ChatGPT Gmail Integration: Set Up Draft-Only Quote Replies

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for ChatGPT Gmail Integration: Set Up Draft-Only Quote Replies.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for ChatGPT Gmail Integration: Set Up Draft-Only Quote Replies.

To make ChatGPT draft a reply when a quote request lands in Gmail, connect Gmail in settings on a Plus, Pro or Business plan, then describe an event-triggered task in ChatGPT Work: when a new email's subject contains "Quote request", read it and draft a reply. Set Gmail to Allow read actions and tell the task never to send.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

The permission level is the real guardrail; the "never send" line in the task is the second layer. With Allow read actions, ChatGPT can read the incoming email without asking but must ask before it changes anything in your mailbox, and saving a draft or sending a message both count as changes. So the worst a confused task can do is ask you for approval, and the task pauses until you answer. Free and Go accounts cannot create event-triggered tasks at all, which is why the plan matters as much as the prompt.

A seven-person translation agency, used as an illustration throughout, receives 15 to 25 quote requests a week, roughly 70% through a form on its website and the rest as ordinary emails. Every request needs the same reply: thanks, a price from the agency's per-word rate card, a delivery date, and a question about anything missing. Writing each one from scratch took about 12 minutes.

If you only want to know whether ChatGPT can send email on your behalf at all, the tutorial on what ChatGPT can and cannot send from Gmail and Outlook answers that. Drafting is the narrower, safer job, and with quotes it is the part that eats the time.

Which ChatGPT plan can run a Gmail quote-request task?

Plus, Pro, Business, Enterprise and Edu plans can run an event-triggered Gmail task in ChatGPT Work, while Free and Go cannot. Event-triggered tasks run in ChatGPT Work and respond to supported Gmail, Slack or GitHub activity. For Gmail, the trigger is a new message, optionally filtered by sender or subject. The Business row's two-seat minimum is what OpenAI's pricing page shows, while its help centre says a new workspace may be able to start with one paid seat (both checked 20 November 2026). Go by the seat selector at sign-up: if it will not go below two, budget for two. Here is what each plan allows:

PlanPriceEvent-triggered Gmail task?Active tasks allowed
Free$0NoNot applicable
Go$8 a monthNoNot applicable
Plus$20 a monthYes5
ProFrom $100 a monthYes15
Business$25 per seat monthly, $20 billed annually (two-seat minimum on OpenAI's pricing page; a new workspace may start with one)Yes10
Enterprise, EduNot on the pricing page (Enterprise via sales)Yes, once an admin enables "Allow event-triggered scheduled tasks"15 (Enterprise), 10 (Edu)

Active task limits count scheduled and event-triggered tasks together, so if you already run weekly reminders (the tutorial on ChatGPT scheduled tasks for business reminders covers those), check you have a slot free. Two more limits apply: event-triggered tasks can run up to 30 times an hour and 720 times a day across all of them, and OpenAI may group several events into one run.

Create the task in ChatGPT on the web or a supported mobile app. The desktop app can show an existing event-triggered task but cannot create or edit its trigger conditions, which catches people out when they try to set it up there first.

How do you connect Gmail to ChatGPT so it cannot send?

Connect Gmail in ChatGPT's settings (under Apps, or Plugins in some versions) and set its permission to Allow read actions, which lets ChatGPT read mail but makes it ask before changing anything in your mailbox.

  1. Open your profile menu, then Settings, then Plugins (some versions and help pages still say Apps). Find Gmail, choose Connect, and sign in with the Google account that receives quote requests.
  2. If you have connected more than one Google account, check which one the quotes arrive in. ChatGPT supports several Google accounts on Plus, Pro, Business and Enterprise, and each connected account has its own permission setting.
  3. Select Gmail, open the connected account's ••• menu, then Settings, then Permissions. Choose Allow read actions and select Save.

