AI Workflows Intermediate 20 minutes to build

Customer Email → AI Reply → CRM

Handle inbound customer email in minutes instead of hours: AI-drafted replies in your own voice, automatic CRM logging, customer tagging and follow-up reminders — built with Gmail, ChatGPT, HubSpot and Zapier.

Difficulty
Intermediate
Completion time
20 minutes to build
Time saved
5–7 hours per week

Workflow overview

Inbound email is where small businesses quietly lose money. Not because the answers are hard, but because the same twelve questions arrive endlessly, replies slip to the next day, and nobody remembers to follow up on the ones that were nearly a sale.

This workflow keeps a human in the loop where judgement matters and automates everything around it. AI drafts, you approve; the CRM logging, tagging and reminder-setting happen without anyone touching them.

Build it once in about twenty minutes. The important design decision is that nothing sends automatically — drafts land in your Gmail drafts folder, so a bad answer is never a customer's problem.

Goal

Reply to every customer email the same day, with every conversation logged, tagged and scheduled for follow-up automatically.

Best for: Small businesses handling 10–100 customer emails a week without a dedicated support team.

Required AI tools

Gmail

Where the customer conversation starts and ends.

ChatGPT

The reasoning and writing engine at the centre of every workflow.

HubSpot

Free CRM for contacts, deals, tasks and email sequences.

Zapier

Connects your apps so the workflow runs without you.

The workflow at a glance

1
Email arrives
2
Classify & draft
3
Human review
4
Log to CRM
5
Tag & schedule follow-up

Step-by-step guide

Work through the steps in order. Every prompt is copy-and-paste ready — swap the bracketed parts for your own details.

1

Capture the incoming email

Gmail 4 min

Create a Gmail label called AI-Draft and a filter that applies it to messages arriving at your support or hello address. In Zapier, create a Zap with the trigger 'New Labeled Email' in Gmail pointing at that label. Filtering by label rather than by inbox keeps newsletters and internal chatter out of the automation.

Prompt

Gmail setup: 1. Settings > Filters > Create a new filter. 2. To: [support@yourdomain.com] 3. Apply label: AI-Draft, and skip nothing else. Zapier setup: 1. Trigger: Gmail — New Labeled Email 2. Label: AI-Draft 3. Test the trigger with a real recent email before continuing.

What you should get back

Every genuine customer email is now labelled and available to Zapier, with subject, body, sender name and address as usable fields.

2

Classify the email and draft a reply

ChatGPT 6 min

Add a ChatGPT step in Zapier that does two jobs in one call: categorise the message and write the reply. Categorisation is what powers the tagging and routing later, so ask for it as a strict single word from a fixed list. Feed in your business facts — hours, pricing, policies — so the draft is accurate rather than plausible.

Prompt

You are the customer support voice for [BUSINESS NAME], a [DESCRIPTION]. Business facts you may use: - Hours: [HOURS] - Pricing: [PRICING] - Delivery/turnaround: [TIMES] - Refund policy: [POLICY] Read the customer email below and return exactly this format: CATEGORY: one of SALES | SUPPORT | BILLING | COMPLAINT | PARTNERSHIP | SPAM URGENCY: HIGH | MEDIUM | LOW SENTIMENT: POSITIVE | NEUTRAL | NEGATIVE DRAFT: [the reply] Reply rules: - Under 150 words, warm and direct, no corporate filler. - Use only the business facts above. If the answer is not covered, write the sentence: "I want to get this exactly right for you — [OWNER NAME] will confirm within one business day." - Never invent prices, dates or policy. - End complaints with a concrete next step and a time. Email: [BODY]

What you should get back

Structured fields Zapier can branch on, plus a reply draft that either answers fully or explicitly hands off — never a confident guess.

3

Put the draft in front of a human

Gmail 3 min

Add a Gmail 'Create Draft' action, threaded as a reply to the original message. Drafts, not sends. In practice you skim, adjust one sentence, and hit send — that is roughly twenty seconds per email versus five minutes writing from scratch, with none of the risk of an automated wrong answer reaching a customer.

