AI Agent The Real cost
Tech Samjho
0:00 / 0:00
AI Agent The Real cost
134 просмотра · 9 дн. назад
Tech Samjho
117 подписчиков
134 просмотра · 9 дн. назад
Designing Agentic Systems | Episode 12 (series finale)
You visit a doctor with a complicated health problem. He orders a blood test, and you come back. He orders an MRI, and you come back. He wants a specialist’s opinion, and you come back again. Five consultations, five bills, one problem. An AI agent works the same way: it calls the model, checks a system, calls the model again, checks another system. Every call has a cost, and unlike a fixed consultation fee, that cost depends on how many tokens the model reads and writes. This video follows one failed UPI payment of ₹50,000 through an AI agent, then asks what happens on salary day, when ten thousand people complain at once.
In this video:
The story: Ravi’s ₹50,000 UPI payment leaves his account but never reaches his landlord, and an AI agent investigates using four tools
Why you pay for every re-telling: the model doesn’t remember the last step, so it re-reads the whole story at every step, and you pay per token
The cost formula: steps × what’s re-told × customers
What 10,000 complaints really do: 50,000 model calls, with tokens, compute and the bill all multiplying, plus rate limits that make genuine cases wait
The fix you’ve already heard on the phone: “if you’re calling about failed UPI payments this morning, we’re aware”. One incident check in plain code turns 50,000 model calls into 5
Don’t put customers on hold: run independent checks side by side, and work behind a reference number
Where the model runs: for an Indian bank, data comes before price. Then it’s a cab (pay per token) or a car (run your own model)
The line to question in your next meeting: “the pilot cost us almost nothing, let’s roll it out to every customer”
Key takeaway: count the re-tellings, find the shared cause, then scale. A pilot on a quiet Tuesday tells you nothing about salary day.
What’s next: this was the final episode of Designing Agentic Systems. The next series, Enterprise Architecture in the AI Era, widens the lens from one agent to the whole enterprise around it. Subscribe so you don’t miss it.