The four levels, and what each means for this task:

Permission levelWhat OpenAI says it doesEffect on the quote task
Always askAsks before reading app information or making changesPauses every run just to read the email. Too strict to be useful.
Allow read actionsReads without asking, asks before making changesReads the request, writes the draft, then asks before saving it to Gmail. Sending would also need approval. Use this.
Allow low-risk actionsApproves low-risk actions automatically; higher-risk ones may need confirmation or be refusedOpenAI does not publish which Gmail actions count as low-risk, so you cannot be sure what runs unasked. Avoid for now.
Allow all actionsSupported actions run without approval prompts; OpenAI flags elevated riskCould send a price to a customer with nobody reading it. Never for this task.

OpenAI's own inbox starter prompt takes the same line: "Don't send, archive, or move anything to Trash without approval." It also lists sending an email among the actions that may require extra review. Changing a permission does not disconnect Gmail; it only changes what needs your say-so.

How do you make quote requests easy for the task to recognise?

A quote-request task is most reliable when the email has a predictable subject line, because the Gmail trigger can filter on sender or subject. The agency's website form sends a notification email for every submission. If yours does too, check whether the form tool lets you fix that notification's subject; the agency set it to "Quote request - website", and the task's condition became "subject contains Quote request". Every form request now matches, and almost nothing else does.

The other 30% of requests arrive as ordinary emails with subjects like "Translation needed", "Can you help?" or "Price for handbook". There are two ways to catch them:

  • Trigger on every new message and let the prompt decide. The task reads each email and only drafts when it is a genuine quote request. This works, but it reads your whole inbox, and every email uses a run against the 30-an-hour and 720-a-day limits. For an inbox receiving 80 emails a day that is well within the limits, but it is a lot more mail passing through the task than you need.
  • Route emailed requests to the form. The agency added a line to its website and email signature asking for quotes through the form, because the form also collects the word count and deadline the draft needs. Within a month, about 85% of requests came through it.

Gmail's own filters and labels can tidy the rest of the inbox around this task; labels, filters and AI replies in Gmail shows how. Keep the trigger simple, though: a subject filter is easy to test and easy to explain to whoever covers your inbox on holiday.

How do you write the ChatGPT Work task for quote replies?

Open ChatGPT Work on the web, describe the event and what should happen, and put your prices in the description, because the task cannot guess your rate card. The agency's version, with its illustrative rates:

When a new email arrives in my Gmail quotes inbox with a subject containing
"Quote request", read it and draft a reply. Rules:

NEVER send any email. Never archive, label, move or delete anything.
Your only job is to draft one reply per request and save it as a draft
in the same Gmail thread. I will approve each save.

If the email is not a genuine request for a translation quote (sales
pitches, suppliers, newsletters), draft nothing and say why in one line.
If the subject starts with Re: or Fwd:, or the email is part of an
existing thread, draft nothing and say so.

Work only from the email text and form fields. Do not open attachments.

Rate card (USD):
- Standard language pairs: $0.14 per word
- Specialist language pairs: $0.19 per word (list: [your specialist pairs])
- Rush (delivery in 2 working days or less): add 30%
- Certified translation: add $30 per document
- Minimum charge: $60 per language, applied BEFORE certification fees
- Standard turnaround: 2,000 words per language per 2 working days

The draft must:
1. Thank them by first name and restate the job in one sentence.
2. Show the price calculation line by line, then the total.
3. Give the delivery date from our standard turnaround (or rush if asked).
4. Say the price is an estimate until we see the final file.
5. If word count, language pair or deadline is missing, ask for it
   instead of guessing, and give no price.
6. Sign off as [your name], [agency name]. Plain, friendly, under 150 words.

Before the task is saved, ChatGPT shows you its Trigger, Condition and Prompt to review, and asks for any authorisation it needs. Read all three. In the agency's first attempt the Condition read "subject contains quote", without "request", which would have woken the task for every email mentioning a quote: supplier quotes, customers' replies, internal notes. One edit narrowed it.

If your rate card lives in a spreadsheet, The Complete Excel Course with AI Integration on Coding Liquids is built for that kind of work.

How do you test that the Gmail task fires before a customer does?

