ChatGPT Prompts for Virtual Assistants: Inbox, Scheduling, Reports

Three groups of prompts cover most virtual-assistant work: triage prompts that sort an inbox into reply, delegate, schedule and ignore; scheduling prompts that turn messy availability into proposed slots; and report prompts that turn your activity notes into a client update. Every output should pass a human approval step before it reaches a client.

That approval step is the part most prompt lists leave out. ChatGPT drafts quickly and confidently, and confident drafts are how a wrong date or a misread tone lands in a client's contact's inbox under your name. The eight prompts below each end with a "what I check before it goes out" note, and the final sections turn those notes into a routine you can run in two minutes.

Follow me on Instagram@sagnikteaches

Set the ground rules before the first prompt

You work inside someone else's business, so the rules come from the client, not from you. Before any of these prompts touch client material, settle three things in writing. First, whether the client agrees to you using an AI assistant at all, and for which tasks. Second, what stays out of ChatGPT whatever the task: passwords, bank details, medical or legal matters, anything the client has marked confidential. Third, who approves AI-assisted drafts, and how fast they will answer.

Connect on LinkedInSagnik Bhattacharya

A simple clause covers it: "I use an AI writing assistant to draft some emails, schedules and summaries. Nothing is sent on your behalf without your approval, and I never enter passwords, payment details or documents you mark confidential." If you want wording for a wider contract, how to mention AI use in client contracts has examples, and the confidentiality checklist for virtual assistants goes further on what to keep out.

Subscribe on YouTube@sagnikteaches

On the tool itself, the plan sets the default. OpenAI's enterprise privacy page says it does not train on business data by default across ChatGPT Business, Enterprise and the API. On the individual plans, its help centre says data sharing stays on until you switch off "Improve the model for everyone" under Settings, then Data controls. For a solo assistant the catch is the seat minimum: a Business workspace needs at least two seats, which OpenAI's help page prices at $25 per Standard seat a month, or $20 billed annually, as of November 2026. Many solo assistants therefore stay on Plus ($20 a month on the pricing page), switch the training setting off and use Temporary Chat for one-off client drafts. OpenAI says a temporary chat is not used to improve its models while it stays temporary, though a copy may be kept for up to 30 days for safety checks.

Placeholders: how to keep real names out of the prompt

Most of the risk disappears if you never paste real identifiers. Build the habit of swapping them for tokens. Instead of pasting the customer's real first name, the printer's trading name and the exact invoice total, write "[client first name], the invoice from [supplier] for [amount] is overdue". ChatGPT drafts around the tokens happily; you paste the real values back into the final text yourself.

Keep a small mapping note on your own machine: "[client] = the florist owner, [supplier] = the vase wholesaler". It takes a minute to set up per client, and it stops you pasting a real name when you are rushing at the end of a long day.

Inbox triage prompts

Prompt 1: sort a batch of subject lines into four piles

Triage works best on what you already know to be non-sensitive: subject lines and the first line of each message, with sender names replaced by roles. Paste ten of them and give ChatGPT the client's own rules, not generic ones.

You are helping a virtual assistant triage the inbox of an independent florist.
Rules: anything about a wedding or funeral order is URGENT (reply today).
Supplier invoices go to FINANCE. Newsletters and promotions go to IGNORE.
Everything else is REPLY THIS WEEK.
For each line below, return a table with: number, pile, one-line reason.
If you are not sure, put it in REPLY THIS WEEK and say why.
Lines:
1. [role: bride] "Change to Saturday's bouquet order"
2. [role: wholesaler] "Invoice 2291 overdue"
3. [role: software vendor] "Your plan renews soon"
...

A realistic output is a four-column table: line 1 URGENT ("order change, event this week"), line 2 FINANCE ("invoice, needs matching to the order"), line 3 IGNORE or FINANCE depending on whether you told it about software renewals. The flaw to watch for is over-confidence: it will file an ambiguous message such as "Quick question about Saturday" under REPLY THIS WEEK when "Saturday" might be a wedding. Check: open every item in URGENT and anything with a date in it yourself. The model sorted words; you are checking meaning.

Prompt 2: draft a holding reply you can send without approval

Some replies are low-risk and the client may pre-approve them: acknowledging receipt, confirming a booking exists, saying the owner will reply by a stated day. Agree that short list with the client, then use a tight prompt.

Draft a 2-sentence acknowledgement from the shop's assistant.
Facts I give you: the customer asked about [topic]; the owner replies within [two working days].
Do not promise prices, delivery dates or refunds. Do not add facts I haven't given you.
Tone: warm, plain, no exclamation marks.

