Inbox, Travel, and Documents: The Daily Machine

Inbox management is the highest-volume task in the role, and the thing being automated is not sorting. It is a model of what one specific person needs to see.

A useful triage produces four outcomes. Needs the executive, now. Needs the executive, not urgently. Handle without them. And nothing required.

What determines which bucket a message lands in is almost never the message alone. It is the sender's relationship to the executive, what is happening this week, what the executive already knows, and whether this thread has history. The same message from two different senders belongs in two different buckets.

Generic importance filtering fails for exactly this reason. It sorts on properties of the email rather than on properties of the relationship.

What does work is calibration from your own decisions. If the system can learn from how you have historically routed messages from particular senders and topics, it becomes useful, because you are the model being approximated.

And the practical arrangement is a queue rather than an action. The system proposes a bucket, you confirm or move it, and the corrections are the training signal.

One asymmetry to design around. The costs of the two errors are wildly different. Wrongly escalating something unimportant wastes a minute of the executive's attention. Wrongly suppressing something urgent can cost a deal or a relationship. So triage should be biased toward escalation, and any automated suppression needs to be conservative in a way that feels inefficient and is correct.

One artefact is worth building deliberately, because it changes how an executive uses their day more than any other output an assistant produces.

The daily brief answers what an executive needs to know before they start. What is on today and what each thing is for. Who they are meeting and what they should know about them. What changed since yesterday in anything they are tracking. What is waiting on them and how long it has waited. And what will become urgent if it is not addressed today.

Most of that is assembly from sources that already exist: the calendar, the inbox, documents circulated, the open-action list from previous meetings. It is precisely the kind of aggregation nobody has time to do manually every morning, which is why executives generally start their day without one.

What a good brief needs beyond assembly.

Editing for relevance. A complete list is not a brief. The value is in the selection, and selection requires knowing what this person cares about.

The outstanding-longest item, which is the single most useful line in it, because things fail by waiting rather than by being rejected.

And context on people. Who this person is, when they last met, what was agreed, what they want. Assembled from your own records, verified before it goes.

The caution. Anything about an external person compiled from public sources needs care, both for accuracy and for proportion. A briefing note is not a dossier, and material about someone's personal life does not belong in it.

The path an inbound message should take, with the automation boundary marked.

It arrives and is classified against what you have taught the system: sender relationship, topic, thread history, and what is live this week.

Routine and answerable from existing information produces a drafted reply, which you review and send. This is most of the volume and almost none of the difficulty.

Needs the executive routes to the brief or to the queue, depending on urgency, with the context attached so they are not opening a thread cold.

Needs a decision you can make gets your judgement, then a drafted response.

And unclear routes to you, always. That default matters more than any classification accuracy, because the messages a system cannot categorise are disproportionately the unusual ones, and unusual is correlated with important.

The boundary to notice. Nothing leaves without a person reading it, and nothing is suppressed automatically. The system is sorting and drafting; the assistant is deciding and sending.

That is deliberately conservative. In a role built on relationships, the cost of an automated message landing wrongly with a senior person is out of proportion to the minutes saved by sending it unread.

Travel arrangement divides sharply into a booking problem, which is largely solved, and a disruption problem, which is where an assistant earns their reputation.

The booking side. Comparing options, assembling an itinerary from scattered confirmations, checking visa and documentation requirements, building the schedule around meetings rather than around flights. Genuinely helped by tooling, particularly the assembly step, since a trip generates confirmations from six providers in six formats and turning them into one coherent document is pure extraction.

One caution on the booking side. Requirements for visas, entry rules and documentation change, and a model's recollection of them is exactly the kind of confident, outdated, consequential answer this catalogue keeps warning about. Someone stranded at a border because a generated answer was a year out of date is a serious failure. Verify against the official source, every time.

The disruption side is different work entirely. A cancelled flight is not a rebooking problem. It is a question about which commitments to protect: which meeting cannot move, who must be told immediately, whether the whole trip is now worth taking, and what the executive would want if you could ask them.

What helps here. Knowing immediately, through monitoring rather than through the executive discovering it. Having the options assembled before you make the call. And having drafted notifications ready for each scenario.

What does not delegate. The choice. It is a judgement about priorities that depends on things no system knows, and it usually has to be made in ten minutes.

The useful framing: automate the watching and the option-gathering, keep the deciding.

Assistants produce a steady volume of documents, and the division of labour is the same one that appears throughout this catalogue.

What works well. Formatting and restructuring content that exists. Producing a presentation from a document, or a document from notes. Adapting one piece of content for several audiences. Proofreading. Summarising long material for someone who will not read it. Turning scattered updates into a coherent report.

