On this page
Know What Your AI Workflows Actually Cost
What Your AI Workflows Actually Cost
Most Organizations Are Flying Blind
Ask any AI team lead what their workflows cost, and you'll get the same answer: the monthly API bill. Maybe broken down by provider. Maybe by project. But almost always limited to the one cost that comes on an invoice.
This captures only model API spend — which is often the smallest component of true workflow cost. Human review time, error correction, rework caused by output quality failures, prompt engineering labor, and monitoring are almost never tracked at the workflow level. They're buried in team labor budgets, attributed to other departments, or simply invisible.
The result: no one knows what their AI workflows actually cost. And you cannot optimize what you cannot see.
< 30%
The average enterprise tracks less than 30% of its true AI workflow cost. (Illustrative estimate)
What Is Baseline Cost per Workflow?
Baseline Cost per Workflow is the true fully-loaded cost of running one AI-assisted workflow today — including model API costs, human review time, error correction cycles, downstream rework, prompt engineering overhead, and monitoring costs.
The key phrase is "per workflow." Not per token. Not per month. Not per team. Per workflow execution. This granularity matters because it's the only level at which you can compare alternatives, identify waste, and make informed decisions about model selection, architecture changes, or staffing.
The other key phrase is "fully-loaded." This means every cost that contributes to producing a usable workflow outcome — not just the compute, but the labor, the rework, the engineering overhead, and the monitoring that keeps it running. Token cost is not workflow cost.
The Baseline Cost Formula
Baseline Cost per Workflow =
Model API Cost
+ (Human Review Rate × Review Time × Hourly Rate)
+ (Error Rate × Correction Time × Hourly Rate)
+ Downstream Rework Cost
+ Prompt Engineering Overhead (amortized per execution)
+ Monitoring and Observability CostMultiply by execution volume to get total workflow cost.
The Six Cost Components
1. Model API Cost
What it is: The direct cost of model inference — input tokens, output tokens, and per-request fees.
Why it's underestimated: It’s the only cost most teams track, which makes it feel like the whole cost. It rarely is. For most enterprise workflows, API cost represents less than 5% of the fully-loaded baseline.
How to measure: Multiply average input tokens by input price per token, add output tokens by output price, and multiply by calls per execution.
For the engineer: You see this on the dashboard. It’s the number you optimize against.
For the CFO: This is the line item on the invoice. Everything else is hidden in labor budgets.
2. Human Review Time
What it is: Any human touching an AI output before it enters production — reviewing, editing, approving, or escalating.
Why it's underestimated: Most teams don’t track review rates at the workflow level. A 40% review rate at 12 minutes per review and $75/hour adds $6.00 per execution — 150x a typical API call.
How to measure: Track the percentage of outputs that require human review, the average time per review, and the fully-loaded hourly rate of the reviewer.
For the engineer: This is the Slack message asking ‘can you check this before I send it?’
For the CFO: This is labor spend you’re probably not attributing to AI.
3. Error Correction Cycles
What it is: Rework triggered when AI outputs fail quality checks — regeneration, human correction, and delay costs.
Why it's underestimated: Teams track error “rate” but not error “cost.” A 15% error rate with 20-minute corrections at $75/hour adds $3.75 per execution for every output that fails.
How to measure: Track error rate, average time to correct, and whether correction involves regeneration (additional API calls), human editing, or both.
For the engineer: This is the retry loop. The one you built ‘just in case.’
For the CFO: This is unbudgeted rework disguised as normal operations.
4. Downstream Rework
What it is: Errors that escape review and cause problems further down the chain — in consuming systems, reports, or customer-facing outputs.
Why it's underestimated: This is the hardest component to quantify and the highest impact. Downstream rework is often attributed to the team that discovers it, not the AI workflow that caused it.
How to measure: Estimate the frequency and cost of downstream failures attributable to AI output quality. Start with incident reports and work backward.
For the engineer: This is the bug report that traces back to a hallucination three steps upstream.
For the CFO: This is the cost center no one can explain — because no one connected it to the AI workflow.
5. Prompt Engineering Overhead
What it is: The ongoing labor cost of writing, testing, iterating, and maintaining prompts.
Why it's underestimated: Prompt engineering is treated as a one-time setup cost. In practice, prompts require continuous refinement as models update, use cases expand, and output quality drifts. For complex workflows, this cost can exceed model API cost.
How to measure: Track engineering hours spent on prompt work per workflow per month. Amortize across execution volume.
For the engineer: This is the time you spend tweaking system prompts after every model update.
For the CFO: This is an engineering cost center with no budget line — because it didn’t exist two years ago.
