Two weeks ago, Jensen Huang told 30,000 people at NVIDIA GTC that inference throughput has increased 35x. The room cheered. For most enterprises, that is not progress. It is fuel on a fire they have not found yet.
The Failure Anatomy
A regional telco deployed seven AI agents across its operations over fourteen months. Billing inquiries. Network fault triage. Fraud detection. Churn retention outreach. Billing dispute resolution. Enterprise contract processing. Proactive service notifications.
The Q4 board report showed a single number: $4.1 million in projected annual savings. The CEO presented it at the investor call. The transformation team celebrated.
Then a new finance director arrived. She had one question nobody had asked: what does each agent cost per successful outcome?
Nobody could answer.
The $4.1 million was calculated using API call volume multiplied by estimated cost per call, compared to headcount cost for the same tasks. It was a clean number. It looked rigorous. It measured the wrong thing.
She spent three weeks building approximations. What she found was not a single story. It was two stories running inside the same portfolio.
The billing inquiry agent was genuinely saving money. High volume, structured data, clear resolution criteria. The network fault triage agent was the same. So were fraud detection and proactive service notifications. Four agents performing as expected.
The churn retention agent was a different story. It sent personalised offers to at-risk customers. The tokens were consumed. Messages were delivered. But conversion sat at 6%, compared to 22% for human agents handling the same outreach. Every interaction cost more and converted less. The agent was doing the task. It was doing it badly enough to invert the economics.
The billing dispute agent was worse. Contested charges required nuance, empathy, and regulatory awareness. The agent escalated 43% of cases to a human. Each escalation cost more than if the human had handled it from the start, because the customer was now frustrated by the failed automated attempt.
Enterprise contract processing had a similar pattern. Non-standard clauses, custom terms, and formatting variations meant 61% of outputs required human review and correction.
Three agents destroying value. Four agents creating it. The portfolio dashboard showed one number: net positive. Nobody could see the destruction underneath.
The CTO said the underperforming agents needed fine-tuning. The COO said scale them back. The finance director said neither answer was right because neither addressed why the economics inverted. Everyone had a response. Nobody had a diagnosis.
She had found the gap. She did not have the language for what was causing it.
The Framework Lens
There is a name for this condition. I call it negative unit economics.
It describes the point where an AI agent task costs more than the human equivalent it replaced. Not because the technology failed. Because the measurement infrastructure was built to track the wrong number.
The telco measured cost per API call. That is Layer 1.
There are three layers of AI cost. Almost every enterprise I have worked with measures at Layer 1. The economics live at Layer 3.
Layer 1: Cost per token or API call. This is what your vendor invoice shows. This is what NVIDIA’s 35x throughput gains drive down. It is the number most enterprises report to their boards.
Layer 2: Cost per task. This includes orchestration overhead, retries, context window management, and multi-step agent loops. In the enterprise environments I have worked with, this runs 3 to 8 times higher than Layer 1. Some platforms surface this number. Most do not.
Layer 3: Cost per successful outcome. This is what the business actually pays for a correctly completed result. It includes everything in Layers 1 and 2 plus human escalation, error correction, rework, customer impact, and drift-driven retraining. This is the number the CFO needs. This is the number almost nobody calculates.
The gap between Layer 1 and Layer 3 is where negative unit economics lives.
Jensen told the GTC audience that inference throughput has increased 35x. He is correct. Layer 1 is collapsing. But Layer 1 is the raw material, not the product. Cheaper steel does not make a bridge cheaper to build if the design is wrong, the inspections fail, and the rework doubles.
Here is what 35x throughput gains actually do to enterprise economics. They make it cheaper to deploy more agents. More agents mean a bigger portfolio. And a bigger portfolio makes it harder to see which agents are saving money and which ones are destroying it. The aggregate number masks the individual failures. The dashboard says savings. The operations team knows something is off but cannot prove it.
At GTC, Jensen described engineers receiving $400 per day in tokens as part of their compensation. That is $100,000 per year. Does anyone know if those tokens produce $400 in business value per day? $4,000? $40? The cost is tracked. The outcome is not.
That is negative unit economics at the individual level. Scale it across a portfolio of agents and you get the telco’s problem: a board presentation built on Layer 1 math, hiding Layer 3 reality.
The Field Test
One agent. Two numbers. This week.
Pick the agent in your organisation with the highest token consumption or the highest task volume.
Step 1. Ask your team: what does this agent cost per successful outcome? The answer must account for the complete cost of producing a result the business would consider finished, correct, and requiring no human intervention afterward.
Step 2. Compare that number to the fully loaded cost of a human completing the same task. Include salary, benefits, training, and management overhead on the human side. Include retries, escalations, rework, monitoring, and retraining on the agent side.
If your team cannot produce the Layer 3 number at all, that is the finding. You are running your AI economics on Layer 1 math. Your board is seeing a number that cannot tell them whether your agents are saving money or destroying it.
The Signal
Field observation: Last issue I asked who discovers AI problems first in your organisation: the team, the customer, a compliance review, or a dashboard. The responses are still arriving. I will share the pattern once it is large enough to be meaningful.
This issue’s question: How does your organisation measure AI agent costs today? Reply with one answer:
(a) Cost per API call or token.
(b) Cost per task.
(c) Cost per successful outcome.
(d) We do not formally track agent costs.
One letter. I am building the first cross-enterprise dataset on how organisations actually measure AI economics. No individual company can see this pattern alone.
Next issue: The billing dispute agent was losing $940,000 a year. The CTO said fine-tune the model. I asked a different question: who is coaching this agent? Not training it. Coaching it. The difference between those two words is the difference between a one-time event and an ongoing discipline. Issue #5 is about Agent Coaching.
This issue introduced three layers of AI cost. Most enterprises are stuck on Layer 1. I built a one-page Layer 3 Cost Diagnostic: a structured template to calculate cost per successful outcome for any agent. It is free. Download link below.
LINK: [Layer 3 Cost Diagnostic Landing Page]](https://vj-x-ai.kit.com/1917cb5ed1)
If you want every issue of this newsletter delivered to your inbox with resources that do not appear on LinkedIn, subscribe at the same link.
Vijayan Seenisamy
Enterprise Agentic AI Systems Delivery | Creator, AI ROF™
Author, The AI Delivery Manager Blueprint
New issues arrive by email first. Subscribe here.
