TL;DR: For any workflow where AI handles some items and people handle the rest, C_eff = C_AI + (1 − R) × C_human, where R is the true resolution rate. Human handling usually costs far more than inference, so each point of resolution is worth more than large cuts in model price. Two adjustments make the formula honest: count false resolutions, and credit good handoffs that make the human faster.
The formula
Most production AI is partial automation. A support agent resolves some tickets and escalates the rest. A document reviewer clears simple cases and routes complex ones to an analyst. A KYC check passes clean submissions and flags the rest for manual review.
The cost per item across the whole workflow is:
C_eff = C_AI + (1 − R) × C_human
- C_AI is what the AI costs per item attempted, resolved or not. Use cost per attempt, including retries and all calls.
- R is the share of items the AI fully resolves, with no human touch and no comeback.
- C_human is the fully loaded cost of a person handling an item: wages, benefits, tools, management, quality review.
Every item pays C_AI. Only the unresolved share pays C_human.
Why resolution dominates
Take a document-verification workflow at a lending company. AI cost per document checked is $0.15. A manual review costs $2.00, a figure that would be higher in a European or US operations centre and lower in many Indian or Philippine ones.
| Resolution rate | Effective cost per document |
|---|---|
| 50% | $1.15 |
| 60% | $0.95 |
| 70% | $0.75 |
| 80% | $0.55 |
| 90% | $0.35 |
Each ten points of resolution saves $0.20 per document. Halving the AI cost from $0.15 to $0.075 saves $0.075. Moving from 70% to 80% resolution is worth more than two and a half times as much as halving the model bill.
This ratio holds in most workflows because C_human is typically ten to a hundred times C_AI. It is the single most useful fact in AI unit economics, and it changes where a team should spend its time: on the cases the AI fails, not on the token price.
Adjustment 1: false resolutions
R must be the true resolution rate. Some items the AI marks as resolved come back: a customer contacts support again, a document passed in error is caught downstream, an approved claim is later reversed. These cost the human handling anyway, plus whatever the error itself cost.
If 5% of the AI's resolutions are false and each costs the $2.00 review plus $1.00 of extra harm (a delayed loan, a complaint), then at 70% resolution:
C_eff = 0.15 + 0.30 × 2.00 + 0.70 × 0.05 × 3.00 ≈ $0.86
That is $0.11 more than the naive $0.75. Measure R with a comeback window (no re-contact, reversal or downstream catch within a set period), not as the share of conversations or items the system closed.
Adjustment 2: handoff quality
Escalated items do not have to cost the full C_human. If the AI hands over a structured summary (what the customer asked, what was checked, what failed, what it recommends), the person starts partway through. If good handoffs cut human handling time by 25%:
C_eff = 0.15 + 0.30 × (2.00 × 0.75) = $0.60
A further $0.15 saving, without changing the resolution rate at all. Handoff quality is an underrated product surface, and one that the formula makes easy to value.
Break-even and comparisons
The formula gives a quick break-even. AI is cheaper than an all-human process when:
C_AI + (1 − R) × C_human < C_human, which simplifies to R > C_AI / C_human
With C_AI = $0.15 and C_human = $2.00, any true resolution above 7.5% saves money. That is why even mediocre automation can pay off in pure cost terms. The harder constraints are usually quality and customer experience on the items the AI gets wrong, not cost.
The same formula compares two AI options. A vendor charging more per item but resolving more can be cheaper overall, and a cheap in-house system that resolves little can be expensive. Put both on effective cost, with the same definition of resolution, before comparing.
Where it misleads
C_human is not always saved. If support staff are not reduced or redeployed, the saving is capacity, not cash. Report it as such.
Segments differ. Resolution may be 90% on simple items and 20% on complex ones. Routing complex items straight to people avoids paying C_AI on attempts that will fail, and often gives those customers a better experience.
Quality costs are real. A cheaper workflow that frustrates customers can lose revenue that never appears in the formula. Pair effective cost with satisfaction and churn on AI-handled items.
In the interview
This formula is the core of most AI unit-economics questions: support agents, document processing, vendor comparisons, "is it worth it?" The scoring jump comes from showing that resolution rate dominates, with numbers, and then adjusting for false resolutions. Mentioning handoff quality as a cost lever is a strong signal that the candidate has thought about the human side of the workflow, not just the model.