ChatGPT, generative AI, RPA, Codex.
In just a few years, the tools for automating work have multiplied.
AI can draft your text. You can ask AI to handle your Excel totals. With RPA (software robots that repeat the clicking and typing a person would do on a computer), you can automate screen operations and copying data from one place to another. With an AI that actually carries out tasks, like Codex, you can push a huge amount of work forward in one go: articles, landing pages, code, documents, data processing, even improvement proposals.
And yet, if you look at real workplaces, this kind of work is still hanging around.
- Open Excel
- Copy
- Paste
- Check it by eye
- Retype it into another system
- Double-check that there are no mistakes
- Confirm it all over again in a meeting
- Get a correction request, then fix it by hand again
This is what I'd call bamboo-spear labor for the AI era. (The name comes from wartime Japan, when civilians were drilled to fight modern weapons with sharpened bamboo poles: lots of effort, wrong tool.)
Of course, Excel itself isn't the problem. Excel is actually great. For spreadsheets, totals, makeshift databases, forms, charts, checklists, tracking sheets, and as a temporary shelter for whatever the workplace needs, it's still excellent.
The problem isn't using Excel.
The problem is that work Excel could handle is still being run on manual effort, eyeballing, grit, and confirmation meetings, year after year.
Usage numbers show we're still at a pretty early stage
You might be thinking, "Doesn't everyone use AI and RPA these days?"
It's true that awareness of ChatGPT has become very high. According to Reuters, Sensor Tower estimates that the ChatGPT app reached 1 billion monthly active users worldwide in May 2026.
On the other hand, the number of people who can hand the actual execution of their work over to AI and automation is still quite small.
According to Axios, OpenAI's Codex had more than 5 million weekly active users as of June 2026. About one in five of them were reported to be knowledge workers.
It's an AI tool for carrying out tasks, available worldwide, and it has just over 5 million weekly users. That's fewer people than live in Aichi Prefecture, the home of Toyota in central Japan.
Compared with the ChatGPT app's scale of 1 billion monthly users, Codex's 5 million-plus weekly users is quite small.
So here's the gap:
- More people ask AI questions
- More people have AI polish their writing
- But few people hand actual work to AI
- And even fewer can hand work over in pieces small and clear enough for AI to carry out
That's where the big business opportunity is.
RPA adoption is still nowhere near high
RPA is the same story.
RPA is a way to have software robots take over routine tasks that people do on a computer. Data entry, copying between files, totals, verification, producing forms, work spanning several systems: it's a good match for manual Excel work.
But in Startear Raise's 2024 survey, the RPA adoption rate for Japanese companies overall was 13.04%. It was 8.51% for small and medium-sized companies, and even large companies only reached 27.69%.
In other words, most Japanese companies still haven't adopted RPA.
The same survey also asked about problems before adopting RPA. The No. 1 answer was "too much manual work (entering, copying, registering data and so on)," and No. 2 was "too many mistakes from manual work."
That's pretty telling.
Workplaces have a lot of manual work. They have a lot of mistakes. They have labor shortages too. And still, RPA adoption is only around 13%.
So the demand is there: "We have so much manual work, we want to automate it." But in reality, most companies are still stuck doing it by hand.
Rookie mistakes get fixed by system design, not by blaming character
This ties into how we train new hires.
Workplaces see things like this all the time:
- A new hire types in the wrong data
- A new hire gets a file name wrong
- A new hire saves a file in the wrong place
- A new hire skips an item on the checklist
- A new hire uses an outdated template
- A new hire misunderstands a deadline or who to send a request to
And the people around them say:
"New people make so many mistakes." "Please check properly." "I already told you before." "Why do you keep making the same mistake?"
But the question isn't only about the newcomer's character or attention span.
What we should really look at is whether the setup lets a newcomer make mistakes so easily.
A new hire doesn't know the background of the job yet. They don't know the unwritten rules. They don't know the habits of the person before them. They don't know why you look at that column, why the file has that name, or why that check is needed.
