Executive operations
How to build an AI executive assistant, in seven steps
Build an AI executive assistant around one routine, not a personality. A morning briefing or a list of promised follow-ups is the right first job. Give it read access only to what that routine needs, make it show the source for every item, hold every message as a draft, and leave every decision with the owner.
Step 1: Which routine is eating your day?
Many owners of small Indian businesses run the day from three places: a calendar, an email inbox and a phone full of WhatsApp chats. The repeated work sits between them. What changed overnight? Who was promised a reply yesterday? What do I need before the eleven o’clock meeting? Pick one of these and write what a useful result looks like.
| Routine | Useful first output | Decision that stays with you |
|---|---|---|
| Morning briefing | Changes, due items and open decisions, each with a source link. | What matters most today. |
| Meeting preparation | Who is attending, past notes, open questions and documents. | What to say or commit to. |
| Follow-up tracking | Promised replies, with drafts held for approval. | Whether and when to send. |
| End-of-day close | Done, not done, and proposed carry-overs. | Who owns what tomorrow. |
Brand Wisdom’s CEO Assistant is scoped around these same four routines. Use them as an outline and keep only the ones that fit how you work.
Step 2: Which sources may it read?
List the sources the routine needs: one calendar, a task list, selected email labels, a notes folder. For each, record what the assistant may read, whether it may change anything, who grants access and how you take it back. Do not connect a whole account when the job needs a slice of it.
Permissions can be narrow. Google’s Calendar API authorisation guide tells developers to “choose the most narrowly focused scope possible”, and separates read-only, free/busy and event-editing access. A briefing needs read-only. Ask why if a builder wants more.
Step 3: What about WhatsApp?
If WhatsApp is part of the routine you want help with, this step needs the most care. A personal or business WhatsApp often holds clients, staff and family in one place. Tools that read or send from an ordinary WhatsApp account by automating the app or WhatsApp Web are a risk to the number itself: WhatsApp’s Terms of Service list “bulk messaging, auto-messaging, auto-dialing” among impermissible uses, and prohibit accessing the service through automated means it has not authorised.
Two safer patterns work. Keep WhatsApp out of the first build, and note promised replies yourself in the task list the assistant reads. Or, for business conversations, move them to a number on the WhatsApp Business Platform, which is built to connect to other systems through an authorised API. Leave personal chats out entirely.
Step 4: What personal information stays out?
An executive assistant can touch private correspondence, staff matters, customer records and finances. Sort each source by sensitivity before you connect it, and leave out anything the routine does not need.
Under the Digital Personal Data Protection Act, 2023, once section 8 commences, your business will be responsible for personal data processed on its behalf, including by a software vendor, which it should engage only under a valid contract, and it must keep reasonable security safeguards. Clients’ and staff details flowing into an assistant will fall within that, so prepare now: read the provider’s data terms before using live records, and test with sample or redacted material first. A prompt that says “keep this private” is not an access-control plan.
Step 5: What are its rules?
Separate preparation from authority. The assistant collects context and suggests next steps; you decide what matters and what leaves the business. Write the rules down:
- Use only approved sources for dates, commitments, names and figures.
- Show the source beside every briefing item.
- Draft messages for review; never send in the owner’s name by default.
- Do not change priorities, assign work or make a commercial commitment.
- Mark conflicting records as a conflict instead of choosing one.
- Escalate missing access, complaints and sensitive matters.
Keep facts and suggestions visibly apart. A calendar entry supports a meeting time; an old message may not prove a promise still stands. The briefing should label what is quoted, what is inferred and what is missing.
| Briefing field | Rule |
|---|---|
| What changed | One plain sentence. |
| Source | A link to the message, event, task or document. |
| Due or promised date | Only when the source contains one. |
| Decision needed | Phrased as a choice, without choosing. |
| Gap | Missing or conflicting information, named directly. |
Here is what a short briefing in that format can look like. It is an illustration: the names, times and details are invented.
| What changed | Source | Decision needed |
|---|---|---|
| The supplier meeting moved from 11:00 to 12:30. | Calendar invite, updated last night. | Keep the 12:30 slot, or ask to move it? |
| A client asked for a revised quote by Thursday. | Email thread with the client. | Who prepares it, and by when? |
| Two task entries give different dates for the site visit. | Task list, two records. | Which date is right? Marked as a conflict. |
Decide where each output lives. A briefing can be a private document or a message to you. A follow-up draft should sit somewhere you can edit and approve it. A task summary should link back to its source record. Avoid creating a second system of record that someone then has to reconcile by hand.
Step 6: How do you test it?
Use sample records that look like your real week, without live customer data. Decide what the right output is before each test, then run them:
- A complete calendar entry, to check the source link is right.
- A task with no due date, to check the date stays unknown.
- Two records with different meeting times, to check the conflict is shown.
- A drafted follow-up, to check it is held, not sent.
- A revoked permission, to check the workflow stops cleanly.
- A failed action, to check it is reported as failed, not done.
“Prepared” and “sent” are different states. The assistant should not report a task complete until the system can confirm the action happened.
Step 7: How do you roll it out and measure it?
Begin with a private briefing or a draft queue that only you see. Each morning, mark items as useful, missing or wrong, and fix the source or rule behind each mistake. Add a second routine only when the first is trusted and shares the same sources and permissions. Keep a short change log of the rules and name who maintains it.
Before launch, note a baseline from your own calendar or task list: time spent gathering context, promised replies missed, meetings walked into unprepared. Measure the same things afterwards, and count your review time honestly. Do not claim time saved until you have measured it. The running cost includes setup, connections, cleaning up sources, upkeep and your approval time, not only a subscription.
QUESTIONS
Common questions
Can an AI executive assistant manage my WhatsApp?
Not safely through tools that automate an ordinary WhatsApp account; WhatsApp’s terms prohibit unauthorised automated messaging. Business conversations can be connected through the WhatsApp Business Platform. Keep personal chats out, and start with drafts you approve.
Should an AI assistant send emails in my name?
Start with drafts held for approval. Sending in your name is a separate permission that should be narrow, deliberate and tested, with an easy way to switch it off.
What data should an executive assistant be allowed to see?
Only the accounts and fields the chosen routine needs. Record the purpose, the permission level, who reviews it and how access is withdrawn, and plan for your business being responsible, under the DPDP Act duties scheduled for mid-May 2027, for personal data a vendor processes for you.
REFERENCES
Sources and further reading
Each link was opened and checked on .
