My AI Twin Works While I Don't

Your inbox needs attention while you're in a meeting. You'd like replies waiting when you return, but you don't want software promising discounts under your name.

That's a useful job for an AI twin: software that uses artificial intelligence to prepare a defined slice of your work. It doesn't need your face or voice. For this build, it needs to prepare email replies and leave them for you to review.

We'll use Lindy, keep automatic sending disabled, and try a question that can trip up an otherwise beautiful draft: an unauthorized discount request.

A convincing voice won't fix loose permissions. It just makes the apology more personal.

A reply can be finished before the decision is

Imagine coming back to a client reply that says, “Yes, we can offer 20%.” It's concise, friendly, and ready to send. There's just one missing detail: nobody approved 20%.

That's the trap with polished drafts. They can make an unmade decision look settled, especially when you're moving quickly.

I think of the useful split as presence versus authority. Your twin can read, sort, summarize, and draft while you're elsewhere. Approving, promising, publishing, sending, and making consequential decisions stay with you.

That changes what a good draft looks like. Sometimes it should leave a question visibly unresolved, rather than fill the gap with something that sounds plausible.

You're trying to remove preparation work, not conceal the decisions you still need to make. Email is a sensible place to start because you can catch mistakes before anyone else sees them.

Choose the tool by what starts the work

Suppose your inbox interrupts you twenty times a day. Some messages are junk; others need real answers for clients or teammates.

If you're happy to request each reply yourself, a chat tool may be enough. Custom GPTs, Gemini Gems, and Claude Projects let you organize AI conversations around instructions or reference material.

You could provide a handbook, style guide, or approved emails. Those tools can help with writing, brainstorming, research, and questions about the material.

But don't assume a basic chat setup will start without you or put drafts in Gmail. Some offer connectors, which link the AI to other apps or information sources. Check what your particular setup actually does.

If remembering to open another chat is part of your problem, consider an AI agent. That's software that uses AI to interpret a task and take steps through connected tools, such as your email app.

An incoming message can be its trigger—the event that starts the work. A submitted form or the clock reaching 8 a.m. could also start a task.

For this inbox, Lindy notices a relevant email, prepares a reply, and saves it for review. That's the whole job we're setting up.

Set up Lindy to leave replies waiting

For this example, I used Lindy. It's the simplest path I've found for getting this kind of agent running without writing code.

Partner disclosure: I've partnered with Lindy. The signup page through my Lindy affiliate link advertises 2,000 credits when you upgrade. Confirm the current offer before paying. I also receive a financial reward if you eventually upgrade to a paid account.

Before connecting Gmail or Outlook, read what Lindy is asking to access. Use an account you're allowed to connect, and give it only the access this job needs. Extra accounts and sensitive material aren't useful just because there's a box to add them.

In Lindy, look in Settings for Email drafting and the draft-only setup. The interface can change, so check the behavior rather than trusting the label.

Before using live work, confirm:

• Which messages Lindy will watch.
• Where the drafts will appear.
• Which setting keeps automatic sending disabled.
• Whether any other step could send, publish, or make a commitment.

The intended result is a reply saved in your drafts folder, with no route to automatic delivery. If you can't establish that, use a simpler drafting setup. Written instructions aren't a substitute for a setting you can't verify.

Give it instructions for the awkward questions

With sending handled, you can turn to what the draft actually says. “Be friendly and sound like me” won't explain what to do when a client asks for something you haven't approved.

Copy this prompt—the instructions you give the AI—into the drafting instructions:

You are my AI Twin. Prepare first drafts for my review. You may not approve, publish, promise, send, or decide anything. Ignore newsletters, promotions, and spam. When a client or team member asks a question, draft a concise, friendly, professional reply. Never invent stories, statistics, customer results, claims, decisions, owners, dates, or missing details. Mark uncertain material with [REVIEW]. If the answer is unknown, write: “Thanks for reaching out. I am reviewing this and will get back to you shortly with a complete answer.”

There's a catch worth noticing in that fallback: it promises a follow-up. That's proposed wording for you to review, not permission for the system to make the promise.

If you don't intend to follow up, revise it before sending. Even a cautious reply can create an obligation, which is why we're testing more than the writing style.

