“Why do I feel more tired after using AI?”
A friend of mine was venting about this the other day: “I used to write a proposal by sitting down, thinking it through, and finishing it in half a day. Now I ask AI to help. Sure, it spits out a huge draft in 10 seconds — but then I spend the next 30 minutes checking the facts, fixing awkward sentences, and correcting all the nonsense it made up with total confidence. This isn’t artificial intelligence. It’s like hiring a clueless intern who needs constant hand-holding.”
I’m guessing a lot of people can relate.
For the past two years, we’ve been surrounded by nonstop AI hype. We’ve been told AI can do everything at the push of a button. But in the reality of everyday work, AI often feels less like a productivity boost and more like a flow-breaking nuisance. So what’s really going on? Are we just bad at using AI — or is AI not as smart as we’ve been told?
A new paper by researchers from MIT, Yale, and Microsoft — Chaining Tasks, Redefining Work: A Theory of AI Automation — gets right to the heart of the problem.
Its central argument is blunt: if we only use AI to improve the efficiency of individual tasks, we’re dramatically underestimating its real value.
So let’s skip the academic jargon and talk about the hard truth this paper reveals in plain English: why AI often feels frustrating in practice, and how we should actually be using it.
How do most people use AI?
Take writing an article as an example: Outline the structure (human) ➔ generate a draft (AI) ➔ revise and polish it (human) ➔ proofread it (AI) ➔ format and publish it (human/AI).
See the problem? In that workflow, the work keeps bouncing back and forth between the human and the AI.
The paper points out that this task-by-task way of using AI hides a massive sinkhole: coordination and verification costs.
Every time work gets handed back from AI to a person, we have to read it, review it, verify it, and adjust it. That process breaks our train of thought and drains our attention. In many cases, the time people spend cleaning up after AI already cancels out whatever time AI saved in the first place.
That’s why using AI so often feels like managing an intern. We’ve chopped the work into pieces that are too small, and the cost of communication, supervision, and correction skyrockets.
Press enter or click to view image in full size
The Real Magic Is Task Chaining
So if one-off task assistance isn’t the answer, what is? The researchers introduce a crucial idea: task chaining.
Instead of seeing work as a collection of isolated tasks, we should treat it as a full pipeline. The real opportunity — for companies and individuals alike — is to bundle adjacent tasks together and let AI run through the entire chain from start to finish without interruption. Humans should step in only at the end to review the final output.
The paper uses a great example: lecturing versus tutoring. At first glance, both are forms of teaching. But their potential for AI automation is completely different.
Lecture-based teaching: A teacher’s prep work is continuous — review source material ➔ build slides ➔ generate classroom examples. These tasks are adjacent and tightly connected. You can hand that whole chain to AI and have it deliver a complete teaching package, with the teacher reviewing only at the end. The efficiency gains can be enormous.
Tutoring: Tutoring is a live, interactive loop — teacher explains ➔ student asks a question ➔ teacher adjusts based on the student’s response. In that setting, the work is constantly interrupted by human interaction. Because it can’t easily become one continuous task chain, AI’s automation value is much more limited.
That leads to a work-design principle that overturns a lot of conventional thinking: How tasks are arranged matters just as much as whether the tasks themselves can be automated. In other words, whether AI can truly transform a job often depends on whether we can stitch together the parts it’s good at into one seamless chain.
A Real Example of an End-to-End Automation Pipeline
Even before reading this paper, I had already come to the same conclusion through hands-on experimentation: if you want to unlock AI’s real power, you have to eliminate the high-friction handoffs between humans and machines.
Based on that insight, I designed and open-sourced a workflow for generating Xiaohongshu/RedNote image-and-text posts. In the past, making a single post meant doing everything in fragments: I would gather source material myself, ask AI to write the copy, manually format the post, then go back to AI again for images. It was the textbook definition of a high-friction workflow.