What needs care. Anything assembled from multiple sources, where the failure is a detail attributed to the wrong source or a figure carried across incorrectly.

And the hard rule, which matters here because assistants frequently handle documents containing numbers they did not produce. A language model does not compute. Any figure in a board paper, a budget summary, an expense report or a forecast comes from the system that calculated it, and the model's job is presentation. A generated number in a document that goes to a board is a serious error with the assistant's fingerprints on it, and it is invisible on the page.

The safe pattern is transcription rather than calculation: moving a figure from a source into a document, then checking it against that source. Even summing a column is better done in the spreadsheet.

One further point specific to this role. Assistants often produce the final version of something an executive will present. An error there is discovered in public, by the person who trusted you with it. That raises the value of a slow final check above almost anything else in the workflow, and it is the one place where speed is not the objective.

A reframing that changes what an assistant should optimise for, and it runs against how the role is often measured.

The implicit objective in much assistant tooling is throughput: more meetings scheduled, faster responses, a fuller calendar. That treats an executive's time as capacity to be filled.

The more useful objective is protecting attention. An executive's scarce resource is not hours; it is uninterrupted thinking time and the capacity to be present in the meetings that matter. A calendar with no gaps is a failure state that looks like efficiency.

What that implies practically, and it is mostly about declining things.

Defending blocks of unscheduled time as though they were meetings, because they are the only slot where anything considered gets done.

Batching the interruptible. Questions that do not need an immediate answer are collected rather than delivered on arrival.

Asking what a meeting is for, and whether it needs this executive or a decision from them by email.

And making the cost of a meeting visible. An hour with six senior people is six hours, and almost nobody counts it that way.

Where tooling helps with this. Analysing where time actually goes, which most executives are wrong about. Identifying recurring meetings nobody has questioned. Drafting the declines, which is the part people avoid because it is socially uncomfortable.

And the honest note. An assistant who fills the calendar looks busy and productive. An assistant who protects two hours a day of thinking time is doing something harder and more valuable, and it is less visible, which is why it needs to be an explicit agreement with the executive rather than a private policy.

The structural question, and the answer is more encouraging than for most administrative roles.

What gets compressed is the mechanical layer: finding times, assembling itineraries, writing routine replies, formatting documents, taking notes, filing, chasing. That is a large share of the hours in the role and almost none of what distinguishes a good assistant from a poor one.

What does not compress is everything the first lesson identified: priority, access, relationships, discretion, knowing what the executive would want, and holding the informal knowledge of how an organisation actually works.

The direction that suggests is toward what is sometimes called a chief of staff function: less coordination, more judgement, more actual ownership of things. An assistant who spends four hours a day on logistics and two on judgement can invert that ratio, and the second kind of work is both more valuable to the organisation and harder to replace.

The risk in the transition is real and worth naming. Organisations that see only the mechanical layer will conclude the role can be cut, because that layer is what appears in the job description. The defence is making the discretionary work visible: the meetings prevented, the problems caught before they surfaced, the commitments tracked to completion, the relationships maintained.

That is difficult to evidence, because good assistance is largely invisible by design. Things not going wrong is the output.

Which is a practical reason to keep a record of what you handled, decided, and prevented. Not for its own sake, but because in a period when the visible part of the role is being automated, the invisible part needs an advocate.

What an assistant can hold in their head, drawn from both lessons.

What I automate. Candidate times, never the choice. Itinerary assembly from confirmations. Meeting transcription and action extraction, with consent. Briefing assembly from documents that exist. Filing, expenses, chasing. Draft replies to routine correspondence.

What I never let a tool decide. Who gets access to the executive's time. What the executive sees and what is handled without them. Who is in a room. How firmly someone is declined.

What never goes out unread. Anything at all. In a role built on relationships, nothing reaches a recipient without a person reading it.

What never goes into an external tool. Pre-announcement material, personnel matters, legal matters, the executive's personal information, and the calendar pattern where it would reveal something. When unsure, assume it is sensitive.

Numbers. From the system that computed them, transcribed and checked. Never generated.

Facts that strand people. Visa and entry requirements verified against the official source, every time, regardless of how confident the answer looks.

Triage bias. Toward escalation. Wrongly escalating costs a minute; wrongly suppressing can cost a relationship.

And the objective. Protect attention rather than fill time, which needs to be an explicit agreement with the executive rather than something you do quietly.

Eight lines, and the whole of it reduces to one idea from lesson one: automate the mechanics of decisions you have made, and keep the deciding.