Research-backed calculator guide

AI Image Review Cost per Approved Asset: Price the Human QA Queue

Calculate AI image review cost per approved asset by combining generation attempts, reviewer minutes, correction work, fixed tooling, and approval yield.

What this metric measures

A raw image price answers, “What does one generation event cost?” The approved-asset metric answers, “What does it cost to get one file that the team can ship?” Those are different business units. OpenAI’s documentation distinguishes image generation and image understanding paths, and notes that image inputs are billed as tokens; its pricing page separately lists image input and output units. Record the model, endpoint, output size, quality, processing tier, and input-image count before copying any number into a worksheet. ^{[1] [2]}

For a provider-rate cross-check, use the Model pricing comparison. If your workflow also includes prompt or vision tokens, compare the assumptions with the AI token cost calculator and the LLM cost calculator before you roll those lines into the approved-asset denominator.

A review queue adds a second meter. A reviewer may check product fidelity, brand elements, text, composition, dimensions, rights flags, and channel requirements. Some candidates fail immediately; others pass visual inspection but need a crop, upscale, or small repair. If you count only the selected files, you hide the work required to find them.

This is a unit-economics problem, not just a model-pricing problem. The FinOps Foundation distinguishes resource-efficiency units such as cost per token from business units such as cost per transaction or case resolved. An approved image is the business unit in this guide because it connects spend to a deliverable. ^[6]

The worksheet formula

Use a batch boundary that everyone understands: one campaign, one product category, one SKU cohort, or one month. Then collect these inputs:

1. Approved assets: the number of files that pass the written rubric.
2. Candidates or attempts: every paid generation and every candidate that enters review. If one API call returns four outputs, count four review objects and record the API billing rule separately.
3. Generation cost per candidate: the dated provider quote or your measured blended rate, including resolution and quality settings.
4. Review minutes per candidate: the average time to open, inspect, label, and route a candidate. Use a sample from real queue logs instead of assuming that a quick glance is free.
5. Loaded reviewer rate: wages plus the employer costs or contractor charge you intend to allocate. A salary divided by 2,080 is not automatically a loaded production rate.
6. Correction minutes per approved asset: retouching, crop, typography repair, file naming, export, and final handoff that occur after a candidate is selected.
7. Allocated fixed cost: subscriptions, review software, integration, storage, or a coordinator’s fixed monthly time divided across the batch under a stated allocation rule.

The core lines are:

• generation spend = candidates × generation cost per candidate
• review labor = candidates × review minutes ÷ 60 × loaded reviewer rate
• correction labor = approved assets × correction minutes ÷ 60 × correction rate
• total batch cost = generation spend + review labor + correction labor + allocated fixed cost
• cost per approved asset = total batch cost ÷ approved assets
• approval yield = approved assets ÷ candidates

If reviewers inspect only a sample, replace candidates in the review line with reviewed candidates, and record the sampling rule. If rejected images are re-reviewed after a repair, add those review events too. The safest ledger has one row per candidate with an attempt ID, model, generation cost, review start and end time, decision, correction minutes, and final disposition.

Worked cost example

Assume a campaign needs 100 approved square product images. The team’s pilot shows 2.4 paid candidates per approved image, so the batch contains 240 candidates. The following numbers are planning assumptions for the example:

The review line uses a transparent benchmark-style assumption, not a market rate. HYPE-EDIT-1 describes a cost model using a $50 hourly rate and 20 seconds of inspection per candidate, which equals about $0.278 per candidate before retries. That study is an image-editing benchmark with its own tasks and retry cap, so use it as a reproducible reference point rather than as a universal staffing quote. ^[5]

The example also shows why “the model costs five cents” is not a useful final answer. Generation is only $12.72 of the $729.39 batch. Correction labor is the largest line because the example assumes six minutes of work for every accepted file. If your workflow performs no correction, remove that line; if only 30% of accepted images need correction, multiply correction minutes by 30 approved assets instead of 100. State the rule rather than hiding it in an average.

How to price the human QA queue

Start with observed minutes, then convert them into queue capacity. If one reviewer can spend 6 productive hours per day on this queue, a 20-second inspection time implies 1,080 first-pass candidates per reviewer-day before breaks, context switches, escalations, and re-review. That is a capacity estimate, not a promise of output. A queue with 2.4 candidates per approval would support 450 approvals per reviewer-day only in that idealized first-pass calculation; corrections, meetings, and hard cases reduce it.