Test the task by sending yourself emails from a different address, then watch Scheduled for the task and any pending approval, and the task's chat for what each run did. Leave a few minutes between tests, because events arriving close together may be grouped into one run. The agency's test plan and results (illustrative):

Test emailExpectedWhat happenedFix
Form request: 2,300 words, two standard pairs, five working daysDraft at $644.00Draft at $644.00, approval request to save itNone
Form request with no word countDraft asking for the word count, no priceDraft asked for the file and word countNone
"Quote request" subject from a software vendor's sales teamNo draftNo draft; one-line reason in the task's chatNone
Certified one-page certificate, 180 words, specialist pair$60 minimum + $30 = $90.00Draft quoted $64.20 (180 x $0.19 + $30)Rule rewritten: minimum applies before certification
Rush job: 1,500 words, one standard pair, in 2 days$210.00 + 30% = $273.00Draft at $273.00None

The certificate test is the one that justified the whole test plan. The first version of the rate card said "Minimum charge: $60" without saying when it applied, and the draft worked out 180 words at $0.19 ($34.20), added the $30 certification fee and quoted $64.20, below what the agency actually charges for a certified page. A real customer would have accepted that price happily. Adding "applied BEFORE certification fees" to the rule and re-running the test produced $90.00. Whenever your prices have a minimum, a cap or a surcharge, test a request that sits right on the edge of it.

The reply-thread trap in a subject filter

The Re: and Fwd: rule in the task above came from a test the agency only thought of in its second week: a customer replying to a sent quote. The reply's subject reads "Re: Quote request - website", which still contains "Quote request", so the original task fired and drafted a fresh quote for a job that had already been priced. Nothing was sent, because of the permission level, but it produced a confusing approval request. Keep that rule in yours from day one, and include a reply in your test plan.

If a test does not fire at all, work through OpenAI's own checklist: open Scheduled and look for a paused task or a pending approval; confirm Gmail is still connected and authorised; check you have not hit your plan's active task limit; and look at the Condition again, letter by letter, against the subject you sent.

What does a good quote-reply draft look like?

A good draft thanks the sender, quotes from your rate card, gives a delivery date and asks about anything missing, and a person still reads every price before it is sent.

A request from the form, lightly anonymised in this illustration:

Subject: Quote request - website. Name: [first name]. Words: 2,300. Languages: two standard pairs. Deadline: 5 working days. Notes: Staff handbook, needs to match our existing glossary if possible.

The draft the task produced (illustrative):

Hi [first name], thanks for asking us to quote for your staff handbook: 2,300 words into two languages. Price: 2,300 words x $0.14 x 2 languages = $644.00. Delivery: within 3 working days of your go-ahead, inside your 5-day deadline. This is an estimate until we see the final file. Could you send the glossary you mentioned? We'll use it for both languages. Best wishes, [your name], [agency name]

The owner changed two things before sending. The draft promised "3 working days", which matches the turnaround rule for 2,300 words split across two linguists, but that week both standard-pair linguists were booked, so the owner wrote 5. And it said "We'll use it for both languages" about a glossary nobody had seen yet; the owner softened it to "We'll check it and use it where it fits." Neither change is something a rate card can encode. That is the reason the task drafts and a person sends.

Review takes about 3 to 4 minutes per draft against the 12 the agency used to spend writing from scratch. At 18 requests in an average week, that frees up roughly two and a half hours, and the first reply goes out sooner because the arithmetic is already done when someone opens the inbox.

When your quotes cannot come from a rate card

Not every business can price from a table. A small manufacturer quoting for laser-cut brackets needs a drawing, a material, a thickness and a quantity before anyone can price the job, and a confident-looking price from a model would be worse than none. For that kind of business, the same task does a different job: the rules tell it never to state a price, and the draft instead thanks the customer, lists which of the four details arrived and which are missing, gives the standard lead time for a priced quote (for example, two working days after a complete drawing arrives), and asks for the missing items in a numbered list. The value is speed and completeness: the customer hears back sooner and the estimator starts with a complete enquiry instead of a three-email chase.

How do you run the quote-reply task day to day?

Day to day, each run ends with an approval request to save the draft into Gmail: approve the save, never the send. If you prefer, leave the drafts unsaved and copy the text from the task's chat into your reply; nothing changes in Gmail until you do.

