cheap or reversible. Coding assistants like GitHub CoPilot's agent mode work because a bad suggestion gets caught by tests. The code either runs or it doesn't. Compare that to an agent emailing your customers unsupervised

Host:where a mistake is a screenshot on social media. So when someone pitches me a fully autonomous agent running my operations, the correct responses? Polite laughter. Ask

Expert:them one question. What happens when step four is wrong and who finds out? If the answer is a human reviews it. Congratulations. You don't have an autonomous agent. You have an assistant with a good marketing budget, which is fine. Assistance are useful. Just don't pay

Host:autonomous prices for supervised work. Is there a version of this that genuinely earns its keep

Expert:right now? Yes, and it's boring, which is why it works. Narrow, repetitive, high volume tasks with a verification step. Klarna talked publicly about their customer service assistant handling the volume of hundreds of agents. That's their figure. So weigh it. But it works because the domain is tight and answers are checkable. The unglamorous stuff is where agents actually pay.

Host:Give me the one thing to remember when the next vendor walks in. Count the steps and ask about

Expert:verification. If a task takes more than five autonomous steps and nothing checks the output before it reaches a customer or your bank account, it will fail more than they claim because the math of chained reliability isn't on their side. Start with tasks where being wrong is cheap. Prove it there, then extend the leash slowly. That single habit will save you more

Host:money than any platform they're selling. Count the steps. Check the work. I can do that. Thanks. That's it. The AI Decision Tools are live at MBA-training.com.

Read and go deeper