What an Assistant Actually Does, and Why It Is Not Scheduling

An executive assistant role is usually written as a list of tasks: manage the calendar, handle the inbox, book travel, prepare documents, organise meetings. Every item on that list sounds mechanical, and each one contains a decision that is not.

Managing a calendar is not fitting appointments into gaps. It is deciding what deserves the executive's time, in what order, and what does not happen. Two requests both fit on Tuesday; only one should get it.

Handling an inbox is not filing. It is deciding what the executive needs to see, what can be answered without them, what needs a holding reply, and what will become a problem if it waits.

Booking travel is logistics until a flight is cancelled, at which point it is a series of judgements about which commitments to protect.

Organising a meeting is coordination until you have to decide who is in the room, which is a political act.

So the role divides cleanly, and the division is what this cursus is built on.

The mechanical layer: finding times, comparing options, drafting standard replies, assembling documents, entering data. Genuinely automatable.

The discretionary layer: priority, access, tone with specific people, and knowing what the executive would want without asking. Not automatable, and it is where the value is.

The risk of tooling in this role is not that it fails at the first layer. It is that removing the first layer can look like removing the job, when the first layer was the visible part and the second was the point.

The function that most defines the role is the one least often written into the job description, and it explains why the position resists automation.

An executive's time is a scarce resource that many people want. The assistant decides, day to day, who gets access to it. That is real organisational power, and it is exercised through hundreds of small judgements: this meeting gets moved and that one does not, this person gets fifteen minutes today, this request gets a warm holding reply and that one gets a firm no.

Those judgements draw on knowledge that exists nowhere in writing. Which relationships are politically important this quarter. Who is difficult and needs handling carefully. Which internal request is genuinely urgent and which is someone managing their own anxiety. What the executive is actually worried about this week, which shapes what deserves their attention.

A scheduling tool given the same calendar makes none of those distinctions. It sees availability.

Which produces the central design principle for this cursus. Automation should handle the mechanics of a decision the assistant has made, not make the decision. Finding three viable times is mechanical. Choosing which of the two competing requests gets Tuesday morning is not.

And there is a second-order effect worth naming. When access is mediated by an automated system, people notice, and it changes how they experience the executive. Being told no by a person who explains is different from being declined by software, and the difference accumulates into how an organisation feels about its leadership.

Calendar coordination looks like a solved problem and keeps not being solved, which is worth understanding before adopting anything.

The visible constraints are easy: who is free when, time zones, room availability, meeting length. Software has handled these for years, and scheduling links solve the simple case well.

The constraints that break it are invisible to any system reading a calendar.

Recovery. Four consecutive difficult meetings is not the same as four meetings, and a good assistant will not book the fourth even though the slot is free.

Sequence. Some meetings must precede others because of what gets decided, and the calendar records no dependency.

Preparation. A slot is not free if the meeting after it requires two hours of reading.

Energy and pattern. Which executive does hard conversations in the morning, who is useless after a long flight, who needs a gap before speaking publicly.

Status. A meeting with a board member is not moved to accommodate an internal request, whatever the calendar says about priority.

And the personal. The thing that is not in the work calendar and is non-negotiable.

So the honest reading of scheduling automation. It solves the constraint satisfaction and leaves the constraints that matter unmodelled. Given three candidate slots that all technically work, an assistant picks the one that works for reasons the system cannot see.

Which again suggests the same pattern: let the tool propose, let the person choose.

Sorting the work by whether a tool should act or prepare.

On the automate side. Finding candidate times across calendars and time zones. Assembling a travel itinerary from confirmations. Extracting action items from a meeting recording. Drafting a standard reply. Compiling a briefing pack from documents that already exist. Filing and expense categorisation. Chasing outstanding responses.

On the prepare-and-let-a-person-decide side. Which request gets the scarce slot. What the executive sees and what is handled without them. Who is invited to a meeting. How firmly to decline someone. What gets moved when something breaks. And anything where the recipient's relationship with the executive is at stake.

The useful diagnostic is a question. If this output goes out wrong, is the cost an inconvenience or a relationship?

A wrongly categorised expense is an inconvenience. A curt decline to someone senior, or a meeting invitation that excluded the person who needed to be there, is a relationship, and relationships are what an assistant is actually managing.

So the boundary is not drawn by difficulty. Extracting structured actions from a two-hour recording is technically harder than deciding who gets Tuesday. The boundary is drawn by whether the output touches a person's standing.

A feature of this role that has no equivalent in most of the professions in this catalogue: the assistant frequently writes in the executive's name, and readers know it and accept it.

