“Manual or API?” is the wrong question for most high-volume image operations. The useful question is how much repetitive cutout work software can remove while keeping people in control of quality. We saw that distinction clearly in an enterprise pilot with a company processing graduation portraits for yearbooks.
The company had been outsourcing the manual cutouts and already knew that operation’s production rate and hourly labor cost. Some returned cutouts still required touch-ups before they were ready for the yearbook workflow. During the pilot, the company also recorded how long reviewers spent on every automated output, how often they denied the first result, and how long corrections took when the model did not perfectly isolate the subject. That gave us a comparison based on accepted work rather than API response time.
The subject matter was graduation photography, not a product catalog, but the operating question is the same one ecommerce teams face: when thousands of images need consistent cutouts, what happens to cost and throughput if automation performs the first pass and people retain final approval?
Pilot result
Automation handled the first cutout; people reviewed every image.
- The outsourced workflow averaged 8.5 manually cut images per labor hour, before separately accounting for occasional touch-ups.
- At $9.15 per hour, direct cutout labor was about $1.08 per image.
- Reviewing an automated output took about 15 seconds.
- Reviewers denied 4.32% of 5,000 outputs, or 216 images.
- Correcting a denied image averaged another 90 seconds.
- Processing plus review and corrections came to about $0.053 per image.
01
The graduation-portrait pilot
The customer’s team was cutting people out of graduation portraits for placement in yearbooks. This is repetitive work, but it is not consequence-free. Hair, gowns, caps, hands, and low-contrast edges still need to look right, and a bad cutout can become a printed defect rather than a temporary screen artifact.
We asked for the operating numbers the company already used to manage its outsourced workflow: average completed cutouts per labor hour and the average hourly rate for that work. We then replaced the initial outsourced cutout with BackgroundErase on our enterprise platform while preserving human review of each and every output.
This was a hybrid workflow, not unattended publishing
Every image still passed in front of a person. Automation changed the amount of work required before approval; it did not remove the customer’s quality gate.
02
The recorded manual baseline
Before the pilot, they used outsourced manual cutous at 8.5 completed cutouts per labor hour. One image therefore consumed about 7.06 minutes of labor. At an average labor rate of $9.15 per hour, the recorded initial-cutout labor cost approximately $1.08 per image.
That baseline did not guarantee a completely finished image. Some outsourced cutouts came back needing additional touch-up work. Because the customer did not provide a separate average for that touch-up time, the calculations below do not add it to the manual cost. The manual comparison is therefore conservative: it prices the recorded cutout operation without inventing a value for its remaining rework.
Manual throughput
8.5/hr Recorded outsourced cutouts per labor hourTime per image
7m 4s 60 minutes divided by 8.5 imagesDirect labor cost
$1.08 $9.15 divided by 8.5 imagesManual baseline
time per image = 60 minutes ÷ 8.5 images
= 7.06 minutes
labor per image = $9.15 ÷ 8.5 images
= $1.076Applied to a 5,000-image batch, that baseline represents about 588.24 labor hours and $5,382.35 in direct cutout labor. Those are the customer’s recorded averages projected across the same batch size used in the pilot comparison.
03
What changed in the hybrid workflow
BackgroundErase produced the initial cutout for $0.005 per image under the pilot’s enterprise pricing. A reviewer then inspected every result. When the output was acceptable, that review averaged about 15 seconds. When the model did not perfectly isolate the subject, the image was denied and received a human correction pass averaging 90 additional seconds.
Approve after review
Most images required only the automated cutout and a roughly 15-second human quality check.
Deny and correct
A small minority required the same initial review plus an average 90-second human correction pass.
This is the main difference between an API-only benchmark and a production workflow. The processor’s response time did not determine the labor savings. The important measurements were the time to approve an ordinary result and the time to repair an exception.
