An AI Executive Assistant for a CEO That Lives in Your Chat App

An AI executive assistant for a CEO works best when it lives where the CEO already types: the chat app on their phone. You send it a message, a voice note, a photo of a whiteboard or a PDF. It reads the input, picks the right tool (calendar, task list, notes, email drafts), acts, and replies in the same thread. No new dashboard to open, no login to remember.

We run one for our own CEO, and this post covers what it does step by step, the night it went quiet, and what it costs to install as a fixed-scope build. It is a personal assistant for one person, which is a different job from the shared knowledge assistant we covered in our post on internal AI knowledge assistants.

The job: a leader’s day is a stream of small requests

An executive assistant exists because leaders produce a constant trickle of small asks: move that meeting, remind me Thursday, what did I promise Dana, draft a reply to this supplier. None is hard. Together they eat the day.

Two public data points frame the problem. Microsoft’s 2025 Work Trend Index, published in June 2025 from about 31,000 knowledge workers plus Microsoft 365 telemetry, found employees are interrupted every two minutes during core hours by meetings, emails or chats. Microsoft notes that figure reflects the top 20% of users by ping volume, which is roughly where a CEO sits.

Prialto’s 2025 Executive Productivity Report, a survey of 508 leaders, found executives spend 20 to 30% of their time on administrative tasks. Prialto sells virtual assistants, so read it as a vendor survey with an interest in the answer. The direction still matches what we see.

Why a chat app, and not another tool

A leader will not open a new application at 10 p.m. to log a thought. They will send a message. The messaging platform we build on documents bots that receive text, voice messages, photos and general files as ordinary message types, according to Telegram’s Bot API documentation (version 10.3, August 2026). That means one input channel covers everything a leader actually produces: a typed line, a spoken note while walking, a snapped receipt, a forwarded document.

Putting the assistant in a channel the person already lives in removes the adoption problem. Nobody trains on it. The only rule is that it accepts messages from exactly one sender, the owner, and ignores everyone else.

What the agent does, step by step

Here is the loop we run, described generically.

1. Receive. A message arrives in the chat. The system checks the sender against an allowlist of one, then drops duplicates.
2. Normalise. Voice notes are transcribed. Photos and documents are attached as context. Everything becomes text plus files.
3. Decide. A language model reads the message with the assistant’s standing rules and its memory of past notes, then chooses from a fixed set of tools. It cannot invent a tool.
4. Act. Calendar changes happen automatically and come back with an undo button. A business task goes into the task backbone with an owner. A personal reminder goes into a private list that never touches the company system. A “remember this” goes into long-term memory.
5. Reply. The answer lands in the same chat, short, with what it did and what it did not do.
6. Report. Every run is logged, so we can see what happened when something looks wrong.

Beyond the on-demand loop, it runs scheduled jobs: a morning brief that pulls the day’s calendar, the tasks waiting on the CEO and anything overdue, and an inbox pass that sorts and labels mail and prepares reply drafts.

Request	What the agent does	Who has the final say
”Move Thursday’s call to Friday”	Edits the calendar, offers undo	Owner can reverse in one tap
Voice note about a supplier issue	Transcribes, files a task with an owner	Task owner confirms in the task system
Photo of a handwritten list	Reads it, splits into tasks or reminders	Owner sees the result in the reply
Inbound email needing an answer	Prepares a draft reply	Owner taps to send; nothing sends alone
”Delete the 4 p.m. meeting”	Asks for explicit confirmation	Owner

Where it stops

This section matters more than the feature list. As of October 2026 the assistant has these hard limits, set by design:

• It never deletes mail. It labels and archives at most.
• It never sends an email on its own. It drafts, and a human taps send.
• Deleting a calendar event always needs an explicit yes.
• Personal items stay out of the company task system entirely.
• It answers one person. It is not a team inbox.

Our status record for it also says plainly that the evidence is young. It was built and deployed in September 2026, and its manifest, as of late September, reported it as reporting runs but not yet proven across a long stretch or across a failed run with real error text. We would rather say that than call it battle-tested. Our chief-of-staff agent post covers what a mature version of the surrounding oversight looks like.

This lines up with the trust gap in Prialto’s data. Its 2025 survey found AI use among executives rose from 51% in 2023 to 78% in 2025, while only 26% said they fully trust these tools. A hard-limits list is how you close that gap in practice: the assistant earns trust one boundary at a time.

The night it went quiet

One honest incident belongs here. In October 2026 the assistant went silent overnight. Messages sent late in the evening sat unanswered until the next morning.

The cause was not the model. The primary worker ran on a machine that went to sleep, and the cloud fallback that should have picked up the slack never ran. Its sweeper only looked back a couple of hours, so late messages aged out of the query and were abandoned. Worse, the old dispatch path used a request that returns “accepted” no matter what happens afterward, so the failure looked like success in the logs.

The fix, merged in October 2026, widened the lookback to a full day and replaced the fire-and-forget request with a direct, inline drain that either replies or fails visibly. We have not yet confirmed the fix end to end with a real overnight message, so we are not claiming it is solved. We wrote up the general failure pattern in our post on what breaks in production agents.

The lesson for any chat-based assistant: a “202 accepted” is not a reply. Check that a person got an answer.

What changed

The change is smaller than a headline would like. The CEO stopped switching between a calendar, a task tool and a notes app for small items. One thread takes all of it. Tasks that used to live in memory or on a sticky note now carry an owner and a date.

We will not quote a time-saved figure. We have not measured one we would defend, and a made-up number is worse than none.

What it costs as a Sprint

A personal chat-based assistant is a standard AI Workforce Sprint: from $9,500 fixed, four to six weeks, built on your accounts and handed over with documentation and a recorded walkthrough. The agent is a tomte, our word for one production agent with one defined job.

Three things move the scope. The first is which tools it touches: calendar, mail, your task system. Each integration adds work and each needs its own permission rules. The second is the standing rules, meaning what it may do alone and what needs a tap. The third is where it runs, since a fallback host (the lesson of the night above) is part of a sound design.

The written scope fixes the price before you commit. You own the delivered code, data and documentation, and you can revoke our access whenever you like. The same fleet pattern is on the agent fleet proof page.

If your day is mostly small requests that never make it into a system, send us a list of the ten you handled last week. Start async and a written reply comes back within one business day, with no meetings at any stage.