Prompt

Zapier action: Gmail — Create Draft Reply To: {{trigger sender}} Thread: {{trigger thread id}} Body: {{ChatGPT DRAFT}} Add a Filter step before this action: Only continue if CATEGORY does not equal SPAM.

What you should get back

A threaded draft waiting in Gmail, correctly addressed, ready to review and send in seconds.

4

Log the conversation and tag the customer in HubSpot

HubSpot 4 min

Add two HubSpot actions: find-or-create the contact from the sender's address, then log an engagement note containing the original message, the category and the sentiment. Write the category into a contact property so your list segments itself — sales enquiries, support cases and complaints separate automatically with no manual tagging.

Prompt

Zapier actions: 1. HubSpot — Create or Update Contact Email: {{sender email}} | First name: {{sender name}} Property last_enquiry_type = {{CATEGORY}} Property last_enquiry_date = {{date}} 2. HubSpot — Create Engagement (Note) Body: Category {{CATEGORY}} / Urgency {{URGENCY}} / Sentiment {{SENTIMENT}} Original message: {{email body}}

What you should get back

A CRM contact record that builds itself, with a full enquiry history and segmentable properties you never had to type.

5

Create the follow-up reminder

Zapier 3 min

The highest-value part of the whole workflow. Add a Paths step: SALES enquiries create a HubSpot task due in two days, COMPLAINT creates one due tomorrow with high priority, and everything else creates a task due in five days. This is the step that stops warm enquiries quietly dying in the archive.

Prompt

Zapier Paths: Path A — CATEGORY = SALES → HubSpot Create Task, due in 2 days, title: "Follow up sales enquiry: {{sender name}}" Path B — CATEGORY = COMPLAINT → HubSpot Create Task, due tomorrow, priority HIGH, title: "Resolve complaint: {{sender name}}" Path C — everything else → HubSpot Create Task, due in 5 days, title: "Check in: {{sender name}}"

What you should get back

A self-populating task list ordered by commercial value, so the enquiries most likely to convert are the ones you chase first.

Copy-and-paste prompt library

Every prompt used in this workflow, in one place. Tap copy and paste straight into your tool of choice.

Prompt 1 — Classify and draft

Return exactly: CATEGORY (SALES/SUPPORT/BILLING/COMPLAINT/PARTNERSHIP/SPAM), URGENCY, SENTIMENT, then DRAFT. Reply under 150 words, warm and direct, using only the business facts provided. Never invent prices, dates or policy. If unknown, promise owner confirmation within one business day.

Prompt 2 — Voice calibration

Here are 10 replies I have written to real customers. Describe my writing voice in 5 rules — sentence length, greeting style, formality, how I say no, how I close. Then rewrite this draft to match those rules exactly.

Prompt 3 — Complaint handling

This customer is unhappy. Write a reply that: acknowledges the specific problem in their words, takes responsibility without over-apologising, states exactly what happens next and by when, and offers [REMEDY]. Under 120 words. No defensiveness, no excuses.

Prompt 4 — Sales enquiry qualification

From this enquiry, extract: what they want, budget signals, timeline signals, decision-maker status, and the single question I should ask to qualify them. Then write a reply that answers their question and asks that one question.

Prompt 5 — Follow-up nudge

Write a follow-up email for a customer who enquired [X] days ago and did not reply. Under 60 words, no guilt, no "just checking in", one specific reason to reply now, easy exit if they are not interested.

Prompt 6 — Canned response library

From these 30 past customer emails, identify the 10 most common questions and write a reusable reply template for each. Mark variable parts with [BRACKETS]. Keep each under 120 words.

Expected output examples

Classified email

CATEGORY: SALES URGENCY: HIGH SENTIMENT: POSITIVE DRAFT: Hi Sam — yes, we can cover a 40-person event on the 12th. Our standard package runs at £480 including setup...

Escalation draft

Hi Dana — thanks for flagging this. I want to get this exactly right for you — Priya will confirm the warranty position within one business day. In the meantime, here is what I can tell you now...

