Let the agent draft, you send — a safer way to automate email
AI can answer most routine email. It shouldn't press send on its own, at least not at first. How to set up drafting that actually saves time.
· 2 min read · AmplifyIQ
Email is the obvious first thing to hand to an AI agent. It's where the day goes, most of it is routine, and language models write perfectly good email. But an agent that sends on your behalf, unchecked, is one bad afternoon away from an apology to a customer.
The pattern that works is simple: the agent drafts, a person sends.
Why drafting wins
- Most of the time is in the writing, not the sending. Reading the thread, finding the order number, checking the price, composing a polite reply. An agent that does all that and leaves a draft saves nearly all the time.
- Mistakes stay private. A wrong draft gets fixed or deleted. A wrong email is out there.
- You learn what it gets right. After a couple of weeks of approving drafts you'll know which kinds of email it handles perfectly, and which it doesn't.
Setting it up
Give it the facts, not just the inbox. A reply is only as good as what the agent knows: your prices, opening hours, delivery times, returns policy, the answers to the ten questions everyone asks. Put them in one document the agent can read.
Teach it your voice. Show it twenty of your own sent emails. It will pick up how you greet people, how formal you are, and how you sign off.
Tell it when not to draft. Complaints, legal letters, anything from a solicitor or a regulator, and anything the agent isn't sure about should be flagged for a person, not answered.
Label as well as draft. Every incoming email gets a label (enquiry, order, complaint, supplier, spam), so what needs a human stands out.
When to let it send
Once a category of email is right every time for a few weeks (say, "what are your opening hours?" or "has my order shipped?"), you can let the agent send those on its own. Keep drafting for everything else. Most businesses end up with a small set of fully automatic replies and a much larger set of one-click approvals, and that combination is where the hours come back.
Things to check before you start
- Access. The agent needs permission to read your mail and create drafts, and nothing more. It doesn't need to delete or forward.
- Where the data goes. Your customers' email is personal data. Know where the agent's provider processes it and under what agreement.
- An off switch. One click to pause the agent, for the day something unusual happens.
Done this way, email automation doesn't feel like handing over the keys. It feels like arriving in the morning to find the replies already written.