The CXO’s Guide to Strategic AI Agent Deployment

By Yan Nuriyev, CTO, Flushing Bank

We’re being sold a compelling vision: AI agents as digital employees who plan, reason, and execute autonomous partners who’ll finally free us from the tyranny of routine work. Deploy them everywhere, we’re told, and watch your organization transform. And technically, they’re not wrong. By combining perception, planning, and tool use, agents can achieve complex goals with minimal human intervention.

AI agents can do almost anything, which is precisely why you shouldn’t let them. The constraint isn’t capability, it’s economics. When your AI agent takes 5 minutes for what code does in milliseconds, you’re not innovating, you’re burning capital. The future belongs to those who know the difference.

Efficiency

The physics of AI agents reveals an uncomfortable truth: intelligence is expensive. A 10-word LLM response consumes 0.3 watt-hours of energy or roughly 1,000 joules, approximately the energy required to keep a standard LED lightbulb lit for nearly two minutes.

(How Much Energy Do LLMs Consume? Unveiling the Power Behind AI: https://adasci.org/how-much-energy-do-llms-consume-unveiling-the-power-behind-ai/)

Traditional code is literally billions of times more efficient for logical operations. We’ve spent 70 years optimizing CPUs for deterministic math. Now we’re using probabilistic neural networks to approximate what a pocket calculator does instantly.

AI agents are cognitive luxury goods. Deploy them surgically, for perception, creativity, and genuine problem-solving.

Non-Deterministic Logic

Another problem is that AI reasoning is non-deterministic, the output can vary slightly each time due to randomness in the model. Fine for brainstorming ideas, but in regulated, repeatable business operations — it’s a governance nightmare.

The reliability problem gets worse at scale. Simple mathematics reveals why AI workflows are inherently fragile: even with 95% reliability per step, a five-step process succeeds only 77% of the time. Extend that to ten steps, and you’re down to 60%.

(Salesforce study finds LLM Agents Fail 65% of CX Tasks: what this means for AI in Customer Support: https://www.usefini.com/blog/why-salesforce-s-ai-fails-65-of-cx-tasks-and-why-b2c-cx-leaders-are-re-thinking-ai-support)

Errors compound, and at enterprise scale, compounded error becomes systemic risk.

AI Code Generation

When executives say “we need an AI agent,” they’re maybe confessing something else: “we don’t understand our own process.” The process is undefined, the data chaotic, the logic undocumented. So, executives punt to AI, hoping it will untangle what they never bothered to understand.

Here’s the counterintuitive solution: use AI to write code, not make decisions. First, map your process properly, inputs, outputs, decision trees. Then let AI generate the procedural code to execute it. This flips the entire equation. Instead of paying for intelligence for every transaction, you pay once for development.

The resulting code runs at zero marginal cost, executes in milliseconds, and slots perfectly into existing governance frameworks. Your compliance team already knows how to audit code. Your security team already scans for vulnerabilities. The version control is your audit trail. You get AI’s development speed without sacrificing operational predictability.

	Live AI Agent	Agent-as-Coder
Primary Cost Driver	Continuous Inference (OpEx)	One-time Generation (CapEx)
Energy Usage	High (Joules per transaction)	Low (Joules per build; negligible per run)
Latency	Seconds to Minutes (Variable)	Milliseconds (Constant)
Determinism	Probabilistic (High Variance)	Deterministic (Zero Variance)
Governance	Complex (Guardrails needed)	Standard (SDLC / Code Review)

I’ve seen this play out firsthand at my institution. Our credit officers began using Claude to help with annual business reviews, a process that requires analyzing income statements going back multiple years. While Claude could handle the task, each analysis took five minutes and the results varied in quality and format.

So, we tried a different approach: we used Claude Code to write a traditional Python program that would perform the same analysis. After a few days of iteration, we had robust, reliable code that completed each review in under a second.

This experience revealed something remarkable: we built custom software in three days for a problem we’ve had for years. Multiply this across every department, every workflow, every “we just deal with it” process. The long tail of enterprise software needs, previously uneconomical to address, has just become a greenfield opportunity.

When Intelligence Is Worth Every Joule

There are vast swathes of economic activity that are defined not by logic, but by entropy, chaos, nuance, and the messy reality of human interaction. We are witnessing the democratization of “high-touch” services, sectors where value was previously capped by the scarcity of human attention. By deploying live agents, we can unbundle complex, high-value services from the constraints of human labor hours.

For example, earlier this month, I attended a presentation by Bridget Mary McCormack, CEO of the American Arbitration Association. Working with McKinsey’s QuantumBlack team, they’ve launched an AI arbitrator that transforms how disputes are resolved. The system reviews filings and supporting documents, breaks down claims into their component arguments, and generates draft awards based on decades of case precedent. Traditional arbitration is notoriously slow, tedious, and opaque; parties often wait months for decisions they struggle to understand. Their new AI Agent accelerates dispute resolution from months to days while providing transparent reasoning for every decision, giving arbitrators greater confidence in outcomes and parties clearer insight into how conclusions were reached.

The Hybrid Future

We’re entering an economic phase transition where the winners will be those who master the arbitrage of intelligence, knowing exactly when to pay the energy cost for a probabilistic mind, and when to use that mind to build a deterministic machine.

AI agents are cognitive luxury goods. Deploy them surgically, for perception, creativity, and genuine problem-solving. For most tasks, though, traditional code (including AI-generated) can solve problems far more efficiently. Most business decisions don’t need bottom-up reasoning; they just require clarity and rules.

AI agents are powerful, but not magical.

Use them wisely, and they’ll transform your productivity. Use them everywhere, and they’ll just transform your cloud bill.

---