CRM note

Category SALES / Urgency HIGH / Sentiment POSITIVE Original message: "Hi, do you cover events for around 40 people? Looking at the 12th of next month..." Contact property last_enquiry_type set to SALES.

Auto-created task

Follow up sales enquiry: Sam Whitfield Due: in 2 days · Priority: Medium · Linked contact: sam@…

Real-world example

Two-person catering business, 60 enquiries a week

A catering pair were spending most of Monday and Thursday answering the same questions about capacity, dietary options and pricing, and admitted that maybe one in five enquiries never got a second reply.

They built this Zap in an afternoon. The single most valuable change was not the drafting — it was Path A, which created a follow-up task for every sales enquiry two days later.

Average reply time fell from around eleven hours to under one. Because every enquiry now produced a CRM contact tagged by type, they could finally see that corporate enquiries converted at three times the rate of private events, which changed where they advertised. Bookings rose without a single extra pound of ad spend.

If you're a student

Freelancing or running a side business while studying? Build a lighter version: Gmail label plus the classify-and-draft prompt, no CRM. It keeps client email off your mind during lectures without letting anything slip.

If you run a business

Add the canned response library prompt after a month of running this. You will find that ten templates cover 70% of your inbound, and the AI step becomes a matter of picking and personalising rather than writing from scratch.

Best practices

  • Create drafts, never auto-send. One wrong automated reply costs more than the time it saves.
  • Feed real business facts into the prompt — hours, pricing, policy — so drafts cannot hallucinate them.
  • Give the model an explicit escape sentence for anything outside its knowledge.
  • Calibrate the voice on ten of your own past replies before going live.
  • Filter SPAM out before the CRM step so your contact list stays clean.
  • Review the tasks list every morning; the automation creates the reminders, but you still have to act on them.

Common mistakes

  • Letting AI send unreviewed replies

    It will eventually quote a wrong price or agree to something you cannot deliver. Keep the human approval step permanently.

  • No fixed category list

    Free-form categories produce twelve spellings of the same thing and break your CRM segmentation. Enumerate the options in the prompt.

  • Skipping the follow-up path

    Fast replies without follow-up still lose the deals. The reminder step is where most of the revenue in this workflow lives.

  • Running everything through one generic prompt

    A complaint and a partnership enquiry need different tones. Branch on category once the volume justifies it.

Variations

No-code-lite version

Skip Zapier entirely: use a saved ChatGPT project with your business facts, and paste emails in manually. Slower, but running in ten minutes with zero subscriptions.

Shared inbox with routing

Add a Slack step that posts HIGH urgency or NEGATIVE sentiment emails into a team channel so nothing sensitive waits for one person to log in.

Multilingual support

Add a detect-language instruction and reply in the customer's language, while logging an English summary to the CRM so the whole team can read the history.

Frequently asked questions

Is it safe to send customer emails through ChatGPT?

Use a business or API plan with data training disabled, tell customers in your privacy policy that you use automated tools to assist with support, and never process payment card or health data this way.

What does this cost to run?

HubSpot's CRM is free at this scale, Zapier's starter plan covers a few thousand tasks a month, and OpenAI API usage for a few hundred emails is typically a couple of dollars. Well under the cost of an hour of your time.

Will customers know a machine drafted the reply?

Not if you calibrate the voice and actually read every draft before sending. The tell is generic corporate phrasing — which the voice calibration prompt is designed to remove.

Can I use a different CRM?

Yes. The pattern is identical in Pipedrive, Zoho or Airtable — find-or-create contact, log a note, set a task. Only the Zapier action names change.

What if an email needs information the AI does not have?

That is exactly what the escape sentence handles. It hands off honestly with a deadline instead of guessing, which customers respond to far better than a wrong confident answer.

Get the next workflow in your inbox

One practical AI workflow every week — no fluff, no hype, just the steps and prompts.

Put this workflow to work today

Compare the tools used here, then go deeper with our written guides.