The safe way to use AI on customer emails is to let it write drafts and keep a person on the send button. Most support tools now offer that mode: Help Scout AI Drafts, Shopify Inbox suggested replies and Zendesk auto assist all put an AI reply in front of your staff instead of sending it to the customer. You stay in control by deciding which emails AI may draft, giving it accurate store information, checking every order fact before sending, and keeping disputes, fraud and angry repeat complaints in human hands.
Below: how to set that up in the tools small stores use, what each vendor’s documentation says about review and data, and a model for checking whether drafts save you time. Vendor details were checked on the vendors’ own pages on October 18, 2026. We have not tested these tools; everything below comes from their documentation.
Draft mode or autopilot: the setting that decides control
AI support tools work in one of two ways.
- Draft mode. The AI writes a reply, a person reads it, edits it and sends it. The customer only sees what a human approved. Mistakes cost you review time, not customer trust.
- Autopilot. An AI agent replies to the customer directly and hands over to a person when it cannot answer or a rule says it must. Mistakes reach the customer before anyone sees them.
Some vendors offer both. Shopify, for example, says its suggested replies are “a composing aid for staff and don’t drive the automatic responses that your Inbox agent sends.” Those are two separate features. For control, start in draft mode and move only a narrow, low-risk email type to autopilot once its drafts prove reliable.
Choosing a helpdesk in the first place is a different question. Our Gorgias vs Tidio vs Intercom comparison covers that. This article assumes you already have an inbox and want AI help inside it. For assistants that sit on your storefront rather than in your inbox, see our AI chatbots for Shopify comparison.
What each tool’s documentation says about review
The table below is built only from each vendor’s help pages. It compares how drafts reach the customer, where the AI gets its answers, and the availability limits the vendor states.
Help Scout AI Drafts No. Replies are generated “in a draft state” and users “review and revise the draft as needed before sending” Previous conversations and Docs articles; Docs and Beacon AI sources can be enabled per inbox Included with contact-based billing plans and available to current Plus and Pro plans. Needs about 100 conversation replies before drafts start. Trial covers 100 unique conversations
Shopify Inbox suggested replies No. Shopify says you need to read the reply closely and edit it if necessary Products, store policies, shipping, payments and pages. Terms of service and privacy policy are not processed English only. Needs an English store and admin, at least one Inbox conversation and 200,000 products or fewer. The page covers Inbox conversations and does not mention email
Zendesk auto assist No. Suggested replies and actions “require agent approval before being sent to a customer” Procedures you write first, then help center articles, similar solved tickets and connected knowledge Listed under the Copilot add-on. Does not work on AI agent tickets
Gorgias AI Agent Yes, it is built to reply automatically and “hands over to your team when it can’t confidently respond” Help center articles, website content, documents, guidance and Shopify data Works on email, chat, SMS and social channels. Order actions are opt-in. Test conversations in the Playground are not billed and send no real messages
Gemini in Gmail Google’s pages describe drafting and refining text in your mailbox, not automatic sending The email thread and your prompt Needs an eligible Google Workspace or Google AI plan for most features; some features on free personal accounts are US only. Google warns Gemini “may suggest inaccurate or inappropriate information”
Built from Help Scout, Shopify, Zendesk, Gorgias and Google help pages, checked October 18, 2026. Prices change often, so confirm them on each vendor’s pricing page.Vendors put the responsibility on you. Shopify says: “You’re responsible for the accuracy of the information that you provide your customers.” Draft quality also depends on the knowledge you give the tool. Zendesk calls procedures “your primary knowledge source,” and Help Scout will not even start drafting until an inbox has about 100 past replies to learn from.
Sort your inbox before you switch anything on
Control starts with a decision about which emails AI should touch at all. Use three lanes. The table is our own decision framework, not a vendor rule.
1. Draft and send after a quick read Pre-sale product questions, sizing, shipping times, how to use a discount code, store policy questions Full draft Product facts and policy wording match your current pages
2. Draft, then verify against the order Where is my order, returns and exchanges, address changes, missing items Draft only; a person confirms the order record Order number, tracking status, dates, amounts and what the policy allows for this order
3. Human only Chargebacks and disputes, suspected fraud, legal or regulator threats, injury or safety reports, anyone who has already complained twice, media requests None, or a private summary for the agent Handled by the owner or a senior person
Decision framework by E-commerce Partners. Adjust the lanes to your own products and risk.In lane 3 a confident but wrong sentence costs real money, for example a reply that admits fault in a dispute or promises a refund the bank has already paid. If a dispute is already open, follow the steps in our guide to what happens when a customer files a chargeback on Shopify instead of letting any tool answer.
