Before you take on more AI delivery work: split review time 3 ways
Your AI delivery capacity comes from human review time, measured as routine checks, exceptions and redelivery.
Key points
- AI capacity is set by human review time
- Log review time 3 ways for 2 weeks
- Find jobs you can take: hours ÷ job time
- Before taking more, cut the biggest time
Contents
Human review time sets AI delivery volume
A&A perspective
How much AI delivery work you can take on is set by human review time, not by generation speed. AI delivery here means delivering documents and materials made with AI to clients. Human review time is the time people spend checking and fixing that work so it meets the quality you promised.
Hypothetical example
Say AI now produces a draft in ten minutes, and you wonder whether you can double your orders. If the people who look at the work before delivery are still just you and one contractor, faster drafts do not raise the number of pieces the two of you can check.
From the sources
In December 2024, Anthropic published an article on how to build systems that hand work to AI, using coding as an example. It says automated tests can verify that the code works, but human review remains crucial for making sure the solution fits the broader system requirements.
From the sources
Gumloop, a company that makes a tool for building AI automations without code, wrote in a February 2026 post on its own blog that its support team is two people. By the time a person opens a support ticket, the customer's information has already been gathered. A system that watches error rates on its own product also alerts a person only when it finds a problem that needs action.
A&A perspective
Automate & Augment (A&A), the company writing this article, reads both the Anthropic and the Gumloop articles as cases where human review has not been removed. Gumloop gathers the information before its two people look, and narrows down what it alerts them about.
A&A perspective
If you plan around deliveries that no person touches at all, you cannot estimate how much you can take on. What you need first is a count of how many minutes the remaining review takes. Deciding whether to keep a product based on its cost per job is covered in "Low-ticket service profit rests on human touches, not AI speed". This article looks at the number of jobs you can take on in a month.
Split review time into 3 kinds and log it for 2 weeks
A&A perspective
Split review time into three kinds, routine checks, exception handling and redelivery, and log it for two weeks. The reason for three is that each one applies to a different set of jobs.
A&A perspective
Routine checks are time spent on everything you deliver. Checking typos and figures and matching them against the source material count here.
A&A perspective
Exception handling is time spent only when the input falls outside what you expected, such as digging again for missing material or asking the client.
A&A perspective
Redelivery is time spent fixing work that was sent back after you delivered it. It applies only to the jobs that were returned.
A&A perspective
If you lump all of this together as "review", the difference disappears. Exceptions stand out, so they feel heavy. Even so, once you add it up, routine checks, which apply to every job, can be the larger share.
A&A perspective
You do not need a stopwatch. For each job, one line saying which of the three it was and how many minutes it took is enough. Two weeks is a rough guide for filling the table with actual counts rather than memory.
Hypothetical example
For example, take a two-person company that delivers 600 product descriptions a month. Its log would have lines like "routine check, 3 min", "exception, 25 min, model number not in the material" and "redelivery, 15 min".
A&A perspective
What the log gives you is your own current figures. They are not numbers to compare with other companies.
| Time | Jobs it applies to | What to log |
|---|---|---|
| Routine check | Every job you deliver | Minutes per job |
| Exception handling | Jobs whose input falls outside what you expected | Number of cases and minutes per case |
| Redelivery | Jobs sent back | Number of returned jobs and minutes per job |
Divide by review time per job to count jobs
A&A perspective
The number of jobs you can take on is your monthly review hours divided by review time per job. Review time per job is the routine check time plus the exception handling and redelivery times, each multiplied first by how often it happens.
A&A perspective
As a formula: jobs ≈ monthly review hours ÷ (routine check + exception rate × exception handling + return rate × redelivery). Generation time is not in the denominator. Even if generation gets twice as fast, the number does not move.
Hypothetical example
Let us put hypothetical numbers on the company above. It delivers 600 pieces a month and has 60 hours for review. A routine check takes 3 minutes per piece, or 30 hours for 600. Exceptions occur in 8%, or 48 pieces, at 25 minutes each: 20 hours. Returns are 5%, or 30 pieces, at 15 minutes each: 7.5 hours.
Hypothetical example
The total is 57.5 hours, almost all of the 60 hours available. Take on more orders in this state and what grows is late delivery. All of these numbers are a hypothetical example, not measurements from A&A or its clients.
From the sources
The December 2024 Anthropic article says it is common to give AI that keeps acting on its own a stopping condition, such as a maximum number of iterations. The purpose is to keep control.