That convention is long established and it is not deceptive, because everyone understands how it works. What it does mean is that the assistant is producing text that carries someone else's authority, and getting the register wrong has consequences the assistant does not personally bear.

Where generation genuinely helps. Producing several versions of a message at different levels of warmth and firmness, so the assistant can pick the one that fits this recipient. Adapting a message the executive has already approved for a different audience. Drafting the routine correspondence that has no discretion in it at all.

Where it fails, specifically.

Register with a particular person. How the executive writes to a board member, a direct report they are unhappy with, and a peer at a competitor are three different registers, and the difference is relational rather than stylistic.

What is not said. Much executive communication is careful about what it does not commit to, and generated text drifts toward completeness, which can concede something nobody intended.

And firmness calibration. Generated declines tend to be either warmer or more abrupt than intended, and both are costly in different directions.

The practical arrangement. The tool drafts, the assistant calibrates, and anything going to someone whose relationship with the executive matters gets the assistant's full attention rather than a skim. Which is to say the tool speeds up the writing and not the judging.

Worth stating explicitly, because it is the reason confidentiality in this role is unusually serious.

An assistant sees, as a matter of routine, material that almost nobody else in an organisation sees. Board papers before they are circulated. Reorganisation plans. Individual compensation. Who is being performance managed and who is being promoted. Acquisition discussions. Legal matters. The executive's personal circumstances, health, and family situation. And the pattern of who the executive is meeting, which is itself sensitive: a series of meetings with an investment bank tells a story before any document does.

That last category is the one people forget. Calendar metadata reveals strategy even when no content is disclosed.

Which means the confidentiality analysis for tooling in this role is stricter than for most.

Pasting a board pack into a consumer tool is a disclosure of material non-public information, which for a listed company is a serious matter.

Uploading a calendar for analysis discloses the meeting pattern.

Processing correspondence about a personnel matter discloses information about an identifiable individual, engaging data protection obligations.

The practical positions. Use tools the organisation has procured with terms covering retention and training use, and confirm that with whoever owns the contract rather than assuming. Hold a hard category that does not go into any external tool: pre-announcement material, personnel matters, legal matters, and the executive's personal information. And be alert to inference: material that names nobody may still identify a person or a deal from context.

An assistant who is careless here can cause harm out of all proportion to the convenience gained.

The highest-value automation available in this role is the one that runs after a meeting rather than before it.

The traditional problem. Someone takes notes, badly, while also participating. Actions are captured inconsistently. The follow-up is written from memory two days later. Half the actions have no owner and no date. Nobody checks whether the actions from the previous meeting happened, so the same items recur for months.

With consent to record, that changes substantially. A transcript exists. Actions can be extracted with owners and dates. The follow-up is drafted from what was actually said. And the open-action list persists across meetings, so the recurring item becomes visible.

That last point is worth emphasising, because it is where the real value sits. The gain is not the twenty minutes writing the note. It is that commitments stop evaporating, which changes how an organisation executes.

The conditions. Consent from participants, obtained openly, and in some jurisdictions from everyone present. Recordings and transcripts are organisational records and belong in the organisation's systems under its retention policy. And the extracted actions are a draft: models misattribute commitments between speakers and drop the conditions attached to them, which is the specific failure to check for.

And a judgement that stays human. Some things said in a meeting should not be in a written record. A candid remark about a colleague, a half-formed idea, a piece of information shared in confidence. An assistant knows which parts of a conversation are minuted and which are not, and that discretion is a real part of the role.

An order that follows the cost of error rather than the size of the task.

First, meeting capture and follow-up, with consent. Largest gain, operates on your own record of your own meeting, and it improves the organisation's execution rather than just your workload.

Second, information assembly. Briefing packs from documents that already exist, background on people the executive is meeting, summaries of long reports they will not read in full. Internal, high value, and the failure mode is an inaccuracy you catch before it reaches them.

Third, the mechanical layers of scheduling and travel. Candidate times, itinerary assembly from confirmations, options comparison. The tool proposes, you choose.

Fourth, drafting routine correspondence, calibrated by you before it goes.

Fifth, the administrative tail: expenses, filing, chasing, data entry. Small individually, large in aggregate, and almost no risk.

And not without you in the loop: anything that decides who gets access, anything declining someone whose relationship matters, anything that reaches a recipient without you reading it, and anything touching the confidential categories.

The organising principle for the whole lesson. Automate the mechanics of decisions you have made. Do not automate the deciding. The mechanical layer is most of the hours and almost none of the value, which is a good thing: it means the compression lands where you would want it to.