6. Monitoring and Observability Cost
What it is: The tools, dashboards, and labor required to watch workflows in production and detect degradation.
Why it's underestimated: At early stages, monitoring is often zero — teams ship and hope. At scale, underfunded monitoring means problems compound silently until they become incidents.
How to measure: Include SaaS tooling costs, engineering time spent reviewing dashboards, and the cost of any custom observability infrastructure.
For the engineer: This is the Datadog bill and the on-call rotation you built around it.
For the CFO: This is infrastructure spend that scales with AI adoption — whether you budget for it or not.
What It Looks Like in Practice
Illustrative figures for a mid-sized company running an AI-assisted contract review workflow at 500 executions per month. These numbers are for methodology demonstration only.
| Cost Component | Per Execution | Monthly (500 runs) |
|---|---|---|
| Model API Cost | $0.04 | $20 |
| Human Review | $1.20 | $600 |
| Error Correction | $0.60 | $300 |
| Downstream Rework | $0.80 | $400 |
| Prompt Engineering | $0.15 | $75 |
| Monitoring | $0.05 | $25 |
| Total Baseline | $2.84 | $1,420 |
What teams think they pay
$20/mo
API cost only
What they actually pay
$1,420/mo
Fully-loaded baseline
The ratio is 71:1. For every dollar this team thinks they're spending on AI, they're actually spending $71. And this is a single workflow. Scale to ten workflows, twenty, fifty — and the gap between perceived and actual AI spend becomes an organizational blind spot with real financial consequences.
This gap is not unusual. It's structural. API costs are the only component that arrives on an invoice. Everything else is distributed across labor budgets, attributed to other teams, or simply untracked.
Why API Cost Is Just the Tip of the Iceberg
API costs are visible because they arrive on a single invoice from a single provider. They have a line item, a date, and a dollar amount. This makes them easy to track, easy to report, and easy to optimize against — which is exactly why teams default to them as the measure of "what AI costs."
But human review, error correction, and rework are distributed across team labor budgets. Prompt engineering is treated as engineering time, not AI cost. Monitoring is an infrastructure line item. No one is connecting these dots at the workflow level — which means no one has a complete picture of what any individual workflow actually costs to run.
This blind spot becomes critical when new model tiers arrive. When a model like Claude Mythos enters the market at 3-5x the per-token price of current models, teams will evaluate it on API cost alone — and many will reject it. But if that model reduces human review rates from 50% to 15% and error rates from 25% to 5%, the fully-loaded baseline cost per workflow may actually decrease. Without baseline measurement, you'll make the wrong call.
How Baseline Cost per Workflow Connects to Croms and TCWO
Baseline Cost per Workflow
You can't optimize what you can't see. Baseline is the starting point.
(This page)
Croms™
Croms identify where the waste is — scoring across Cost Waste (35%), Latency Overhead (25%), Failure Risk (25%), and Structural Gaps (15%).
TCWO
TCWO translates your Croms score into dollar terms over time — giving finance the number they need to make AI investment decisions.
You cannot reduce what you cannot measure. Baseline Cost per Workflow is where measurement begins.
How to Calculate Your Baseline
- 1
Identify the workflow
Define the AI-assisted workflow you are measuring. Include all inputs, outputs, model calls, and human touchpoints. Be specific — "contract review workflow" is measurable; "legal AI" is not.
- 2
Calculate model API cost
Multiply average input tokens by input price per token. Add average output tokens multiplied by output price per token. Multiply by the number of model calls per workflow execution. This is the one number you probably already have.
- 3
Calculate human review cost
What percentage of outputs require human review before use? How long does each review take? What is the fully-loaded hourly rate of the reviewer? Multiply all three. This is usually the largest component.
- 4
Calculate error correction cost
What is the error rate — the percentage of outputs that fail quality checks? How long does each correction take? Multiply error rate by correction time by hourly rate. Include the cost of any additional API calls for regeneration.
- 5
Add downstream rework cost
Estimate the labor and opportunity cost of downstream failures attributable to AI output quality. Start with incident reports and customer complaints, then work backward to the workflow that caused them.
- 6
Add prompt engineering overhead
Track engineering hours spent on prompt work per workflow per month — writing, testing, iterating, and maintaining. Amortize across the execution volume for that month to get a per-execution cost.
- 7
Add monitoring and observability cost
Include the cost of any tooling, dashboards, or labor used to monitor this workflow in production. If you're not monitoring it, this cost is zero — but the risk isn't.
- 8
Sum all components
Add all components to produce the fully-loaded Baseline Cost per Workflow execution. Multiply by monthly or annual volume to get total cost. Now you have a number you can compare, optimize, and track over time.