A conversation keeps repeating itself in companies that have put AI agents to work. The AI inference bill arrives, larger than last month's, and nobody in the room can say where it came from. Not for lack of competence, but because the number arrives consolidated, with no split by process, by agent, or by unit of work. This article is about getting out of that position. It is not a vendor negotiation problem. It is an instrumentation problem.
The paradox that confuses companies
Intuition says that if an input gets cheaper, the bill falls. With AI agents, the opposite happens:
These numbers are worth unpacking, because they explain almost everything. Three things multiply:
- Calls to AI models per task. An AI agent does not ask one question and stop. It reads, makes a decision, calls a tool or a model, evaluates what comes back, and tries again. That is 5 to 30 calls to complete one unit of work.
- Price per token. The per-token price gap between the cheapest and the most expensive model is roughly 4,500-fold. Which model runs each step is not a technical detail: it is the largest cost lever there is.
- Volume. Whatever works becomes routine, and that routine runs thousands of times, not always in an optimized way.
A falling unit price does not offset those three factors rising at once. That is why "tokens are cheaper now" is a true statement, but an irrelevant one for your budget.
- 01Read
- 02Decide
- 03Call a tool or a model
- 04Evaluate what comes back
Try again
Unit of work complete
Calls per task
How many times the loop runs.
Price per call
Which model runs each step.
Volume
How many times the routine runs.
Why the bill does not answer the CFO's questions
The inference bill is organized in the way that best suits the vendor — by account, by API key, or by period. The business needs that information organized by process: what it costs to reconcile an invoice, review a contract, or triage a claim.
Without that translation, there is no decision to make. There is only a shock when the bill lands.
How the invoice arrives
- By account: costs from all three processes, mixed together
- By API key: costs from all three processes, mixed together
- By period: costs from all three processes, mixed together
How the business asks
- 1Reconcile an invoice: costs from this process only
- 2Review a contract: costs from this process only
- 3Triage a claim: costs from this process only
And this problem is neither rare nor being solved adequately:
Read those two numbers together: nearly half of CFOs name the uncertainty of AI costs as their biggest concern, and only a minority can see that cost in real time. Cost, for that matter, is among the reasons Gartner gives for cancellations:
The four numbers that make cost and benefit visible
Every completed unit of work should produce four measured numbers:
- Token cost per unit of work — what the AI reasoning consumed.
- Infrastructure cost per unit — what execution consumed outside the model: compute, storage, calls into systems.
- Human intervention rate — how often a person had to step in on a problematic result.
- Accuracy rate — how often results were accepted with no correction needed.
Unit of work
Invoice or entry reconciled
What it cost
- Token cost per unit
- Infrastructure cost per unit
How well it worked
- Human intervention rate
- Accuracy rate
Measured and published, per unit.
The first two numbers, in isolation, can mislead. An AI agent can be very cheap in token and infrastructure cost and very expensive in the final bill, because it needs human review in most cases — and a professional's hour is the most expensive item on the list. The four numbers together tell the real story: what it cost, how much of it worked, and how much human labor it took.
This is why AInda charges per unit of work delivered rather than per hour or per license: in agentic AI, the unit of work is the smallest thing that can be measured, compared, and optimized.
How to measure this in practice
It does not depend on new platforms. It depends on discipline, in four steps:
- Define the unit of work before automating. "Invoice reconciled," "contract reviewed with an opinion," "claim triaged." If you cannot define the unit today, you will not be able to measure it later.
- Measure the human baseline. What it costs and how long it takes today, without AI agents. Without that measurement, any later result is only an opinion — and contracting based on results and improvements is off the table.
- Tag every step with the process that triggered it. This is the step that is rarely formalized, and its absence is what makes cost hard to measure and attribute.
- Measure the four numbers per process, monthly, building a history. Numbers nobody sees do not change behavior.
The potential gain justifies the work:
Note McKinsey's qualifier, though: in early pilots. The conversion from pilots that impress into operations that sustain themselves is almost always very low:
What changes once measurement exists
AI inference bills stop being a monthly disappointment and become portfolio decisions: run more of these processes, fewer of those; use this AI model for the simple steps, that one only where the step genuinely demands it. It is the same reasoning any COO already applies to other costs per unit produced, now applied to a new input.
Companies that measure can scale predictably. Companies that do not are funding a team of AI agents without knowing what it costs or what it delivers — and that bill shows up in executive committees, where budget refusals come from.
AInda's unit-of-work engagement models are described under Offerings.