Lane 2 is where most time is saved and most mistakes happen. A draft can sound right and still give the wrong delivery date. Unless your tool pulls live order data into the draft, treat every date, tracking number and refund amount as unverified until a person checks it.
Setting up drafts in the tool you already use
Help Scout
Turn AI Drafts on for the whole account under Your Plan > AI Drafts, or per inbox under Settings > Inboxes > [Inbox] > AI Drafts. Staff need the “Reply to Conversation” permission. Adding your Docs content as a source helps drafts match your written policies. On user-based billing, a hard spending limit stops drafts once it is reached each cycle, a useful cap during a trial.
Shopify Inbox
Suggested replies are on by default once the requirements are met. To change it, open Shopify Inbox settings, choose Message preferences, and set Automatically generate a reply to On or Off. Because the replies are built from your products, policies, shipping and payment settings, fix any outdated shipping times or return windows in Shopify first. The AI will repeat whatever your store says.
Zendesk
Auto assist leans on procedures: short written instructions for a type of ticket. Write procedures for your lane 1 and lane 2 emails before relying on suggestions. Zendesk says actions are only suggested if they are defined in a procedure (setting a ticket to Solved is the exception), so leaving refunds out of your procedures keeps refunds out of the suggestions.
Gorgias
Gorgias AI Agent replies on its own, so it is an autopilot tool by design. If you use it, run test conversations in AI Agent > Playground first, enable channels under AI Agent > Deploy only when the answers hold up, and set handover rules for your lane 3 topics. Order actions such as cancelling an order, processing a return or updating a shipping address are opt-in; leave them off until you trust the replies. Gorgias suggests keeping handover topics minimal at first, but a store that wants control should keep its full lane 3 list.
If you paste emails into ChatGPT, Claude or Gemini
Many small stores copy emails into a general chatbot instead. Check your plan’s data setting first, because customer details go with the paste. Gemini in Gmail can also draft or refine text from a prompt, but it has no view of your store data, so use it for lane 1 unless you paste verified order facts in.
- ChatGPT. OpenAI says that for ChatGPT Business, Enterprise, Edu and the API, “by default, we don’t use provided inputs and outputs to train our models.” On Free, Plus and Pro, data sharing is on by default and you can opt out.
- Claude. Anthropic says consumer chats on Claude Free, Pro and Max are used to improve Claude if you choose to allow it in your privacy settings, and incognito chats are not used for improvement even if that setting is on.
- Gemini in Gmail. Google says it does not use Workspace data to train its generative AI models, and that for users with Google AI plans, Gemini in Gmail does not use your content to train its models.
Whatever the setting, strip what the draft does not need. A shipping update never needs a card number, password, home address or phone number. Use placeholders such as [ORDER] and [TRACKING] and fill them in from your admin after approval. What the law in your market requires is a separate question this article does not cover.
A prompt that keeps the AI inside its lane:
- You write draft replies for [store name]. Use only the facts in the policy text and order details I paste below.
- If a fact is missing, write [CHECK] instead of guessing.
- Never promise a refund, credit, replacement or delivery date that is not stated in the pasted text.
- If the customer mentions a chargeback, fraud, a lawyer, an injury or a safety problem, do not draft a reply. Write “ESCALATE” and one line explaining why.
- Keep the reply under 120 words, in plain, friendly English.
The [CHECK] line turns a silent guess into a visible gap your reviewer cannot miss.
Review rules that keep you in control
- Check every fact the customer can act on. Order number, tracking status, delivery estimate, refund amount, return window. If the draft states it, someone confirms it in the admin.
- Read for promises. Look for “we will refund”, “free replacement”, “by Friday”. A promise the AI made up is still a promise from your store.
- Log edits for two weeks. Note whether each draft was sent as is, lightly edited or thrown away. That gives you your rewrite rate for the model below.
- Fix the source, not the reply. When drafts keep getting one thing wrong, update the policy page, help article or procedure the tool reads.
- Spot-check after store changes. New shipping rates or a new return policy can make good drafts wrong overnight.
If you sell to customers in the EU with a chatbot that replies on its own, there are disclosure rules to consider; we cover them in our article on the EU AI Act and store chatbots.
Does it actually save time? A worked model
Drafts save time only if quick reviews outweigh the drafts you throw away. Here is a hypothetical model with assumed inputs; replace them with your own.
