Handing Routine Tasks to AI: I Run a Comparison First, Then Catch Exceptions

I wrote “Machine as Maker, Human as Checker” last month, mentioning that the Checker needs to look at the underlying execution logic. This piece completes that section: How I actually verify the things AI delivers.

For things like expense reporting and contracts where mistakes are very troublesome, I always review them once myself; AI only helps me take the first look.

For high-volume routine tasks, like the mandatory training rosters I handle, running hundreds of people at once means if a human has to check every single entry, they easily end up just clicking “confirm” all the way through.

Wingerter’s team conducted an experiment in 2025 using the Cognitive Reflection Test. The results showed that participants who received flawed AI assistance scored less than half the correct answers compared to the control group without AI assistance¹.

It is the same in the workplace. The tidier the spreadsheet delivered by AI, the easier it is for people to feel there are no problems, and they conveniently just click approve. For high-volume data, relying on “looking closer” is not enough; I switch to a different verification method.

Let the AI Shadow First, Write into a Skill Only When Results Align

Almost everything that can be operated on a computer now can be delegated to AI, making it a very reasonable choice for execution. But precisely because of this, how to verify the output becomes crucial. It is like being a manager; you always have to review the work submitted by your subordinates. Now, I just spend my time verifying the results delivered by AI.

In practice, I first let the AI shadow me. I tell it the method and watch it execute step by step. Every intermediate step, which column needs to align with which, and what conditions require separate judgments are all clarified during this time.

After shadowing, you cannot just accept it because the spreadsheet produced by AI looks neat. I manually complete the task myself and compare my results with the AI’s output. Only when the total count, amount, or number of people perfectly align on both sides does it mean the AI truly did it right.

Only when I confirm the results align do I ask it to write the previous process into a Skill (which is the SOP it will execute from then on).

The person who comes to compare the results is usually the original person responsible for this task—they hand the work over to the AI for execution, so naturally, they personally verify it. AI being honest is useless on its own, but AI plus experience can amplify output. To recognize errors, you must first know where this type of business usually makes mistakes, which only someone who has done it firsthand knows. Letting the AI handle execution while the human focuses on using experience to control quality—this is true human-AI collaboration.

Continue Parallel Comparison for the First Month or Two After the Skill Goes Live

flowchart TD;

subgraph S1["Stage 1: Establish Execution Rules"];

A["Human leads AI in shadowing"];

B["Clarify steps and columns"];

C["Compare with manual output"];

D["Solidify into Skill upon consistent output"];

A --> B;

B --> C;

C --> D;

end;

subgraph S2["Stage 2: Live Parallel Comparison"];

E["Human and AI each do it once"];

F["Catch exceptions the Skill missed"];

G["Modify Skill to add rules"];

E --> F;

F --> G;

end;

subgraph S3["Stage 3: Exception Sampling and Automation"];

H["Specifically sample high-risk edge cases"];

I["Human steps back to Checker for judgment"];

H --> I;

end;

D --> E;

G --> H;

Having this Skill does not mean you can immediately let go completely. For the first month or two after it goes live, when encountering routine tasks, I even ask the AI to “shadow” me again—I do my manual work, it runs its Skill, and then we compare and verify the results from both sides.

The initial shadowing was to establish the rules; the mutual shadowing now is to catch omissions. During this period, my comparison focus is: In the real environment, did any exceptions pop up that the Skill didn’t anticipate initially? As soon as a new exception occurs, I go back and modify the Skill to add the rule. After modifying and rerunning, this “month or two” comparison period must also restart.

Only when we’ve run in parallel for a month or two without any errors, confirming the Skill is stable enough, do I switch to sampling.

When sampling, I only sample exceptions. The AI naturally handles normal data well; if I sample ten entries, all ten will probably be correct, which is reassuring but fails to catch loopholes. The things that go wrong are usually new situations not written in the Skill. When I was doing the mandatory training platform before, I focused my sampling verification on places prone to exceptions, like “source file format changes” and “regulatory hour modifications.”

Besides sampling, I also actively look for counterexamples, not waiting for sampling to happen to catch one. For the logic part, I pass it to another model for a second look. In my experience, switching to a model from a different source for checking is more useful than having the same model rerun it, because a model usually still cannot spot the mistakes it made itself upon rerunning.

This happens to coincide with the conclusion of an academic study from July this year². The research confirmed that having different models check each other can significantly improve the accuracy of catching errors; moreover, the model that is best at answering questions is not necessarily the best auditor for catching others’ mistakes. This explains one thing: in practice, the point is not to find the smartest model to carry the whole load, but to establish a cross-model cross-checking mechanism to effectively break homologous blind spots.

Hand Over Execution So Humans Can Focus on Being the Checker

This is the true face of human-AI collaboration. At first, you must personally step in and collaborate with the AI for a period, guiding it through step by step.

Once the AI thoroughly masters the complete workflow, you can solidify it into a Skill, or even build it directly into an automated workflow like a Python script, ensuring it stably produces the same results every time it executes.

Once this production line is established, the labor of routine tasks is truly stripped away from humans. And “humans” can finally step away from endless execution, focusing on playing the Checker role well, spending time on the highest-value verification and judgment. This is the true purpose of handing our work over to AI.

If you are also handing routine tasks over to AI, you can find more methods and observations on human-AI collaboration in Field.

Extended Reading

• Machine as Maker, Human as Checker: How Routine Tasks Can Be Reassigned in Human-AI Collaboration
• You Cannot Be a Checker Without Having Been a Maker: Newcomers Skipping Implementation Face a Professional Gap

References