Sample output: "Thanks for getting in touch about your order. I've passed your message to [owner first name], who will reply within two working days." Check: that the reply contains none of the facts you didn't provide. If ChatGPT adds "we'll make sure it's perfect", delete it; promises are the client's to make.

Prompt 3: summarise a long thread into a decision list

Long threads hide the one decision the client needs to make. With permission, and with names replaced, ask for the decision first and the history second.

Below is an email thread about a bookshop's event booking, with names replaced by roles.
Return: (1) the decision the owner must make, in one sentence;
(2) the options on the table with who proposed each;
(3) any date or amount mentioned, quoted exactly;
(4) anything that looks unresolved.
Do not infer anything that isn't written.

Here the key line is "quoted exactly". Models sometimes round or reformat dates and amounts when summarising. Check every number and date in section 3 against the original thread, because section 3 is the part the client will act on.

Scheduling prompts

Prompt 4: turn tangled availability into three proposed slots

A familiar mess: the client is free "Tuesday afternoon unless the courier comes, or Thursday before 11", the guest suggests "late next week". ChatGPT is good at laying this out as options, and poor at knowing today's date and time zones, so give it both.

Today is [date]. Client time zone: [zone]. Guest time zone: [zone].
Meeting: 45 minutes, video call.
Client availability, in their words: "[paste]".
Guest availability, in their words: "[paste]".
List up to three slots that satisfy both, written in both time zones,
and flag any slot where the guest's local time is before 8am or after 6pm.
If nothing fits, say so and suggest what to ask the guest.

Sample output is three bullet lines, each with two times. The failure that actually happens: daylight-saving changes between the two zones, so a slot "9am for them" turns out to be an hour off in the calendar. Check: put each proposed slot into the client's real calendar as a tentative hold, and let the calendar's own time-zone display confirm it. Never send a time you haven't seen in the calendar.

If you want to go further and let software book on the client's behalf, can AI manage your calendar and book meetings covers what that involves and where it breaks.

Prompt 5: a week plan with the client's non-negotiables first

For a driving-school owner, a week has fixed points (lessons booked, a test day, a van service) and flexible ones (admin, quotes, social posts). Give ChatGPT the fixed points as a list and ask it only to arrange the flexible ones.

Fixed this week (do not move): Mon 9-5 lessons; Wed 2pm vehicle service (drop-off 1.15pm);
Thu all day: test day, no calls.
Flexible, with time needed: reply to 6 enquiries (45 min); update the price list (1 hr);
chase two unpaid invoices (30 min); plan next month's social posts (1.5 hr).
Working hours for admin: 7.30-9am and 5-6.30pm only.
Arrange the flexible items into those windows. If they don't all fit, say which to drop.

Expect a day-by-day list with a running total, and expect the total to need checking. An illustrative output: "These tasks total 3 hours 45 minutes and your windows hold 4 hours 15 minutes across the three non-test days, so the price list should be split." The first figure is right (45 + 60 + 30 + 90 minutes). The second is wrong twice: Monday, Tuesday, Wednesday and Friday are four non-test days, and two 90-minute windows a day give 12 hours, so nothing needs splitting. Check the arithmetic yourself, because models add up durations carelessly, and check that nothing sits in a protected window or on the test day.

Report prompts

Prompt 6: turn your activity log into a client update

Clients want to know what happened, what is waiting on them and what is next. Keep a plain running log as you work ("Mon: replied to 14 enquiries, booked 3 viewings; Tue: chased invoice 2291"), and let ChatGPT shape it.

Turn the activity log below into a weekly update for a photography studio owner.
Structure: three headed sections: Done, Waiting on you, Next week.
Under "Waiting on you", list only items that need a decision, each as a question.
Use only the log; add no numbers or claims that aren't in it.
Keep it under 200 words.
Log: [paste]

A good sample output: "Done: 14 enquiries answered (11 within two hours); 3 studio hire bookings confirmed. Waiting on you: shall I accept the 12 March hire request given the double booking? Next week: ..." Check: every number against your log, and the "Waiting on you" questions against what the client really needs to decide. A report that invents a metric, say "response rate improved", is worse than no report.

Prompt 7: monthly report with a variance explained

If you track simple counts, such as enquiries, bookings and invoices chased, ask ChatGPT to compare two months, and to separate what the numbers show from what it guesses.

Here are two months of counts for a bookshop's admin: [table].
Write a five-sentence summary. Put anything not directly shown by the numbers
in a separate section headed "Possible reasons (unverified)".
Do not calculate percentages; I will add them myself.

