Digital Cobra Farms: How to <<NOT>> Fail at AI!
Disclaimer: Based on a true story, articulated with the help of Kimbho-AI
After 8 long years, I had an opportunity to meet an amazing friend of mine from college days. We were friends through thick and thin in college, good times, bad times, and even worse grades. No matter what, we always found a way to enjoy ourselves.
Today, after decades of climbing the corporate mountain (and only occasionally getting lost in the break room), he’s finally steering the ship as the Grand Poobah of Technology. Now an executive vice president at one of those skyscraper-sized, glass-and-steel money palaces, he’s sharper than a COBOL implementation and more ambitious than a zero-touch onboarding, and these days, he’s surfing the AI tsunami like it’s his personal wave pool.
Over a plate of Japanese omakase sushi (which I am still learning to pronounce) that cost more than my first car, he was practically vibrating with excitement. “Sid, we’re finally cracking it,” he said, swirling the Chianti in his glass. “I’ve got the board completely bought in. I’m heading up the new ‘AI Agent Velocity’ initiative. Our North Star metric for the next six months is to deploy 25 new autonomous agents across the business.”
I raised an eyebrow. “Twenty-five? That’s… aggressive. What kind of agents?”
He leaned forward, his voice dropping to a conspiratorial whisper. “The really sexy stuff. We’re spinning up a team to build a COBOL Modernization Agent. It’s supposed to automatically translate our ancient mainframe code into Python. Can you imagine? We’ll finally kill the dinosaur!”
He took a triumphant sip of wine. “And that’s not all. The retail division is prototyping a ‘Zero-Touch’ Onboarding Service. A customer can go from application to a fully funded account without ever interacting with a person. Pure automation. We’re counting the agents as we deploy them. It’s a revolution.”
I took a slow bite of my food, trying to choose my words carefully. “Buddy,” I said gently. “Can I tell you a story? It’s a classic that Pandey sir used to narrate almost every day, but it feels more relevant now than ever. It’s about cobras in Delhi.”
He looked at me, a little confused, but nodded. (He didnt know he is going to now be subjected to my amazing storytelling, which puts my kids to bed like instantly.)
I started:
“So, back in the day, the British government in Delhi had a cobra problem. They wanted to get rid of them. Their goal was simple: a safer city with fewer live snakes. So they came up with a measure: the number of dead cobras. Seemed logical. To make it a target, they offered a cash bounty for every cobra skin people brought in.” “Smart,” my friends said. “Incentivize the outcome.” “Exactly,” I replied. “And at first, it worked. But the officials were paying for dead snakes, not for a safer city. Enterprising locals soon realized that hunting dangerous wild cobras was a pain. It was far easier to set up cobra farms, breed them in their backyards, kill them, and cash in. The number of dead cobras being turned in skyrocketed. The government officials celebrated hitting their targets, even as the underlying problem got worse. They were just funding a thriving cobra-breeding industry.”
I let that sink in. “Of course, when the government scrapped the bounty, the farmers released their now-worthless snakes. Delhi ended up with a bigger cobra problem than ever before. This is the essence of Goodhart's Law: When a measure becomes a target, it ceases to be a good measure.”
I leaned forward, mirroring his earlier posture. “You, my friend, are not building an institution of the future. You are building digital cobra farms.”
My dear friend stared at me, his fork frozen mid-air. I continued. “Let’s look at your COBOL Modernization Agent. The goal isn't to ‘translate COBOL.’ The goal is to have a robust, secure, and agile core banking system. Your team, tasked with just creating an ‘agent,’ will build a fragile translation layer. It will look great on a PowerPoint. You’ll check the box. But you won’t have a modern system; you’ll have the old, creaking mainframe with a fancy, unauditable, and buggy Python wrapper on top. You’re breeding a cobra that will bite you during a system failure or a regulatory audit.”
“And the ‘Zero-Touch’ Service? Your real goal isn't to have no intervention; it's to provide fast, frictionless, and secure service to happy customers. By making ‘zero touch’ the target, your team will create a chatbot nightmare that can’t handle a single edge case (which you cant always predict). Imagine A high-net-worth client trying to wire a down payment for a house gets stuck in a loop, gets furious, and takes their eight-figure account to your competitor. You’ll hit your ‘zero touch’ target for 99% of low-value accounts, but the fallout from the 1% can be catastrophic.”
My buddy put his fork down and stared into his wine. The excitement had faded from his eyes, giving way to a growing sense of horror. “So what’s the alternative?” he asked quietly. “How do we measure this?”
“You stop counting the dead snakes,” I said. "You stop counting the dead snakes and start measuring if the city is actually safer. You stop chasing sexy demos, and you attack the real, ugly, and profitable problems.”
I held up a hand and started ticking them off on my fingers.
“One: You don't want a ‘COBOL Modernization Agent.’ You want stability and efficiency. So, you target a 20% reduction in system-failure-related Priority 1 incidents. You build an AI that does predictive maintenance, analyzing system logs to flag a potential failure before it takes down your wire transfer system on the last day of the quarter.”
“Two: Instead of a ‘zero-touch service,’ you want faster, more secure service. So, you target a 75% reduction in false positives from your fraud detection system. Every time you block a legitimate transaction, you create a furious customer. An AI that can tell the difference between a dad buying a video game for his kid in another state and a genuine threat is worth a hundred flashy chatbots.”
“Three: You have an army of analysts drowning in paperwork for anti-money laundering checks. Forget a flashy ‘AML Bot.’ You target a 60% reduction in the manual review time for low-risk AML alerts. You build an AI co-pilot that pre-processes the data, summarizes the case, and presents a recommendation, letting your human experts focus on the truly suspicious cases.” He was leaning in now, the wine forgotten.
“Four: Your commercial lending division is slow. A small business waits three weeks for a loan decision while your nimbler competitors do it in three days. You don’t need an ‘Underwriting Agent.’ You need to reduce the time-to-decision for small business loans to under 48 hours. Use AI to ingest financials, run credit models, and automate the initial risk assessment. Your target is market share, not a bot count.”
“Five: Trade settlement failures cost you millions in penalties and counterparty risk. This is as un-sexy as it gets, but it’s pure gold. You target a 40% decrease in the rate of trade settlement failures. You build an AI that validates trade details before they are sent, catching errors that would normally take days and a team of ten people to unravel.”
“And six, my personal favorite,” I said, leaning back with a final flourish. “Your call centers are a treasure trove of data you ignore. You target the automatic identification of the top 5 drivers of customer complaints, delivered in a report every Monday morning. An AI that listens to and analyzes every customer call can tell you that a new feature in your banking app is broken before Twitter does. That’s not a cobra farm; that’s an early-warning radar.”
I took a final sip of my water, and he asked me, "Dude, so you still do not drink?" , I ignored and continued :)
“Don't build a theme park. Your job is to build a fortress that makes money. Stop counting your shiny new toys and start counting the bricks you’re adding to the wall and the gold you’re putting in the vault.”
He was silent for a long time, just looking at the tablecloth. Finally, he picked up his phone and typed a brief message. “Just cancelled my 8 AM meeting,” he said, looking up at me with a new kind of clarity. “The one where we were going to finalize the ‘Agent Velocity’ dashboard. I think we need a new North Star. Let's talk tomorrow.”