The Economics Question
After four pillars of work, NorthRidge had accomplished what few organizations manage: prioritized AI opportunities, validated data readiness, established governance frameworks, and designed experiences that users actually wanted.
But Sarah Williams, the CFO, had a question that stopped the room.
"We've built a compelling vision. Now tell me: what does it cost to run, and how do we know it's working financially?"
This wasn't skepticism—it was fiduciary responsibility. Sarah had seen too many technology initiatives launch with enthusiasm and wither without economic accountability.
"I was here for the cloud migration that went 340% over budget," she reminded them. "Nobody modeled the true operational costs. I'm not making that mistake again."
The FinOps Session
Michael convened a working session with Sarah, her finance lead Kevin Park, and the operational owners of each agent—Lisa, David, and Jennifer.
He opened with a principle that resonated with the finance team.
"An agent without financial controls is a liability, not an asset. Today we're going to define four things for each agent: What's its economic purpose? What will it cost to operate? What controls prevent runaway spending? And how do we see the economics in real-time?"
Sarah appreciated the framework. "That's the right order. Start with why we're spending, then worry about how much."
Defining Economic Purpose
For each agent, the team had to complete a simple sentence: "This agent exists to [reduce cost X / generate revenue Y] by [specific mechanism]."
Michael explained the distinction.
"Some agents save money—they reduce labor costs, cut errors, or accelerate cycles. We call these bottom-line agents. Others generate revenue—they enable faster delivery, improve conversion, or unlock new services. Those are top-line agents. The economics are completely different."
Jennifer spoke first about the Pre-QA Validation Agent.
"This is clearly bottom-line. My team spends 480 hours a month on mechanical validation checks. If the agent handles 75% of that, we're saving real money."
Kevin did quick math. "At $65 blended rate, that's about $23,000 a month in labor if you hit that 75%."
David framed the Field Note Normalization Agent differently.
"This is also bottom-line, but the savings are harder to see. My surveyors waste time cleaning up notes. If they save an hour a day, that's capacity we get back—but we're not cutting headcount."
"Capacity is still value," Michael noted. "You either absorb more work without hiring, or you improve quality with the same people. Both are measurable."
Lisa raised the Exception Routing Agent.
"This one's tricky. The volume is low, but when we miss a high-risk case, the rework cost is enormous. One bad delivery last year cost us $180,000 in remediation plus the client relationship damage."
Sarah understood immediately. "That's risk avoidance. The economics are about what we don't lose, not what we save."
Understanding AI Operating Costs
Michael pulled up a slide that made Kevin lean forward.
"Traditional software has fixed infrastructure costs. AI agents have variable costs driven by usage. Every time the agent thinks, it costs money."
He introduced the concept of tokens—roughly 4 characters or 0.75 words—and explained how costs accumulate differently for AI than for traditional systems.
Sarah asked the critical question: "So if we're successful—if people actually use these agents—our costs go up?"
"Yes. That's exactly why we need forecasting and controls. Success without financial visibility is dangerous."
Kevin had seen this pattern before. "It's like cloud computing. Pay-per-use sounds great until you get the bill."
Michael showed them the cost drivers specific to AI agents:
- Input tokens: The context and prompt sent to the model
- Output tokens: The response generated (typically 2-5x more expensive)
- Calls per task: Unlike simple chatbots, agents reason through problems—each "thought" is a separate cost
"A single user question like 'validate this report' might trigger 15-25 internal model calls as the agent works through the validation steps. Track calls-per-task, not just tokens-per-response."
Building the Forecast
For each agent, the team built a 12-month financial forecast.
Michael walked through the Pre-QA Validation Agent as an example:
"Based on your volume—150 reports per day, average 2,500 tokens input, 800 tokens output—we can model the monthly cost."
Kevin ran the numbers as Michael talked:
- Daily token usage: ~495,000 tokens
- Monthly token usage: ~10.9 million tokens
- Estimated monthly API cost: ~$65 at current rates
- Infrastructure and monitoring: ~$200
- Total monthly operating cost: ~$265
Sarah looked up sharply. "$265? Against $23,000 in labor savings?"
"For this agent, yes. The economics are compelling. But"—Michael paused for emphasis—"this is a simple agent. One model call per report. Your more complex agents will cost more."
He showed the forecast for the Exception Routing Agent, which required autonomous reasoning:
- Multiple model calls per exception analyzed
- Higher token counts for context retrieval
- Monthly cost estimate: ~$850
"Still profitable against the risk it mitigates," Kevin noted. "But the cost profile is different."
Financial Controls
Sarah was satisfied with the economics but not ready to sign off.
"What happens when something goes wrong? What if an agent gets stuck in a loop? What if usage spikes unexpectedly?"
Michael outlined the control architecture:
Budget Caps:
- Daily spend limits with alerts at 80% and hard stops at 100%
- Monthly budget allocation by agent and department
- Quarterly review triggers when variance exceeds 20%
Rate Limiting:
- Requests per minute caps to prevent sudden spikes
- Concurrent session limits
- Token limits per request to prevent context window abuse
Circuit Breakers:
- Automatic pause if costs spike unexpectedly
- Disable agent if error rates exceed thresholds
- Escalation when controls trigger
Sarah tested the logic. "So if the Pre-QA agent suddenly starts costing $500 a day instead of $10, it stops automatically?"
"Yes. And it alerts the finance team and the operational owner. You find out in minutes, not at month-end."
Kevin appreciated the design. "This is what we needed for cloud. Nobody built it. Nice to see it designed in from the start."
Real-Time Visibility
The final component was observability—Sarah needed to see the economics, not just trust the controls.
Michael showed the dashboard specifications:
- Cost per agent, per day, per department
- Token consumption trends
- Cost vs. forecast variance
- ROI tracking connecting cost to value delivered
"Every Monday, you'll get a summary: what each agent cost, what value it delivered, and how that compares to forecast."
Sarah nodded. "And if I want to dig in?"
"Real-time dashboard. Same data, drill-down capability, exportable for your board reports."
Lisa added a practical note. "This also helps us. If I see one agent costing three times what we expected, I can investigate before Sarah asks me about it."
The Sign-Off
As the session wrapped, Sarah summarized what they'd accomplished.
"I came in asking whether I could explain this to the board. Now I can. We know what we're spending, why we're spending it, what controls are in place, and how we'll know if it's working."
Michael delivered the final principle:
"Most AI programs fail economically not because the technology is expensive, but because nobody modeled the costs before deployment. You've done that work. When these agents go live, there won't be surprises."
Marcus, who had joined for the final discussion, appreciated the journey.
"Five pillars ago, I asked a simple question: how do we figure out AI? Now I have answers I can actually use. Priorities backed by evidence. Data we can work with. Governance I can defend. Experiences people want. And economics I can explain to shareholders."
Sarah closed her notebook. "I've signed off on a lot of technology initiatives. This is the first time I've understood the economics before we spent the money."