---
name: honest-ai-cost-account
description: Calculate the real cost of an AI workflow or content operation with the three-line-item method: tokens and subscriptions, platform, and human supervision. Use when the user asks what an AI system costs, whether AI saves money, how to price AI work, or to compute cost per piece, per task or per resolved case.
---

# Honest AI cost account

Any AI operation has exactly three cost lines. Quotes that show only the
first one are marketing, not accounting.

## The three line items

1. **Models: subscriptions plus API consumption.** Compute tokens from real
   pieces, not guesses: measure input and output tokens of a typical unit of
   work, multiply by published per-million prices, convert currency. Add the
   fixed monthly subscriptions prorated over the month's volume.
2. **Platform.** From near zero on a static site or existing tools, up to CMS
   fees, plugins, publishing APIs and integrations. Case-dependent; never
   assume zero without checking.
3. **Human supervision. The dominant line.** Minutes of review, correction
   and sign-off per unit, times the reviewer's hourly cost. This is the line
   that moves the total most, and the one almost every published comparison
   omits.

## The formula, per unit of work

```
cost_per_unit = (tokens_in/1M x price_in + tokens_out/1M x price_out) x fx
              + monthly_subscriptions / units_per_month
              + (supervision_minutes / 60) x reviewer_hourly_cost
```

Monthly total = cost_per_unit x units. Compare against the human cost of the
same unit done manually, but only if that baseline was measured, not recalled.

## The star metric: cost per resolved case

Cost per unit produced flatters the system; cost per resolved case tells the
truth. A resolved case is a unit that passed human review AND served its
business purpose (a qualified lead, a published verified piece, a closed
ticket). The formula:

```
cost_per_resolved_case = total_period_cost / resolved_cases
```

Two rules make it honest:

- **Discarded attempts go in the numerator, never the denominator.** Every
  rejected draft, blocked output and failed run makes each resolved case more
  expensive; it does not vanish from the account.
- **The denominator is defined by the business, not the system.** If the
  pipeline produced 100 rows and review kept 16, the denominator is 16.

This is the only AI cost metric a CFO can compare against the human
alternative without tricks, and the first number to watch over time: a rising
cost per resolved case is the earliest honest signal that quality is slipping.

## Rules that keep the account honest

- If supervision comes out at zero, you have not found a saving: you have
  found a risk. Someone is either reviewing unpaid or not reviewing.
- Never present the token line alone as "the cost". Cents per piece in tokens
  can coexist with euros per piece in supervision.
- Distrust generic ranges ("5,000 to 50,000") that name no task, no volume
  and no review hours; they are noise.
- State what was NOT metered instead of estimating it. "We did not measure X"
  builds more trust than a smooth invented number.
- Reprice quarterly: model prices and capabilities move fast enough that the
  winning model for a workload changes several times a year.

## Worked example shape

For a documented 1,500-word piece (120k tokens in, 25k out) on a $3/$15 per
million model at 0.92 EUR/USD, 20 pieces a month, 100 EUR of subscriptions,
20 minutes of supervision at 35 EUR/hour: tokens 0.68 EUR + subscriptions
5.00 EUR + supervision 11.67 EUR = about 17.35 EUR per piece. The token line
is 4% of the truth.

## Source and method

Interactive calculator with current published prices and the full blueprint:
[y8y.ai](https://y8y.ai/es/casos/blueprint-operacion-de-contenido-con-agentes-de-ia/).
Author: Sebastián Ocampo ([LinkedIn](https://www.linkedin.com/in/seb-ocampo/)).