Let me share some numbers that should make you uncomfortable.
A client came to us last quarter processing 1 million customer service interactions per month. They'd deployed an agentic AI solution — state of the art, impressive demos, leadership buy-in.
Their monthly LLM inference cost: $47,000.
Their previous solution (rules + basic ML): $3,200/month.
The agents were more capable. They were also 15x more expensive. And costs were growing 20% month over month as usage expanded.
The CFO had one question: "What exactly are we getting for that $44,000 difference?"
Nobody had a good answer.
The Agentic AI Gold Rush
We're in the middle of an agentic AI gold rush. Every enterprise wants agents. Every vendor is selling them. Every conference is talking about them.
The pitch is seductive: Deploy intelligent agents that reason, decide, and act. Automate complex workflows. Replace expensive human judgment with cheap AI judgment.
Except it's not cheap. And most companies are discovering this the hard way.
The Math Nobody Wants to Do
Let's do the math that most POCs skip.
Basic LLM economics:
- Input tokens: ~$0.01-0.03 per 1K tokens
- Output tokens: ~$0.03-0.06 per 1K tokens
- Typical agent interaction: 2,000-5,000 tokens total
- Cost per interaction: $0.02-0.10
Sounds cheap, right? Let's scale it.
At 1 million interactions/month:
- Low end: $20,000/month
- High end: $100,000/month
But agents don't make one call. They reason, call tools, check results, retry. A single "agent decision" might be 5-10 LLM calls.
Realistic agent economics:
- Calls per decision: 5-10
- Cost per decision: $0.10-0.50
- At 1M decisions/month: $100,000-500,000/month
Now factor in:
- Retry logic when agents fail
- Fallback to more expensive models for hard cases
- Growing usage as you expand deployment
- Price increases from providers
One year cost for a single agent use case: $1.2-6 million.
Is your use case worth that?
The Uncomfortable Question
Here's what nobody wants to ask: Does this decision actually need an agent?
I've reviewed agent deployments across banking, insurance, healthcare, and telco. A consistent pattern emerges:
- ~70% of decisions could be handled by rules
- ~20% of decisions could be handled by traditional ML
- ~10% of decisions genuinely benefit from agent reasoning
But most deployments route 100% of decisions through agents.
That's not a strategy. That's waste.
The Trust Cascade: Right-Sizing Intelligence
The solution isn't to abandon agentic AI. It's to use it where it matters.
We call this the Trust Cascade — routing each decision to the cheapest sufficient intelligence.
| Layer | Handles | Cost | Volume |
|---|---|---|---|
| Rules | Known patterns, simple decisions | $0.0001 | ~70% |
| ML | Moderate complexity, pattern matching | $0.001 | ~20% |
| Single Agent | Complex reasoning | $0.01 | ~7% |
| Multi-Agent | High-stakes, ambiguous | $0.03-0.05 | ~3% |
The math changes dramatically:
Instead of: 1M × $0.10 = $100,000/month
You get:
- 700K × $0.0001 = $70
- 200K × $0.001 = $200
- 70K × $0.01 = $700
- 30K × $0.04 = $1,200
Total: ~$2,200/month
Same outcomes. 98% lower cost.
"But Agents Are Getting Cheaper!"
I hear this objection constantly. "Token costs are dropping. Just wait."
It's a trap. Here's why:
Token costs drop. Token usage explodes.
As costs drop, you deploy more agents. Each agent gets more capable (more tokens). You expand to more use cases. Total spend keeps growing.
This is Jevons paradox applied to AI. Efficiency improvements increase total consumption.
The companies that will win aren't waiting for cheaper tokens. They're building architectures that use tokens intelligently — regardless of price.
The CFO Conversation You Need to Have
If you're deploying agentic AI, your CFO will eventually ask hard questions. Here's how to prepare:
"What's our cost per decision?"
You need to know this number. Not approximately — precisely. With breakdown by decision type, complexity, and model used.
"What's driving cost growth?"
Usage growth? More complex decisions? Retry rates? Model changes? You need attribution.
"What's the ROI?"
What value does each decision generate? What's the break-even? Can you prove it?
"What's our cost at 10x scale?"
Does cost scale linearly? Sub-linearly? Super-linearly? What's the architecture?
If you can't answer these questions, you don't have a strategy. You have a science experiment with a credit card attached.
Sustainable Agentic AI
Sustainable agentic AI isn't about avoiding AI. It's about deploying it intelligently.
Principles:
- Route by value. High-value decisions justify agent reasoning. Low-value decisions don't. Route accordingly.
- Measure everything. Cost per decision, by type, by outcome. You can't optimize what you don't measure.
- Build the cascade. Rules for simple. ML for moderate. Agents for complex. The architecture matters more than the model.
- Self-improve. When agents solve problems, extract the pattern. Push it down to cheaper layers. Costs should decrease over time, not increase.
- Set budgets. Per agent, per use case, per month. Alerts when approaching limits. No blank checks.
The Strategic Question
The strategic question isn't "How do we deploy more agents?"
It's "How do we deploy AI sustainably at enterprise scale?"
The companies that answer this question will build durable AI capabilities. The companies that don't will eventually face a reckoning — when the CFO finally asks why AI costs grew 10x while everything else stayed flat.
Agentic AI is powerful. But power without economics is just expense.
Build the architecture for sustainable AI. Or explain to finance why you didn't.