Throw them into free-form typing, checking by eye, verbal explanations, hand-copying, and checks that depend on one particular person, and mistakes are only natural.
Scolding them every time is less effective than building a setup where mistakes are hard to make in the first place.
For example, changes like these make a big difference:
- Use dropdown lists instead of typing by hand
- Use multiple choice instead of free text
- Add input rules for dates and amounts
- Show a warning if a required field is blank
- Generate file names automatically
- Fix the save location
- Make it impossible to use an old template
- Automatically check the entries for contradictions
- List only the likely errors
- Show a confirmation screen before sending
- Keep a log when the work is done
- Turn common mistakes into a checklist
This isn't about going easy on new people.
It's about building what needs to be taught into the system itself.
Instead of telling people to "be careful," it's faster to build a path where mistakes are unlikely, using Excel's input rules, VBA (Excel's built-in macro language), RPA, forms, checklists, and AI review.
And it's easier on the newcomer, too.
They don't get scolded every time. Whoever trains them doesn't have to repeat the same warning every time. Managers get less buried in cleanup after mistakes.
So building mistake prevention into the system isn't just about efficiency.
It's a safety design that lowers the mental cost for both the newcomer and the person teaching.
If you're saying "new people make too many mistakes," the first thing to check isn't their personality.
It's whether the job itself is a bamboo-spear design that invites mistakes.
Generative AI is still stuck at the company level, too
With generative AI, too, how individuals feel about using it doesn't match how far companies have actually adopted it.
In a Nikkei Research survey reported by Reuters in July 2024, about 24% of Japanese companies had already adopted AI, 35% planned to, and 41% had no plans to.
In other words, however much AI is in the news, at the company level only about one in four has adopted it. Around 40% answered that they have no plans to use it.
We see the same pattern here.
It's a hot topic. But it hasn't made it into daily work.
People think it's useful. But they haven't built it into their workflows.
More people have tried ChatGPT, but few can actually break their work into steps and hand it off to AI, RPA, VBA, or Codex.
So why is everyone still doing Excel by hand?
The reason isn't simply that people are lazy.
There are five main reasons.
1. They can't break the work down
To automate anything, you first have to break the work into steps.
For example, to automate copying data between Excel files, you'd have to decide things like:
- Where is the input data?
- Where does the output go?
- Which column goes into which column?
- What are the exceptional cases?
- When an error happens, do you stop or just record it in a log?
- What part does a person check at the end?
Without that sorting-out, you can't get VBA, RPA, or AI to do anything.
Many people can do the work itself. But they struggle to give the work a clear structure and hand it to a machine in a form it can use.
That's where they stall.
2. The work depends on particular people
Manual Excel work almost always ends up depending on one person.
"That person does it." "I inherited it from my predecessor." "Better not touch this cell." "I don't know why this formula is like this." "We just kind of do it this way every month."
Once it gets like this, you don't even know what the work means, let alone how to automate it.
If you bring in RPA without understanding what the work means, all you get is a faster version of a mystery task.
Automation doesn't mean swapping a machine in for the work. First, you make the meaning of the work visible.
3. Few people can use VBA or RPA
Plenty of people can use Excel functions, but once you get to VBA and macros, the numbers drop sharply.
I couldn't find reliable public data on how many people use VBA. But from what I see on the job, many people can use Excel, and very few can build real work processes with VBA.
And even when someone can use VBA, work tends to pile up on that one person.
The result looks like this:
- There's a lot of manual work
- Few people can automate it
- Requests all go to the few who can
- Those people become the office handyman
- No one else can maintain what they built
- In the end, everyone goes back to doing it by hand
It's a zoo full of holes. (My own nickname for a workplace that's understaffed and patched together with holes everywhere.)
4. They use AI only as someone to consult
Even people who use ChatGPT mostly still use it as someone to consult.
- Fix my writing
- Summarize this
- Give me ideas
- Write an email
That's useful in its own way.
But if you stop here, AI ends up as a handy notepad.
The real power is beyond that.
