Recently, in a conversation, I said something rather blunt:
“Many people spend their whole lives as workhorses. Give them AI, and all they can think to do is ask it to teach them how to be workhorses.”
That remark came from a frustration I keep encountering at work.
I introduce people to Codex and explain that they can hand work over to an agent. They often respond with complete conviction: “My work is exceptionally complicated. AI can’t do it.”
They have not seriously tried it, supplied the necessary material, explained the result they want, or even found out which tools it can use. Yet they have already reached their conclusion.
Explain an approach, and they say their business is different. Suggest a concrete solution, and they introduce another condition they never mentioned before. By the end, the discussion feels little different from arguing with an internet troll who has no intention of being reasonable. They are defending the same conclusion throughout: “It could never do my work.”
I am losing patience with these discussions. This matters because it concerns what these people will be able to earn a living from next.
When I say “word-processing work,” I mean something much broader than typing documents.
Accountants organizing supporting documents and reconciling accounts; salespeople maintaining customer records and preparing proposals; legal staff reading contracts and comparing clauses; customer service staff organizing issues, looking up policies, and drafting replies; operations staff consolidating data and preparing reports—if the work mainly takes place on a computer, its source information can be digitized, and its final deliverable is also digital, it falls within the scope of this discussion.
“Words” here comes closer to the broad meaning of language in a large language model: text, numbers, tables, rules, code, and the relationships they express.
These things can be read, compared, calculated, transformed, and recombined. Many of the operations you perform by hand every day belong to this category.
At the World Governments Summit in February 2024, Jensen Huang put it strikingly:
“the programming language is human: everybody in the world is now a programmer”
In other words, human language becomes the programming language, and everyone can become a programmer. Original: NVIDIA’s official coverage ↗
The direction he describes is one in which ordinary people can instruct computers in natural language to do work. It also reinforces my belief that a person’s productive capacity increasingly depends on whether they can express their intent clearly.
Meanwhile, many people still treat “I have lots of steps to perform every day” as evidence that their work cannot be automated.
Being busy does not establish technical complexity.
Building a business system that actually enters service means dealing with requirements, data structures, business rules, interfaces, permissions, exceptions, testing, deployment, and the problems that arise once it is running. It has to work in the real world. Looking reasonable in a document is not enough.
Agents can now complete the entire workflow of coding, testing, and deploying even complex commercial software and websites. Such systems can involve hundreds of thousands of lines of code or more, with numerous modules, business rules, data, interfaces, permissions, and exception handling working together to run reliably in the real world. When agents can already deliver work at that level, I find it hard to accept someone insisting: “My office work is too complicated. Codex can’t do it.”
In my view, in terms of the workload and complexity of execution, engineering a complete commercial system is a hundred times more complex, or more, than the document work performed in all other office roles outside programming. Rating these everyday office tasks at one hundredth of that complexity would already be generous. This is my judgment about the difference in scale. But the question I am asking is very specific:
Can you actually explain where the complexity in your work lies?
Is the information unavailable? Are the rules undecided? Do several sources contradict one another? Is a business decision still outstanding? Or must someone with the appropriate qualifications and authority sign off on the result and take responsibility?
All of these can be discussed, and all of them need to be discussed.
A blanket statement that “my work is complicated” hides every one of these questions. It neither helps the agent complete the task nor proves that the agent cannot complete it.
The difficulty of professional judgment, the severity of the consequences of an error, and the tediousness of operating procedures need to be considered separately.
A contract may carry enormous liabilities. An accounting entry may affect an important decision. That is precisely why someone must judge whether the goal makes sense, the evidence is reliable, and the result meets the standard. It does not automatically establish that every act of reading, organizing, comparing, calculating, and drafting must continue to be performed manually by a person.
The change I really want to emphasize is that people must occupy a different position in the work.
Make AI the workhorse. Don’t ask AI to teach you how to become a more capable workhorse yourself.
Many people now use AI by asking a chatbot how to do something, getting a set of instructions, and then opening the software themselves and working through the steps.
They learn more complicated formulas, more proficient operations, and more efficient procedures. Then they continue doing the work by hand.
If the goal is to hand computer-based work over to an agent, this approach starts from the wrong position. The human still performs all the execution. AI merely helps that executor become more proficient. The executor improves, but the work remains tied to that person.
This misses what an agent means entirely.