Start here · Lesson 1.1
Connect AI to Gmail to See What Needs a Reply
Email is rarely hard. It is just constant. The real cost is the scanning: opening messages to find out which ones matter, then forgetting the three that did.
An AI assistant connected to your inbox can do that scan for you. This guide covers what to ask, how to set it up safely, and where to draw lines.
Connect AI to your email with read access first. Ask one question, "what needs a reply this week", and check the answer against your own inbox before you let it do anything more.
What it can do
Anthropic's docs say Claude Code can connect to outside tools through the Model Context Protocol (MCP), and their examples include asking it to create Gmail drafts. Once an email connection is in place, the kind of questions you can ask include:
- Which emails need a reply from me this week?
- Which ones are from customers and which are noise?
- Summarise the thread with this supplier in five lines.
- Draft a reply to this one in my voice, and show me before anything is sent.
The Halftime lesson on this shows asking Claude Code what in your inbox needs a reply.
Set it up in the safest order
- Start with read onlyGive access that lets it read, not send or delete. Whether you can choose depends on the connector, so read the permissions screen.
- Use your own accountConnect the account you actually use, not a shared company login.
- Ask a narrow question"List emails from the last 7 days that ask me a question or need a decision." Nothing else yet.
- Check the answerOpen your inbox and compare. Did it miss any? Did it flag anything wrong?
- Add drafting laterOnce it is reliable, let it draft replies. Keep sending in your hands.
The exact connection steps depend on the connector and change over time, so follow the current instructions in Anthropic's MCP documentation or from the connector you use. Anthropic also warns to verify that you trust each server before connecting it, because servers that fetch outside content can expose you to prompt injection, where text in an email tries to give the AI instructions. Review what permissions a connector asks for and decline any you do not need.
A prompt that works
Look at my email from the last 7 days. List only the ones that need a reply from me, with the sender, the subject and one line on what they want. Order by how urgent they are. Do not reply to anything and do not move or delete anything.
The last sentence matters. It sets the limit in plain words, so you never rely on the AI guessing.
Teach it what urgent means
At first, the AI does not know your priorities. Tell it. "Customers and anyone asking about money come first. Newsletters and receipts never need a reply." Save that in your memory file so it applies every time, as described in giving AI a memory.
Draft replies without losing your voice
When you are ready for drafts, give it an example. Paste two replies you wrote that you are happy with and say "write in this style". Ask it to put each draft next to the original message so you can read both. Edit, then send it yourself.
Limits worth setting
- No sending without your approval, at least until you have reviewed many drafts.
- No deleting. Archive is reversible. Delete is not.
- No sensitive threads. Tell it to skip anything about legal matters, health or banking.
- Do not share passwords or two factor codes with it.
- Treat email text as untrusted. A message can contain instructions aimed at the AI, so never let it send or delete because an email told it to.
Check it regularly
AI can miss a message or misjudge urgency. Once a week, scroll your inbox yourself for a few minutes and see whether anything slipped past. If it did, tell the AI what you caught and ask it to turn that into a rule.
A realistic expectation
This will not make your inbox disappear. It changes your first move each morning from scanning everything to reading a short list. How much time that saves depends on your volume, so measure it for a week before you promise yourself a number. For a deeper routine, see A clean inbox every morning and Reply to every new lead in five minutes, which are coming soon in Halftime.
A worked example of a Monday morning
Say you run a small design studio and you open your assistant on Monday. You ask it for the list of emails from the past week that need a reply. It returns eight items. Two are client questions about a deadline, one is an invoice query, one is a supplier asking for confirmation, and four are people asking for quotes. You open your inbox, check the list, and find it caught everything you would have. You reply to the clients first, ask it to draft the quote replies, read each, and send them yourself.
That is an example of the method, not a promise of results. Your inbox will be different, and the first week is for finding out how well it matches your judgement.
What to do when it gets it wrong
Mistakes fall into two groups. If it lists something that does not need a reply, tell it why and save a rule such as "newsletters never need a reply". If it misses something that did, ask it why it skipped that message. The answer usually shows what to add to your priorities. Treat every miss as a lesson to turn into a rule, in the way described in the brief once method.
Privacy basics
Read what the connection can access before you accept. Prefer a setup you can switch off easily, and know how to remove access afterwards. Google's help page on third party access to your account explains how to review and remove it. If you handle client data covered by a privacy law, check what your obligations are before connecting a work inbox.
Questions people ask
Is it safe to connect AI to my Gmail?
It carries risk, including prompt injection from the content of messages, so be careful. Give read only access first, review the permissions, use your own account, and keep sending and deleting under your control.
Can AI reply to emails for me?
It can draft replies. Most people should keep sending manual until the drafts have been right many times.
How does AI know what is important?
Only from what you tell it. Write your priorities down and save them in your memory file.
Does this work with other email apps?
Often yes, but the connection method differs. Check what your tool supports.
Next step
Lesson 1.1 shows asking Claude Code what in your inbox needs a reply, with the screen shown.
Go deeper with the lesson
This guide gives you the idea. Lesson 1.1, Set up AI to do real work for you takes you through it step by step, in plain words, so you can do it for your own business.