Separate decision time from handling time. Opening a file, loading references, recording the rejection reason, and moving an asset to the right folder can take longer than the visual decision. Measure both. If reviewers must compare a generated image with source product photos, policy documents, or claim evidence, a 20-second assumption may be too low.

Use a loaded rate that matches the operating model. For a U.S. benchmark, the Bureau of Labor Statistics’ May 2025 OEWS publication documents the occupational-wage data program and its national/state/metro tables; it does not determine your company’s loaded rate or contractor quote. Pick a rate, date it, and disclose whether benefits, payroll taxes, management, and vendor margin are included. ^[4]

If you outsource the queue, keep the vendor’s per-review quote separate from internal QA management and escalation time. AWS’s Amazon A2I pricing page illustrates the distinction: billing can be per human-reviewed object, while workforce payments may be additional for Mechanical Turk or a marketplace vendor; reviews by employees have a different charge structure. Treat those as workflow/vendor prices, not as evidence that all human review costs the same. ^[3]

Yield and sensitivity checks

Approval yield is the fastest diagnostic: 100 approved ÷ 240 candidates = 41.7%. It lets you compare a cheaper model with a more expensive one without pretending the lowest API rate wins. A model that costs $0.03 per candidate but needs four candidates per approval can be more expensive than a $0.08 model that usually passes first or second attempt once review and correction are included.

For the worked example, holding labor and fixed cost constant gives this sensitivity:

The table rounds to cents and leaves $650 for the example’s $500 correction line plus $150 fixed allocation. It is not a forecast of any model’s pass rate. Re-run it with your own measured candidates-per-approval and review minutes.

Track at least four companion measures: first-pass approval rate, candidates per approved asset, human minutes per approved asset, and post-approval defects. The last measure matters because a fast approval that misses a label or product feature can create rework downstream. Keep the rubric stable while comparing models; changing the standard midway makes the unit metric hard to interpret.

What to exclude or disclose

Do not silently mix taxes, subscription minimums, model-training costs, storage, creative direction, legal review, and campaign-performance outcomes into a single number. Either include them in named lines or exclude them explicitly. A monthly subscription should be allocated by a declared denominator, such as approved assets in that month, not by a theoretical quota that was never used.

The metric also does not prove that an AI workflow is cheaper than a human-only workflow. It measures the cost of this defined route. Compare like with like: same output dimensions, same product references, same approval bar, same correction policy, and same time period. OpenAI states that image-generation and image-input billing use different units, so a pricing comparison that ignores input images or token accounting can understate request cost. ^{[1] [2]}

Finally, label every number as a vendor quote, benchmark observation, or internal estimate. The $0.053 generation value, $50/hour labor rate, 20-second review time, 6-minute correction time, and $150 fixed cost in the worked example are assumptions for demonstrating the calculation. They are not a promise, an industry average, or a quote from a provider.

FAQ

What is the simplest AI image review cost formula?

Divide all generation, review, correction, and allocated workflow costs for a defined batch by the number of approved images. Rejected candidates remain in the numerator and are excluded from the denominator.

Should I count every generated image in the review cost?

Count every candidate that a person actually inspects. If your process samples candidates, use the reviewed count and document the sampling rate. Add re-review events after corrections rather than counting one review per file by default.

Is reviewer salary the same as review cost?

No. Salary is only one possible input. Use a loaded employee rate or an all-in contractor/vendor rate and state whether benefits, payroll taxes, management, tooling, and vendor margin are included.

How do retries change cost per approved AI image?

Retries increase both the generation numerator and usually the review numerator. Measure candidates per approval or approval yield. Dividing one generation price by the pass rate is a useful first approximation, but it will miss correction, escalation, and fixed workflow costs.

Can an AI judge replace human QA in this calculation?

It can change the workflow, but it does not make review free. Add the judge’s input/output or per-call cost, calibration, exception handling, and the human time spent reviewing disagreements. Keep a human audit sample if your risk policy requires it.

What is a reasonable approval rate for AI images?

There is no universal rate that can be safely applied across models, prompts, products, resolutions, or rubrics. Run a representative pilot, record first-pass and final approval separately, and use your observed yield in the worksheet.