Are these numbers a quote?

No. Every default on this page is a clearly labeled assumption for illustration, not a benchmark and not a price. Actual costs vary with the task, the reviewer, the model, and the vendor.

Why is there a 5 to 30 times token toggle?

Some teams report that an agentic run burns many times the tokens of a single direct model call. That range is not a fact about your stack. The toggle exists only as a sensitivity test, so you can see how much the math moves if your runs turn out token-heavy.

What counts as supervision here?

The human minutes spent reviewing each output, the loaded cost of the person doing it, and the retries those reviews trigger. Incident response, audit trails, and liability sit outside this page's math.

When does the agent still win?

When review minutes stay low, rework stays rare, and the manual alternative is expensive or slow. Run your own numbers above; the threshold usually shows up in seconds.

Free cost planning tool

The agent costs pennies. Supervising it costs the real money.

Two columns of honest math: what a successful agent task really costs once you add human review, rework, and infrastructure, and how much of the promised saving supervision eats. Every default on this page is a labeled assumption, not a benchmark.

The calculator

Run your numbers

Tasks per week40

How many agent runs you put through review each week. Assumption: 40.

Minutes of human review per task10 min

Reading, checking, and signing off on one output. Assumption: 10 min.

Rework rate20%

Share of tasks that fail review and get retried. Assumption: 20%.

Model spend per taskC$0.40

What the model call costs you per run, at your rates. Assumption: C$0.40.

Minutes a person needs without the agent30 min

The manual baseline the pitch is measured against. Assumption: 30 min.

Test the token multiplier Sensitivity only

Some teams report an agentic run burning 5–30x the tokens of a single direct model call. That range is not a fact about your stack; it only shows how sensitive the math is to token-heavy runs.

Show your work

How the math works

Reviewer cost: C$70/hr loadedAssumption
Tools + infrastructure: C$0.15 per taskAssumption
Manual time per task: 30 min, adjustable aboveAssumption
Rework: One retry + one more review per reworked taskAssumption
Token multiplier 5–30x: Sensitivity test only, never a fact about your stackAssumption
Retry success: Every reworked task succeeds on retryAssumption

Assumptions used by the calculator
Assumption	Value	Status
Reviewer cost	C$70/hr loaded	Assumption
Tools + infrastructure	C$0.15 per task	Assumption
Manual time per task	30 min, adjustable above	Assumption
Rework	One retry + one more review per reworked task	Assumption
Token multiplier 5–30x	Sensitivity test only, never a fact about your stack	Assumption
Retry success	Every reworked task succeeds on retry	Assumption

Cost per successful task = (model spend × token multiplier) + tools + infrastructure + human oversight + rework × (one full retry). The multiplier is 1 unless the sensitivity toggle is switched on.
Human oversight per task = review minutes ÷ 60 × reviewer cost, counted once for the first pass and again for every retry.
The supposed gain = manual cost per task − machine cost per task. The pitch prices the machine only; supervision is the part it leaves out.
Share eaten by supervision = (human oversight + rework) ÷ supposed gain per task.

Common questions

Before you decide

Are these numbers a quote?: No. Every default on this page is a clearly labeled assumption for illustration, not a benchmark and not a price. Actual costs vary with the task, the reviewer, the model, and the vendor.
Why is there a 5–30x token toggle?: Some teams report that an agentic run burns many times the tokens of a single direct model call. That range is not a fact about your stack. The toggle exists only as a sensitivity test, so you can see how much the math moves if your runs turn out token-heavy.
What counts as supervision here?: The human minutes spent reviewing each output, the loaded cost of the person doing it, and the retries those reviews trigger. Incident response, audit trails, and liability sit outside this page's math.
When does the agent still win?: When review minutes stay low, rework stays rare, and the manual alternative is expensive or slow. Run your own numbers above; the threshold usually shows up in seconds.

Talk it through

If the math makes the business case wobble, bring your agent rollout to a focused conversation. No pitch, no obligation.

Discuss your AI governance needs

Disclaimer: all figures are estimates for illustration only, not a quote, offer, or commitment. Actual costs vary by task, reviewer, model, and vendor. Confirm current legal requirements and obtain qualified advice for your facts and jurisdiction. This tool is maintained by Mohammad Movahedi.

If oversight is the real cost, buy a service that includes it, or do not buy the agent.