04
The measured 5,000-image result
Across 5,000 images, reviewers denied 216 of the automated outputs. That equals 4.32% denial rate and 4,784 images accepted after their initial review.
| Measure | Manual-only workflow | API + human review |
|---|---|---|
| Images | 5,000 | 5,000 |
| Initial cutout | Performed by the outsourced cutout team | Performed by BackgroundErase |
| Human review and touch-up | Some returned cutouts still required touch-ups; separate time was not recorded | Every image |
| First-review approvals | Not separately recorded | 4,784 |
| Denied automated outputs | Not applicable | 216 |
| Denial rate | Not applicable | 4.32% |
| Calculated human hours | 588.24 hours | 26.23 hours |
| Direct batch cost | $5,382.35 | $265.04 |
| Direct cost per image | $1.076 | $0.053 |
The hybrid calculation conservatively counts a 15-second review for all 5,000 images, then adds a 90-second correction pass for each of the 216 denied outputs. Direct cost includes pilot processing and labor at the supplied $9.15 hourly rate; it does not include integration, storage, management, or other overhead. The manual calculation excludes unmeasured touch-up work on returned outsourced cutouts.
The hybrid workflow reduced calculated human time from about 588.24 hours to 26.23 hours. That is a 95.5% reduction in direct labor time. Direct batch cost fell by approximately $5,117, or 95.1%, under the pilot inputs.
05
The cost and throughput math
The review team spent 20.83 hours inspecting all 5,000 outputs. The 216 denied images added another 5.4 correction hours, bringing total calculated human time to 26.23 hours.
Hybrid pilot calculation
denied images = 5,000 × 4.32%
= 216
review time = 5,000 × 15 seconds
= 20.83 hours
correction time = 216 × 90 seconds
= 5.40 hours
total human time = 20.83 + 5.40
= 26.23 hours
labor cost = 26.23 × $9.15
= $240.04
processing cost = 5,000 × $0.005
= $25.00
total direct cost = $240.04 + $25.00
= $265.04
direct cost per image = $265.04 ÷ 5,000
= $0.053Weighted human time
18.9s Average review and correction time per imageEffective throughput
190.6/hr Finished images per human labor hourThroughput change
22.4× Compared with the recorded 8.5-image manual rateThe effective rate is not the number of API calls a server can complete in an hour. It is the number of reviewed, corrected-when-needed images produced per human labor hour. That is the more useful throughput measure when every output still requires human approval.
06
What the denial rate means
A 4.32% denial rate does not mean 216 images were unusable or had to restart from a blank canvas. It means reviewers decided those automated cutouts were not ready to accept without intervention. The automated result still reduced the average correction pass to about 90 seconds, far below the roughly seven-minute manual baseline.
Review did not disappear
The company retained a human decision on all 5,000 images. The savings came from shortening that decision and concentrating editing on exceptions.
Denial is a workflow metric
It measures whether the first output met the customer’s acceptance standard, not whether the model found any part of the subject.
Corrections still benefited from automation
Denied outputs averaged a 90-second correction pass instead of a full manual cutout averaging just over seven minutes.
Category mix still matters
Graduation portraits have a particular distribution of hair, clothing, poses, and backgrounds. Another image program should measure its own denial rate.
07
How to evaluate your own workflow
This pilot should not be treated as a universal denial rate or labor forecast. All image catalogs have different edge conditions and acceptance standards. The method transfers more reliably than the exact result. In this case we were especially cost-effective compared to other pilots.
Record these values before and during the pilot
- Measure the current number of accepted images completed per labor hour.
- Use the fully burdened labor rate appropriate to your cost analysis.
- Review every pilot output, or define and document a statistically sound sampling policy.
- Record review time separately from correction time.
- Track denials with specific reason codes such as lost hair, clipped clothing, halo, or wrong subject.
- Count processing fees, retries, human labor, integration, storage, and operational overhead.
- Calculate cost per accepted image rather than cost per API request.
A hybrid workflow makes sense when the first automated pass is usually acceptable and the exceptions are faster to repair than to create manually. A pilot gives you both sides of that equation: the denial rate and the cost of resolving each denial.
Keep the quality gate
Move human time from repetitive cutouts to fast decisions.
In this 5,000-image pilot, every portrait still received human review. Automation reduced the weighted human requirement to about 18.9 seconds per image and direct cost to roughly $0.053 per finished output under the recorded assumptions.
Become an enterprise customer