Asking it not to calculate percentages is deliberate. You do the sums in a spreadsheet and paste them back, so a wrong figure can only be your error, which you can catch. The "unverified" section is where you decide with the client whether any reason is worth saying aloud.

Prompt 8: a polite chase for a late client reply

Not a report as such, but the end of every report: the item waiting on the client. Ask for three versions of increasing firmness, then send the one that suits the relationship.

Write three versions of a follow-up asking the client to approve the attached quote.
Version A: friendly nudge. Version B: states the deadline. Version C: says work will pause.
Each under 60 words. Do not invent a deadline; use [date].

Check version C against your service agreement. If your contract does not say work pauses, you cannot promise it, however firm the draft sounds. Version A usually wins for a first chase; keep B and C in a note for the second and third.

The approval step: a two-minute routine before anything goes out

Everything above ends in the same routine, so make it a habit rather than a decision each time.

1. Sort by risk. Pre-approved holding replies go out; everything else is "needs approval".
2. Verify facts. Read every date, number, name and promise against the source. This takes thirty seconds and catches nearly everything.
3. Restore the real details. Replace placeholders by hand, and re-read once for sense after you do.
4. Send to the client for approval in the channel you agreed, with the draft and your one-line note ("changed the Thursday slot because of the courier").
5. Log it. Keep a line in your own records: what was drafted with AI, who approved, when.

The log matters more than it sounds. If a client ever asks "who wrote that?", you can say exactly: drafted with an AI assistant, checked by me, approved by you on the date shown. That answer protects the relationship.

Monday morning in a florist's inbox, counted

You look after a florist's inbox for ninety minutes each morning, with 38 new messages overnight. Using Prompt 1 on subject lines, 4 land in URGENT, 6 in FINANCE, 15 in IGNORE and 13 in REPLY THIS WEEK. You open all four urgent ones yourself: one turns out to be a complaint, not an order change, which the model had misjudged from its subject line. That alone justifies the check.

You then use Prompt 2 for three straightforward enquiries (pre-approved holding replies, sent directly) and Prompt 3 on one long wedding thread, producing a decision list for the owner. Time spent: about 40 minutes instead of the usual 75, with the saving coming from sorting and first drafts, not from skipping reading. The report on Friday uses Prompt 6 on your log and takes ten minutes.

The numbers are invented to show the shape of the work; yours will vary. The point is where the time moves: less typing, the same amount of reading and checking.

One rescheduling request, by hand and with Prompt 4

Take a message from a guest: "Can we move Thursday? Something's come up. Next week any time apart from Monday would be fine." Without AI, you read the client's calendar, work out three options, type a reply and check the dates. With Prompt 4 you paste the client's free windows and the guest's constraint, and get three candidate slots in a minute.

Before: a 12-minute job with a risk of typing the wrong weekday. After: a 4-minute job where the drafted reply reads "Thanks for letting me know. [Client first name] could do Tuesday at 2pm, Wednesday at 10am or Friday at 11am next week. Which suits you best?" Your check is that Tuesday is actually a Tuesday in the calendar and that Wednesday at 10am does not clash with the standing meeting. The draft is quicker, but the calendar still has the last word.

The summary that credited the wrong person with a yes

A common slip: you ask ChatGPT to summarise a thread and it merges two people with similar roles, so the summary says the bookshop's events manager approved a quote when it was the accountant who queried it. Nothing looks wrong in the prose. It surfaces only when the client reads "you approved" and replies "I never did".

The fix is built into Prompt 3: ask for options "with who proposed each" and quote exact lines. If a name or role appears in the output that you did not expect, go back to the source message before the summary goes anywhere. Treat every attribution ("she agreed", "he asked for") as a claim to be verified.

Building the prompts into a per-client kit

Eight prompts are a start, but the real time saving comes when each client gets their own saved version. Put the client's rules at the top of a document, such as the four inbox piles, the pre-approved reply list, their working hours and the tone notes, and paste that header above any prompt. If you use ChatGPT Projects, which OpenAI's pricing page lists on every individual plan from Free to Pro, one project per client keeps those instructions attached to the work without retyping them, and keeps one client's material out of another client's chats.

A filled-in header for the bookshop might read: "Owner prefers short replies, signs off with first name only, never discounts by email, replies to event enquiries within one working day, and does not want anything sent on Sundays." Four lines like that remove most of the editing you would otherwise do by hand. Review the header every quarter with the client; businesses change their minds about discounts and opening hours faster than anyone updates a document.

