Ask whether AI is coming for this job and you will get two answers, both confident and both wrong. Here is what the employment data shows, and the test I use to work out whether any given task is genuinely at risk.
What the employment data actually shows
The US Bureau of Labor Statistics projects employment for secretaries and administrative assistants at roughly 3.5 million jobs in 2024, with little or no change expected through 2034. It also projects about 358,300 openings each year over that decade, arising mostly from people leaving the occupation rather than from new positions being created. BLS names productivity-improving technology as the force holding demand flat.
Read that carefully, because the headline and the detail say different things.
The occupation is not collapsing. It is also not growing. What is happening is a long, quiet substitution: the same number of seats, filled by people doing a different job than the one their title describes.
The risk is not that the role disappears. It is that the role changes underneath you while the title stays the same, and nobody tells you.
Why the honest answer is boring
Two claims are made constantly and both are wrong.
The first is that AI will eliminate the role. This is usually argued from a list of tasks a language model can perform: drafting an email, summarising a meeting, formatting a document. All true, and all beside the point, because those tasks were never the job. They were the visible surface of it.
The second is that AI changes nothing. This is usually argued by people whose value genuinely does rest on those tasks, and it is the more dangerous position, because it is comforting.
The defensible position sits between them, and it is uncomfortable enough that very little content occupies it: the mechanical half of this work is being automated, the judgment half is becoming more valuable, and the people who struggle are the ones whose role never contained much of the second half.
The test for any given task
Rather than argue in the abstract, take a task you did this week and run it through two questions.
Does the answer ever change?
If the output is the same shape every time regardless of circumstance, a machine can produce it. Producing a weekly report from three systems, generating recurring work, routing an intake form: none of these require anyone to decide anything, and all of them are currently done by hand in most small businesses.
If the answer changes depending on the client, the person, or the week, you are looking at a decision wearing the costume of a task. Which of two conflicting priorities wins. Whether this message should be sent today or on Monday. Whether the founder needs to hear this now. Those do not automate, because the difficulty was never the execution.
Was it ever written down?
This is the question that decides whether automation is available to you at all. A process nobody has documented cannot be automated, only guessed at, and pointing a tool at an undefined workflow produces the same mess faster. In practice this is where most automation projects actually fail, long before any tooling decision is made.
What is genuinely hard to replace
Not because a model cannot generate the words, but because the value is not in the words.
| What is being replaced | What is becoming more valuable |
|---|---|
| Producing the draft | Knowing which draft should never be sent |
| Summarising the meeting | Noticing what was not said in it |
| Scheduling the call | Deciding it should not happen this week |
| Assembling the report | Reading it and knowing which number is wrong |
| Answering the question asked | Answering the question that was meant |
The right column shares a property: every item requires accountability. Somebody has to be answerable for the choice afterwards, and accountability is not a capability you can install.
It also requires context that exists nowhere in writing. Which client is sensitive about tone. Which internal deadline is real and which is aspirational. That the founder says they will look at it later and means it about half the time. None of that is in any system, and a tool cannot retrieve what was never recorded.
What I would actually do about it
Not reassurance. Three things, in order.
Automate your own mechanical half before somebody else does. The person who removes forty hours of manual work a quarter from their own role is not competing with the software. They are the reason it was deployed, and they own the result. I have written separately about doing this with scripts rather than subscriptions.
Document what only you know. Being the sole holder of undocumented knowledge feels like security and is a ceiling, because nobody gets promoted out of a job only they can do. Writing it down is what creates room to take on the half that is getting more valuable.
Then take work that involves a decision. Owning an outcome rather than executing a task is the entire difference between the half that is being substituted and the half that is not.
It was never AI against executive assistants. It is assistants who moved into the judgment half against assistants who did not.
The short answer
No, AI is not replacing the executive assistant, and the employment projections support that rather than the panic. But the version of the role that consisted mainly of executing defined tasks is being substituted, steadily and without announcement, and the projections support that too.
Which half of the job you occupy is largely a choice, and it is still available to make.