- 400 customer emails a month.
- Writing a reply from scratch takes 5 minutes.
- Reviewing a good draft takes 1.5 minutes with a quick read, or 3 minutes if you check every order fact in the admin.
- A bad draft costs the review time plus the full 5 minutes to write the reply yourself.
Writing every reply yourself takes about 33 hours a month here.
10% 13.3 20.0 23.3 10.0
20% 16.7 16.7 26.7 6.7
30% 20.0 13.3 30.0 3.3
50% 26.7 6.7 36.7 3.3 lost
70% 33.3 0.0 43.3 10.0 lost
Hypothetical model, 400 emails a month, 5 minutes per reply written from scratch (about 33.3 hours). Calculated with a formula and checked by counting good and bad drafts separately; both methods gave the same results.The break-even rewrite share is the time saved per good draft divided by the time to write a reply. With a quick read, that is (5 minus 1.5) divided by 5, or 70%. With a careful fact check, it is (5 minus 3) divided by 5, or 40%.
So keep the fact check and make it fast. Order data shown next to the draft, accurate policy pages and a prompt that marks missing facts with [CHECK] all shorten review. Careful review of drafts full of guesses can end up slower than writing by hand.
If returns fill your inbox, a self-service portal may remove those emails entirely. See our roundup of return management software.
What you should check next
- Check whether your helpdesk’s AI drafts, replies on its own, or both, and which is switched on.
- Sort last month’s emails into the three lanes and count how many fall in each.
- Update shipping times, return windows and product details in the pages your tool reads.
- Write your lane 3 list and set it as handover rules or procedures.
- If you paste into a general chatbot, check the plan’s training setting and remove personal details.
- Track sent, edited and discarded drafts for two weeks, then run your own numbers through the model.
Frequently asked questions
Will customers know a reply was drafted by AI?
In draft mode, the customer receives whatever your staff send. Zendesk, for example, says customers aren’t aware of auto assist and agents remain in control of what they send. If an AI agent replies on its own, disclosure rules may apply depending on where your customers are.
How many past emails does a tool need before drafts are useful?
It varies by vendor. Help Scout says an inbox needs about 100 conversation replies before AI Drafts starts generating. Tools that rely on your help articles or procedures need those written first.
Can I let AI send replies automatically for simple questions?
You can with autopilot tools such as Gorgias AI Agent. A cautious approach is to start with one lane 1 email type, test it in the vendor’s sandbox, and keep a handover rule for anything outside that type.
What rewrite rate should I aim for?
There is no published standard. Use the model above: your break-even rate is the minutes saved per good draft divided by the minutes it takes to write a reply. Aim to stay well below it.
The takeaway
Keep AI on drafting and people on sending. Decide which emails AI may touch, feed it accurate store information, make order facts quick to check, and measure how often drafts get thrown away.
Sources
- Help Scout Docs, “AI Drafts”, updated October 18, 2026. docs.helpscout.com
- Shopify Help Center, “Automatically generating suggested replies in Shopify Inbox”. help.shopify.com
- Shopify Help Center, “Set up instant answers for Shopify Inbox”. help.shopify.com
- Zendesk Help, “About auto assist”. support.zendesk.com
- Gorgias Docs, “AI Agent explained”. docs.gorgias.com
- Gorgias Docs, “Set up and go live with AI Agent”. docs.gorgias.com
- Google Gmail Help, “Learn about Gemini features in Gmail”. support.google.com
- Google Gmail Help, “Collaborate with Gemini in Gmail”. support.google.com
- Google Gmail Help, “Learn how Gemini in Gmail, Calendar, Chat, Docs, Drive, Sheets, Slides, Meet and Vids protects your data”. support.google.com
- OpenAI Help Center, “ChatGPT Enterprise and Team data policy”. help.openai.com
- Anthropic Privacy Center, “Is my data used for model training?”, updated March 16, 2026. privacy.claude.com
Last reviewed: October 18, 2026. Feature descriptions, admin paths, plan availability and data settings match each vendor’s documentation as checked on that date.
How this article was made: researched and drafted with AI assistance, with every factual claim checked against the sources above. The three-lane framework, the prompt and the 400-email time model are our own; the model uses assumed inputs, not measured store data.
Related guides
- Gorgias vs Tidio vs Intercom: the right pick for your store
- Best AI chatbots for Shopify: 2026 comparison
- What happens if a customer files a chargeback on Shopify?
- EU AI Act: chatbots serving EU shoppers must disclose AI