Timing the checks: what a sensible week costs

Illustrative sums for one client on a ten-hour weekly retainer, where roughly four hours are inbox and scheduling work. Triage with Prompt 1 takes about five minutes a day instead of twenty; first drafts for ten replies take twelve minutes instead of thirty-five; the Friday update takes ten minutes instead of twenty-five. That is nearer two hours saved than four, because reading, checking and approval do not shrink. If a client expects the whole four hours to vanish, say plainly that checking is the job you are selling.

Track it for two weeks before promising anything. Note the start and end time of each task with and without the prompt on a sample, then keep only the prompts that save at least a third of the time. A prompt that saves two minutes but needs five minutes of fixing is not worth keeping.

Handing over cleanly when the client says no to a draft

Sooner or later a client rejects a draft: too formal, wrong promise, or simply not what they would say. Treat each rejection as data. Copy the client's corrected version into your per-client header as a "good example", and add a one-line rule to the header, such as "never say 'kindly'" or "always offer two dates, not one". After three or four rounds the drafts need far fewer edits, and you have a written record of the client's voice that a colleague or cover assistant can use during your holiday.

That record is also your handover document. If you ever pass a client to another assistant, the header, the pre-approved reply list and the approval log tell them what the client accepts, what they refuse and how quickly they answer. Keep it free of passwords and account numbers, and share it only through whatever channel the client has approved.

When a prompt should not be used at all

• Anything legal, medical or financial in content. Summarising a contract dispute is not a prompt, it is a decision for the client and their adviser.
• Complaints with emotional weight. A model-drafted apology tends to read as scripted. Draft by hand, or ask ChatGPT only to suggest a structure.
• Personal data about third parties. Customer lists, addresses and payment details stay out.
• Anything the client has told you not to run through AI. Respect it without argument.

Prompts also drift. Keep your eight in one document, date each when you last tested it, and rerun it on a sample once a month; ChatGPT's models and defaults change, and a prompt that behaved last quarter may now add polite padding or reorder your table. If you sell this skill to clients, pricing for virtual assistants when AI does the drafting explains how to charge for the checking rather than the typing. To see how these prompts sit alongside other tools, read the five-client tool stack, and for inbox rules beyond prompts see AI email management rules, drafts and handover.

Questions virtual assistants ask before using ChatGPT on client work

Do I need my client's permission to use ChatGPT on their inbox?

Yes, ask first and get it in writing. Clients often have confidentiality clauses or simply expect a human to read their mail. A short line in your service agreement naming the tool, what goes into it and what never does keeps both sides clear and gives you something to point to later.

Should I paste a whole email thread into ChatGPT?

Only if the client has agreed and the thread holds nothing sensitive. Better habit: paste the single message you need help with, replace names, amounts and account details with placeholders, then put the real details back yourself before you send anything.

Which ChatGPT plan should a virtual assistant use?

Most solo assistants start on Plus with the model-improvement switch turned off, because a ChatGPT Business workspace needs at least two seats. Business plans do not train on workspace content by default, which matters if your clients have strict confidentiality terms. Compare both against OpenAI's current pricing page before you commit, since prices and seat rules change.

How do I stop ChatGPT sounding like a robot in client emails?

Give it two or three real emails the client has sent as style samples, then ask for a draft in that voice. Always read the result aloud once. If a sentence is one the client would never say, rewrite it by hand before it goes anywhere near the approval queue.

Further reads

• Questions to Ask Before Buying AI That Touches Client Data — Questions to ask any AI tool that touches client data.
• Virtual Assistant or AI for a Real Estate Agent: What Goes Where — A worked split of tasks between a human assistant and AI.
• AI Appointment Scheduling for Small Businesses: Tools and Setup — Booking tools once prompts are not enough.
• How Photographers Can Handle Client Emails With AI — Client email drafting seen from a studio owner's side.
• How to Stop AI Tools Training on Your Business Data — Switches to check so client content is not used for training.
• Should You Charge Clients Less When AI Speeds Up Your Work? — Whether faster drafting should lower your fee.
• AI Virtual Assistant Services: How a VA Business Can Package Them — Three package tiers, the pricing sums, scope limits and disclosure wording for a VA business selling AI-assisted work.
• Chat Support Virtual Assistant Software: Which Fits a VA Business? — Which shared inbox or live chat tool suits a VA serving several client shops, what each really costs at eight clients, and whether AI drafts pay.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI enterprise privacy page, ChatGPT pricing page, and OpenAI help articles on ChatGPT Business, data controls, temporary chat and projects, checked 4 November 2026.