To eliminate that friction, I broke the process down into the smallest possible set of steps and built a system with one controller plus four independent execution modules:
Source collector: Scripts lock down the crawling logic and focus only on pulling content from the web and extracting useful information. That keeps AI from hallucinating fake news or invented facts.
Business writer: Once it gets a clean source draft, AI focuses on one thing only — turning that material into strong headlines and catchy social copy, complete with emojis and the right internet-native tone.
Image generator: The mechanical parts — calling the image API and saving files — are hard-coded into the workflow. AI only handles the part it’s best at: ideation, visual concepts, and writing prompts for beautiful images.
Review-and-storage module: Finally, scripts automatically package the text and images together and push everything neatly into a Feishu/Lark database for review.
In this pipeline, one controller script runs those four modules in sequence. The output of one module automatically feeds into the next. There’s no need for a human to jump in halfway through. The repetitive, mechanical work gets locked inside SOPs and code, while AI’s strengths — creativity and interpretation — get amplified.
At that point, all I have to do is drop in a link at the front end, then show up at the finish line with a cup of coffee and approve the final result in Feishu/Lark. That is exactly the kind of system-level efficiency the paper is talking about: eliminate friction, reduce handoffs, and let the chain run.
Press enter or click to view image in full size
At this point, some people will ask the obvious question: Does every single step in that pipeline actually outperform a human? For example, is the copy AI writes always more on-brand or more internet-savvy than what an experienced editor would produce?
Of course not. And that leads to one of the most counterintuitive — and most important — findings in the research:
System Efficiency Beats Local Perfection
A lot of managers and perfectionists instinctively push back here. They’ll say, “AI isn’t precise enough. My veteran employee is much better than AI at collecting data. Why would I hand the whole chain over to AI?”
The answer is this: AI does not need to outperform humans at every individual task in order to create enormous value.
Why not? Because the moment you insert that employee into the middle of the chain, the chain breaks. Once that person finishes the data collection, you now have another human-to-AI handoff — which means another round of transfer, verification, coordination, and correction.
The paper makes this point very clearly: even if humans are better than AI at certain intermediate steps, it can still make more sense to give the entire task chain to AI.
Why? Because when AI handles the whole thing from beginning to end, you remove the friction, reduce the handoff cost, and dramatically speed up total output — even if the quality at one step drops a little. What you’re saving is not just time on the task itself. You’re saving one of the most expensive resources in modern work: human coordination time.
It’s a lot like Henry Ford’s assembly line. Individually, a worker on the line might not have the craftsmanship of an old-school master artisan. But as a continuous system, the assembly line was overwhelmingly more powerful. In the age of AI, system-level efficiency beats task-level perfection.
Work Itself Is Being Redefined
Historically, we defined jobs by grouping tasks together in ways that matched human physical and mental limits. AI is now rewriting that equation.
The paper points out that in the early stages of AI adoption, the payoff may not even cover the cost. You have to buy tools, learn prompting, tolerate messy handoffs, and absorb all the coordination overhead. That’s the well-known productivity J-curve.
But once you cross that threshold — once you stop treating AI like a glorified typing tool and start redesigning the entire workflow around it, the real gains begin to show.
As that happens, traditional roles built around typing, formatting, and routine execution will be compressed. The future of work will be less about who can code faster or make prettier slides, and more about who can design better systems, spot better opportunities, and exercise better judgment.
If we can connect routine tasks into end-to-end AI task chains, we can free ourselves up to do the higher-value work that still depends on humans: complex judgment, strategic decisions, and emotional resonance. One person, paired with a rigorous automated workflow, can now do the work that used to require a team.
So maybe the question is no longer: “How can AI help me finish this task faster?”
The real question is: “How do I redesign this workflow into an AI pipeline that can run with little to no human intervention?”
From factory motors to wall outlets, the logic of technological progress has never really changed. The world is changing fast. AI is already powerful enough. What’s been too small isn’t the technology, it’s the role we’ve assigned it.
Top comments (0)