- Break the work into steps
- Turn it into requirements
- Write it up in Markdown (a simple plain-text format)
- Hand it to Codex
- Push it to GitHub (a site where code is stored and shared)
- Reflect it in landing pages, articles, and apps
- Turn it into an automation script
- Even create the test items
Don't ask AI, hand work to AI.
Few people have reached this stage yet.
5. Companies are afraid of risk
When a company uses AI or RPA, there are real risks.
- Personal information
- Confidential information
- Access control
- Wrong outputs
- Who's responsible
- Security
- Pushback from the people on the floor
- Fit with existing work
Given all this, it's natural for companies to be cautious.
But being cautious is not the same as doing nothing.
Don't enter personal information. Process things locally. Try it with dummy data. Keep a human review step. Don't use the output as-is. Keep logs.
If you set up your operations this way, you can start small without rolling it out company-wide on day one.
The problem is when fear turns into "ban everything." Do that, and the people on the floor stay stuck in manual Excel work forever.
The manual Excel economy will stick around for a while
Looking at all this, manual Excel work on the front lines will probably stay for quite some time.
The reasons are simple:
- RPA adoption is around 13% across Japanese companies overall
- Under 10% among small and medium-sized companies
- For generative AI, only about one in four companies has adopted it
- Codex has just over 5 million weekly users worldwide
- Few people can handle VBA or RPA
- Even fewer can break work down and hand it to AI
In other words, the world hasn't fully entered the "age of AI doing the work" yet.
In the news, AI is amazing. On social media, stories about using AI keep flowing by. But on the floor, there's still copy-paste, retyping, checking by eye, and confirming in meetings.
This isn't a pessimistic point. It's actually an opportunity.
The people who'll be valued are "workflow plumbers"
What gains value in the AI era isn't simply someone who knows about AI.
What gains value is someone who can turn work into a form that can be handed to AI, RPA, VBA, and Codex.
In other words, a workflow plumber.
On the front lines, there's a pile of scattered tasks. There's Excel. There's email. There are PDFs. There are handwritten notes. There are meetings. There's tacit knowledge that lives only in particular people's heads.
Turn all of that into the following:
- Purpose of the task
- Input information
- Output format
- Criteria for decisions
- Handling of exceptions
- How to review
- What can be automated
- What a person checks
- Instructions to hand to Codex or RPA
The people who can get it down to this level are strong.
Work in the AI era is shifting from "the person who does everything by hand" to "the person who designs work so it can be handed to AI."
"Can hand off work" beats "can use AI"
What will make the difference from here on isn't whether you've used AI.
The number of people who have used AI will grow. The number of people who have tried ChatGPT will grow. Many people will be able to do text generation and summaries.
But few people can yet do the following:
- Break their own work into steps
- Pick out manual tasks as candidates for automation
- Decide which Excel columns to use
- Define how to handle exceptions
- Write the Markdown to hand to AI
- Give Codex the conditions for "done"
- Reflect the results in landing pages, articles, and apps
- Look at the results and improve
This is where the gap opens up.
So the edge from now on belongs not to "AI users" but to "people who design how AI gets work done."
Which manual tasks to look at first
So where should you start automating?
Start by looking at tasks like these:
1. Tasks you repeat every month, week, or day
The more often you repeat a task, the bigger the payoff from automating it.
Even one hour a month adds up to 12 hours a year. If several people do it, the total is bigger still.
2. Tasks with a lot of copy and paste
Copy-paste is a prime candidate for automation. There's little point in a person doing it.
If there are many cases that need judgment, though, start with partial automation instead of going fully automatic.
3. Tasks with a lot of checking by eye
Checking by eye is tiring. Mistakes creep in.
If the conditions for checking are clear, you can cut it down a lot with Excel functions, VBA, Python, AI, or RPA.
4. Tasks you confirm in every meeting
What you confirm in every meeting can very likely be turned into a table or checklist ahead of time.
Just changing from "ask in the meeting" to "have people fill it in beforehand" can make things much more efficient.
