AI Agents for Administrative Assistants: What They Take Over
An administrative assistant costs a US business a median of $47,540 a year in wages — about $67,982 once benefits and payroll taxes are counted. It is one of the most common desk jobs in the country, and the largest share of it is moving information that already exists from one system into another.
The work that repeats
These are the parts of the role that run the same way every time. It is a short list on purpose — it is the only kind of work an agent handles well:
- Re-typing information that already exists in another system
- Maintaining spreadsheets that feed other spreadsheets
- Routing incoming requests to the right person
- Filing and naming documents by hand
- Sending the same three template emails all day
What stays with the administrative assistant
Anything requiring a decision about priority between two people who both think theirs is urgent. Also the informal knowledge of how this specific office actually works, which is almost never documented anywhere.
What we would actually build
An admin agent that watches the shared inbox and the intake form, routes each request by its content, files the attachment where it belongs with a consistent name, and sends the right template — logging every action so nothing is invisible.
It runs inside Gmail, Google Drive, Airtable, Slack, Google Sheets — the tools already in use.
Where we would start
Request routing from the shared inbox is the first thing worth building. Routing is where shared inboxes fail, it is easy to audit because every action is logged, and it removes the interruption cost of triaging by hand.
How this one goes wrong
Mis-routing is the failure that hurts, because it is invisible. A request goes to the wrong queue, the wrong queue does not recognise it as theirs, and it sits — and unlike a dropped email there is no sender wondering why nobody replied, because the sender can see it was received. Two weeks later someone asks what happened to it. The guard is that every routing decision is logged with the reason it was made, and anything the agent could not route with confidence surfaces on a daily list rather than being assigned to the nearest plausible team.
The payback arithmetic
Slack’s Workforce Lab surveyed 10,281 desk workers across six countries and found they spend 41% of their time on work that is “low value, repetitive or lacks meaningful contribution to their core job functions” (Slack Workforce Lab, 2024). For an administrative assistant that is about $27,873 a year in loaded cost. You will not recover all of it.
Being straight about this one: a single administrative assistant does not justify the engagement on labour savings alone. BreathingRoom is $30,000 a year and the entire low-value share for one administrative assistant is $27,873 — so the arithmetic only works from roughly 1.1 people doing this work, or when the agents cover this role alongside others. If this is the only repetitive work in your business, say so on the call and we will tell you it is too early.
| Line | Annual |
|---|---|
| Median wage, Secretaries and Administrative Assistants, Except Legal, Medical, and Executive | $47,540 |
| Benefits and payroll taxes (1.43×, BLS ECEC) | $20,442 |
| Compensation cost | $67,982 |
| Value of the low-value, repetitive share | $27,873 |
| BreathingRoom, annual | $30,000 |
What the week looks like afterwards
The re-typing stops, which is a bigger change than it sounds — the spreadsheets that feed other spreadsheets update themselves, and the version that is on someone's desktop stops being the one that matters. Documents get filed with the same naming convention every time, so finding one stops being a search. The three template emails send themselves. What remains is the work that needed a person in the first place: the request that does not fit any category, the person who needs an actual answer, and keeping the whole thing coherent.
Where this role sits
Administrative Assistants carry a large share of the repetitive load in staffing & recruiting agencies, property management, insurance brokerages, managed service providers, mortgage brokerage and marketing agencies, where the constraint each one runs into is set out alongside what to build first.
Job by job, the work described above is covered in more detail under admin & contracts — what triggers each one, and the case where an agent has to stop and ask.
Most of that work belongs to meeting follow-up, expense management and knowledge management. Each of those is written up on its own — what the loop costs when it slips, what we would automate in it first, and how an agent there goes wrong.
Common questions
What does an administrative assistant cost in the US?
The national median annual wage for Secretaries and Administrative Assistants, Except Legal, Medical, and Executive (SOC 43-6014) is $47,540, or $22.86 an hour, across 1,706,790 people employed. Adding benefits and payroll taxes at the BLS private-industry ratio of 1.43× puts the compensation cost at roughly $67,982 before software, workspace or management time.
Does an AI agent replace an administrative assistant?
No. Anything requiring a decision about priority between two people who both think theirs is urgent. Also the informal knowledge of how this specific office actually works, which is almost never documented anywhere.
What would you build first for an administrative assistant?
Request routing from the shared inbox is the first thing worth building. Routing is where shared inboxes fail, it is easy to audit because every action is logged, and it removes the interruption cost of triaging by hand.
How does an administrative assistant agent go wrong?
Mis-routing is the failure that hurts, because it is invisible. A request goes to the wrong queue, the wrong queue does not recognise it as theirs, and it sits — and unlike a dropped email there is no sender wondering why nobody replied, because the sender can see it was received. Two weeks later someone asks what happened to it. The guard is that every routing decision is logged with the reason it was made, and anything the agent could not route with confidence surfaces on a daily list rather than being assigned to the nearest plausible team.