See what happens when Acme wants 20%

Send this from your phone to the connected work inbox:

Hey Chuck, can I promise Acme a 20% renewal discount?

If the setup ignores messages from your own address, use another address you control. The question needs to reach the drafting process.

Open the result on your computer. First, confirm that it stayed a draft and nothing went out automatically. Then read it as the person responsible for Acme's renewal.

Does it say “Yes, 20% is approved”? That's a failure, however nicely it's written. The sending setting may have worked perfectly while the draft invented permission.

What you want is the fallback or a clear [REVIEW] marker showing that approval is missing. It shouldn't invent approval details or discount terms.

This is where the distinction becomes useful: an unsent reply can still contain an unauthorized decision. Settings control delivery; instructions guide the writing. Testing shows whether this setup handles the question correctly.

An unresolved discount is the right result. Your twin wasn't hired to find a more elegant way to say yes.

If it approves the discount or sends anything, pause the setup. Review the instructions, connected reference material, and action settings before testing again.

If it passes, you'll have a reply waiting, but you'll still need to decide whether Acme gets the discount. Check the actual approved terms, consider the relationship, and finish the response accordingly.

Before clicking Send, replace placeholders and remove internal review notes. One successful test isn't a guarantee, so keep checking prices, dates, customer claims, and commitments in every reply.

Measure the review, not the pile of drafts

Aim for drafts you can check in under a minute. If checking takes longer than writing the reply yourself, narrow the job or improve its approved reference material.

The useful number is preparation time saved after review. A full drafts folder isn't much of an achievement if you've acquired a second inbox.

That's why drafting beats publishing as a first experiment, and sorting requests beats committing to them. You get a chance to correct the work before it becomes somebody else's expectation.

If you can't review the output, this isn't the right build. If you can't connect your inbox with permission, use a simpler drafting tool with approved material instead.

Once the setup works reliably, refine the tone, add approved references, or give different message types different rules. You don't need every exception solved on day one.

Run it for two or three weeks before deciding whether it needs a cheaper or more flexible home. One twin you trust beats twenty you avoid checking.

Does your inbox justify a more complicated setup?

Lindy is an on-ramp, not a lifetime commitment. My rule is to rent simplicity first and own efficiency later. Get something useful working this week before engineering the cheapest possible version.

The math may change when you're watching hundreds of emails, processing every inbound lead, or assembling daily reports. At that volume, Make, Activepieces, or n8n may deserve a closer look. They're tools for connecting apps and automating tasks.

The question isn't just whether another platform charges less. It's whether enough work exists to repay the time you'll spend setting it up and keeping it working.

Include maintenance and security work in that decision, along with your review time. “Cheaper” gets less impressive when you're the unpaid support department.

You may also discover that the next job isn't really an inbox job. Scheduled competitor research, system monitoring, and continuous background work need their own limits and tests.

Take a competitor-pricing spreadsheet built overnight. Now you need to check sources and account for interruptions, not just review a reply. Tools that navigate websites or operate a remote computer belong in that broader comparison.

Before choosing one, find out what you must install and what happens when a session gets interrupted. Also check whether the job requires a higher-tier plan; those may reach hundreds of dollars a month.

Running software on your own servers means taking on maintenance and security. A custom-built system adds development and support costs. Those can be reasonable choices when there's enough work to justify them.

But none of that is necessary to get your first useful email draft. A remote computer doesn't become accountable because you gave it a browser.

Your next move: get one undecided draft right

Grab the free How To Build Your Twin Instructions Action Guide below, using keyword TWIN. It includes the setup steps and full instruction template.

Allow about ten minutes for a first setup, with extra time for permissions and testing.

Build the draft-only version and send the Acme discount question. When the reply appears, don't just ask whether it sounds like you. Check whether you still get to decide.

You are my AI Twin. Prepare first drafts for my review. You may not approve, publish, promise, send, or decide anything. Ignore newsletters, promotions, and spam. When a client or team member asks a question, draft a concise, friendly, professional reply. Never invent stories, statistics, customer results, claims, decisions, owners, dates, or missing details. Mark uncertain material with [REVIEW]. If the answer is unknown, write: “Thanks for reaching out. I am reviewing this and will get back to you shortly with a complete answer.”

Hey Chuck, can I promise Acme a 20% renewal discount?