A&A perspective
A&A applies the same idea to delivery terms. Without a cap on redeliveries, the back-and-forth on a single job eats into the month's review hours, and the formula can no longer give a number.
Hypothetical example
For example, say a hypothetical company's terms read: "Up to two returns per deliverable. A third, or a rewrite because the source material changed, is quoted separately."
From the sources
Gumloop writes that, to confirm that the solutions its support team provided actually work, it identifies users to follow up with every day.
A&A perspective
If you use the return rate in the formula, you need one extra step after delivery to check that the fix worked. If an unfixed job slips into the next request, it is counted as a new job, and you overestimate the number.
A&A perspective
Some work does not fit this formula. When every job has different specifications and the content of the routine check changes each time, time per job does not settle. In that case, narrow the range of work you accept before counting.
Before taking more orders, cut the largest time first
A&A perspective
Before taking on more orders, reduce the largest of the three times first. Exceptions are what catch your eye, but routine checks, which apply to every job, can be larger.
Hypothetical example
In the example above, halving exception handling frees 10 hours. Cutting the routine check from 3 minutes to 1 minute per piece frees 20 hours. In this example, cutting routine checks first raises the number of jobs you can take on by more.
A&A perspective
Routine checks simply happened to be the largest in this example. In work with few jobs that are each heavy, exception handling can be the largest. In that case, cut exceptions first.
From the sources
In January 2026, Anthropic published an article on how to build ways of measuring how well AI performs, and it divides grading into three methods. Grading by code is fast, cheap and objective. Grading by people is described as the gold standard for quality, but expensive and slow.
From the sources
Grading by a model is described as flexible but more expensive than code, and as needing calibration for accuracy: comparing it with human grading and correcting the gaps. The same article says a good problem for measuring AI is one where two experts judging separately would reach the same pass or fail.
A&A perspective
A&A reads these three as a way to sort the checks you make before delivery. For each item, ask whether you can write the reason for pass or fail in one sentence, and whether someone else reading that sentence would reach the same verdict. If both are yes, give it to a machine that matches; if the reason can be written but people split on the verdict, give it to a model; if the reason cannot be written, keep it with a person.
Hypothetical example
In the product description example, model numbers, dimensions and character counts can be judged by matching against the source material. Whether the text follows the brand's voice, and whether it adds specifications not in the material, can be judged by a model once the criteria are written. Client-specific calls, such as what to push this season, stay with people.
From the sources
The Anthropic article recommends calibrating a model grader closely with human experts. Once the system is robust, it says, using human review only occasionally is enough.
A&A perspective
What is easy to miss here is the calibration period. Calibration means having both people and a model judge the same work and correcting the gaps. During that period, review time goes up for a while. A&A thinks the month you move checks to a machine should not be the month you take on more orders.
From the sources
Gumloop also writes that its current setup was not complete from the start but grew one workflow at a time over months.
A&A perspective
To start, working on the single largest item in your log is enough.
Summary
A&A perspective
How much AI delivery work you can take on is set by human review time, not by generation speed.
A&A perspective
Split review time into routine checks, exception handling and redelivery, and log it for two weeks.
A&A perspective
Divide monthly review hours by review time per job to get the number of jobs you can take on.
A&A perspective
Take on more orders only after you have cut the largest time.
What sets the number of AI delivery jobs you can take on is human review time, not generation speed. Log review time for two weeks split into routine checks, exception handling and redelivery, then divide monthly review hours by review time per job to get the number. Take on more orders only after cutting the largest of the three.
FAQ
Will a faster model let me take on more jobs?
Not while review time is the limit. The number of jobs you can take on is monthly review hours divided by review time per job, and generation time is not part of that calculation. To take on more, reduce routine checks, exception handling or redelivery.
If AI does the review, does review time drop to zero?
No. Anthropic writes that even for code, where automated tests can be used, human review remains crucial. While you move checks to a machine, review time rises for a while because of calibration.
Does freed-up review time turn straight into profit?
No. Revenue rises only when another order fills that time. This estimate is for choosing where to cut first, not a promise of profit.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- Building effective agents
Anthropic · 2024-12-19 (page notes tooling has changed since)
Accessed 2026-10-22 - Demystifying evals for AI agents
Anthropic · 2026-01-09
Accessed 2026-10-22 - Supporting the world's most AI-native companies with a 2-person team
Gumloop · 2026-02-17
Accessed 2026-10-22