Why this job shifts rather than vanishes
Ask whether AI will replace administrative assistants and the useful answer sits in the task mix, not the job title. The role is a bundle of small jobs: calendars, inboxes, travel, expense reports, meeting notes, filing, phones, visitors, supplies. A lot of that is text in and text out, which is where language models are strongest. Drafting routine correspondence, formatting documents and turning a recorded meeting into minutes no longer need a person at a keyboard for every line.
What holds the role together is the part that is not text. Someone has to decide which of two urgent requests wins. Someone chases the signature that never arrives, re-books the room when the room is double-booked, and reads the mood in a department before sending a message to all staff. Those judgment calls are short, constant and specific to one office. They are also the hardest things to hand to software, because the right answer depends on people and politics rather than a rule.
Scale matters here too. About 1.7 million people held this job in the United States, with median pay of $47,540 a year, and the Bureau of Labor Statistics projects employment down 6% from 2025 to 2035 (BLS, 2025). That is erosion, not a cliff. Tasks move into software, job ads ask for more coverage per person, and fewer brand-new seats open at the bottom. You can watch that pattern across junior roles on our entry-level hiring tracker.
What software finishes, what it assists, and what stays with you
Start with the work tools can already complete end to end. Transcribing a meeting and producing clean minutes, drafting a standard letter or confirmation email, keying and cleaning spreadsheet data, and sorting an inbox into categories all run well without a person writing each version. Our scoring puts that slice here: 33%
Then there is assisted work, where AI speeds a task up but a person signs it off. Scheduling across several calendars is a good example: a tool can propose times, while you know that the vice president will not take a 4 p.m. on a Friday. Travel and expense work is similar, since the booking is easy and the exception handling is not. The assisted share sits here: 32%
The third group is work that still runs through a person: greeting and routing walk-in visitors, taking a phone call that turns into a complaint, keeping sensitive records in order, and coordinating a team that is behind schedule. The share of task time left with people is printed here: 35% The headline Can AI do it? figure, 45 out of 100, is built from that split, and how coverage is measured explains what counts as task time.
What the evidence does and does not show
There is no published head-to-head test of administrative assistants against an AI system doing the same office work. The Is it better than a person? grade for this job reflects that gap: D A D grade means the quality question has not been measured here, so we give no parity number rather than guess one. The letter grades and what each one requires are set out in how quality parity is graded.
A test that would settle it is easy to describe. Give assistants and an AI agent the same week: an inbox with real competing requests, two conflicting calendars, a trip to book with a change of plan halfway through, and a set of documents to prepare. Then let the managers who depend on that work grade the results blind. Until something like that is published, the honest position is a coverage estimate from task-level analysis, with the sources and weights shown on our methodology page and in the open dataset.
When the picture could change
Most likely between 2035 and 2047 (8 in 10 of our scenarios). The method behind that window, including what it does and does not claim, is set out in how the replacement year is estimated.
Two things could pull it earlier. The first is agents that act rather than suggest, with real permission to send mail, move meetings and file documents inside a company’s own systems. The second is price: the annual cost of tooling for these tasks sits far below the cost of a seat, and that gap is what makes employers consolidate two roles into one.
Two things slow it down. Over a quarter of the task time in this job is physical, and the only automation that handles it today is fixed automation, so mail, print rooms, supplies and front desks stay stubborn. Access and accountability are the other brake. Giving a tool the keys to an executive’s calendar, payroll files or a legal hold is a trust decision, and small employers rarely have the systems to make it cleanly.
How to stay needed in admin support
Lean into the parts of the job no tool is close to owning. Coordination under pressure, when three people want the same hour and someone has to be told no. Front-line contact, where a visitor or a caller needs a decision, not a menu. And record keeping where accuracy carries consequences, such as audit trails, confidentiality and retention rules.
Two skills raise your floor. The first is supervising and checking AI output: writing clear instructions, spotting a confident error in a draft or a transcript, and knowing what must never be pasted into an outside tool. The second is light operations work, such as building the scheduling workflow, the intake form or the shared tracker that the whole team then uses. Our guide to AI skills employers want covers how those show up in job ads.
What to do: pick one recurring task you own, document it, and automate the dull half yourself before someone else does it for you.
Nearby roles are worth a look, because the task mix changes a lot between them: Executive Secretaries and Executive Administrative Assistants, Legal Secretaries and Administrative Assistants and Medical Secretaries and Administrative Assistants. You can put any two of them side by side on the compare page, see the wider group on the secretaries and administrative assistants family page, or read how the whole administrative support sector scores. The Still needs a human score on this page is 61 out of 100 (higher is safer), and what that headline figure means explains how it is built.