5. Tasks only the previous person understood
Work that only your predecessor understood is dangerous.
First, write it up as a standard procedure. Then look for the parts you can automate.
Automating without standardizing is risky. You'd just be speeding up work nobody understands.
6. Tasks where new hires often make mistakes
Tasks where new hires often make mistakes should be a high priority for automation and standardization.
That's because there's usually one of these behind it:
- The rules are unwritten knowledge
- The input fields are too free-form
- The criteria for decisions aren't written down
- Handling of exceptions depends on people's memory
- There are too many checks
- You only find out about a mistake after you've made it
This is less a problem with the new hire than a problem with how the work is designed.
Where a new hire makes mistakes, a veteran will too when tired. Where a new hire gets stuck, a handover will get stuck too. What a new hire can't understand hasn't been put into a standard procedure, either.
So a new hire's mistakes are valuable debug logs. (A debug log is the record programmers read to find where a program goes wrong.)
Instead of seeing "they made another mistake," see it as "this part demands too much attention from people."
Just having that viewpoint makes training a lot easier.
Small steps to escape bamboo-spear labor
You don't need to launch a company-wide digital transformation overnight.
Just start small.
Step 1: Pick one manual task
Pick one Excel task you do every month. Copying data, totaling, checking, verifying file names, or listing PDFs is fine.
Step 2: Write the steps out in words
Write the steps as if you were explaining them to someone.
- Which file you open
- Which columns you look at
- What conditions you judge by
- Where the result goes
- What to do about exceptions
Step 3: Ask AI which parts can be automated
Ask ChatGPT or similar something like this:
Of the steps in this task, please sort out which parts can be automated with Excel functions, VBA, RPA, Python, or generative AI. Full automation isn't necessary; partial automation is fine. Please propose them in order from lowest risk.
Step 4: Start with partial automation
Don't go fully automatic from the start.
First, make a checking sheet. Show only the likely errors. Total things with one button. Automate only the check before copying data.
That's enough.
Step 5: Once it works, make it a template
Once you've got one working, turn it into a template.
- Procedure manual
- Checklist
- VBA code
- RPA scenario
- AI prompt
- Markdown to feed into Codex
If you keep all of this, your personal work becomes an asset.
Beware of becoming the office handyman
One caution, though: people who can automate things inside a company need to be careful.
The more you can see, the more you get made to pick up. The more you can build, the more you get made to carry. The more you can improve, the more you become the handyman.
This happens a lot in the zoo full of holes.
The company doesn't provide the time for standardization, someone responsible, recognition in evaluations, or a maintenance setup, and just lets the individuals who can do it soak up the improvements.
The result is that the people who can improve things burn out.
So when you propose an improvement at your company, you need to bundle these four points together:
- Who makes the decisions
- Who runs it
- Who maintains it
- How far your own role goes
If you improve things while leaving these vague, your good intentions get drained away.
Spend the minimum at your company. Give your all on your own assets.
Keeping those two separate is what matters.
Wrap-up: While everyone does Excel by hand, build the plumbing
AI is already amazing. There's RPA. There's VBA. There's Codex.
But looking at usage numbers, few people have managed to hand off the actual execution of their work.
The ChatGPT app has spread to a scale of 1 billion monthly users. Meanwhile, Codex has just over 5 million weekly users. RPA adoption is 13.04% across Japanese companies overall. About 24% of Japanese companies have adopted AI.
In other words, the world hasn't completely changed yet.
Many workplaces still open Excel, copy and paste, check by eye, and confirm in meetings.
That's behind the times. But it's also an opportunity.
What will be valuable from now on isn't someone who just plays with AI. It's someone who breaks work into steps and puts it in a form that can be handed to AI, RPA, VBA, and Codex.
While everyone's doing Excel by hand, we build the plumbing.
While the monkey meeting copy-pastes, we run the Markdown behind the scenes. (The "monkey meeting" is my joke for a pointless meeting where everyone just copies and pastes by hand.) While the manual Excel economy lasts, we build the AI execution factory.
That's the winning move right now.