  • Notifications. Under Settings, then Notifications, choose push, email or both for task notifications, so a waiting approval does not sit unseen for half a day.
  • Price changes. Edit the rate card in the task from Scheduled the same day your prices change. A stale rate card in an automated draft is worse than no draft, because it looks authoritative.
  • A weekly ten-minute look. Compare five drafts with what was actually sent. If you keep making the same edit (a softer opening line, a different sign-off, an extra question about file format), add it to the task's instructions.
  • Usage. Event-triggered tasks run in Work, and Work usage allowances differ by plan; the tutorial on ChatGPT Work usage limits and credits explains what a Business plan allows if you add more tasks.

If several people handle quotes, the quote requests themselves can also be logged and tagged in a shared table; Airtable AI fields for incoming quote requests shows that approach, which pairs well with draft replies.

How do you pause, edit or switch off the task?

Manage the task from Scheduled in the sidebar, where you can pause, resume, edit or delete it. A few behaviours are worth knowing before you need them:

  • Pausing for holidays. If nobody will approve drafts for a week, pause the task. Approvals and drafts piling up help nobody, and paused tasks resume with one click.
  • Deleting. Deleting the task from Scheduled does not delete its chat. Deleting the chat pauses the task but does not delete a shared link to it, so delete the link separately if you ever shared the task.
  • Cutting access completely. Disconnect Gmail under Settings, then Plugins (or Apps). That removes ChatGPT's access to the mailbox, which a permission change alone does not.
  • Freeing a slot. Pausing or deleting a task frees one of your plan's active-task slots, so if you hit the limit, pause anything you no longer need before adding more.

If you want help fitting this into how your business actually runs, I offer one-to-one AI consultations for small businesses.

What happens to client emails, attachments and training settings?

The task should work only from the email text and form fields, not attachments, because quote requests often arrive with confidential documents attached. Three settings decide where the rest of the information goes:

  • Training. On Free, Go, Plus and Pro, OpenAI may use information accessed from apps to train its models if Improve the model for everyone is switched on. Turn it off under Settings, then Data controls, before connecting a business inbox. Business, Enterprise and Edu do not use app information for training by default.
  • Memory. ChatGPT may remember relevant information it reads through an app. If you do not want client details in memory, review your memory settings after setting up the task.
  • Scope. The task can reach only the connected Gmail account and what its permissions allow, and workspace app settings and role permissions on a Business plan still apply on top.

Set up this way, the task does the dull part of quoting (reading the request, doing the arithmetic, asking for what is missing) and leaves the commitment, the price that goes to the customer, with a person who has read it.

ChatGPT Gmail quote replies: questions answered

Can ChatGPT send the quote reply automatically if I trust it?

Not in this setup, although it is possible to loosen permissions. OpenAI warns that its Allow all actions level carries elevated risk because supported actions may run without another confirmation. A quote is a price commitment to a customer, so a person should read every one before it leaves. Keep the permission at Allow read actions and the never-send line in the task.

Does this work on the free ChatGPT plan?

No. OpenAI's help page says Free and Go accounts cannot create event-triggered tasks. You need Plus, Pro, Business, Enterprise or Edu, and on Enterprise and Edu an admin must first enable the setting that allows event-triggered scheduled tasks. On Free you can still paste a request into a chat and ask for a draft by hand.

Why did two test emails produce only one run?

Because OpenAI says multiple events may be grouped together, so emails arriving close together can be handled in a single run. Check the run in Scheduled to see whether both requests were drafted. When testing, leave a few minutes between test emails so each one's result is easy to tell apart.

Will my clients' emails be used to train OpenAI's models?

Not by default on Business, Enterprise and Edu, where OpenAI says information accessed from apps is not used for training. On Free, Go, Plus and Pro it may be used if the Improve the model for everyone setting is on, so turn that off under Settings and then Data controls before connecting a business inbox.

Further reads

Sources: OpenAI help pages on scheduled and event-triggered tasks, connected apps, app permissions and data controls; OpenAI's inbox use-case page; chatgpt.com/work; the ChatGPT and ChatGPT Business pricing pages, all checked 17 November 2026.

Want a second pair of eyes on your setup?

Bring the process you're trying to fix. On a 1:1 call we'll work out which AI tools should handle it, which parts stay manual, and how to roll it out without disrupting the team.

Book a 1:1 call with me