This is the single most common question we get, and it’s usually asked with a slightly guilty face. It shouldn’t be. Email is where most owners lose their mornings, and a lot of what’s in there is the same handful of questions wearing different clothes.
So here’s the honest split. AI is genuinely good at writing the reply. It should not be the thing that sends it.
What does an AI email setup actually do?
It sits between the message arriving and you answering it. When something lands, it reads the thread, works out what’s being asked, pulls in whatever it needs to know (your hours, your prices, your policy on deposits, what you told this person last month), and writes a reply in your voice. Then it stops, and puts the draft where you’ll see it.
You open your inbox to answers instead of questions. Most of them you skim and send. Some you rewrite. A few you delete, because they were never the right response.
The saving isn’t the typing. It’s the blank page, the mental gear change, and the twenty minutes you spend finding the thing you said last time.
Which emails is this for?
The repetitive ones. Every business has a set, and you already know yours by heart:
- Do you have availability on the 14th?
- What’s included in that price?
- Can you send the invoice again?
- Where do I park?
- Can we move Thursday to Friday?
These are the ones worth handing over first, because the answer barely changes and you’ve written it two hundred times. They’re also the ones that feel worst to spend a morning on.
The hard emails stay yours. The delicate client conversation, the complaint that needs a human, the negotiation, the bad news. You’ll notice those are also the ones you’re good at.
How do you get it to sound like you?
You feed it your own writing, not a tone setting.
This is where most tools go wrong. They offer you a dropdown labelled “professional” or “friendly” and generate something that reads like a press release. That isn’t your voice, it’s an average of everyone’s.
The setups that work start from your actual sent mail. A few dozen real replies is enough to pull out how you open, how you close, whether you use bullet points, whether you say “no problem” or “of course”, how long your sentences run, and what you never say. That gets written down as a voice profile the system reads before every draft.
Then it gets corrected. The first week you’ll be rewriting a lot. Note what you keep changing, put those rules in the profile, and the rewriting drops off fast. This is the refinement step every AI build needs, and email is the easiest place to see it work, because you get feedback on every single message.
Why shouldn’t it send on its own?
Because the cost of a bad reply is not symmetrical.
A draft you have to fix costs you thirty seconds. A wrong reply that already went out costs you a client, a booking, or an afternoon of untangling. Those two mistakes are not the same size, so the design shouldn’t treat them the same way.
There’s a second reason that matters more over time. The approval step is where you find out what the system is getting wrong. Turn it off and you stop looking, and a setup nobody checks quietly drifts until someone outside the business notices before you do.
We build with a human at the decision point on purpose, not because the technology can’t manage without one. It’s the same pattern behind an executive briefing system we built for an energy policy consultancy: the ingestion, the summarising, and the formatting all run themselves, and a person still approves what makes the cut. That’s what took an eight-hour weekly job under three hours without anyone losing control of what went out the door.
What does it need to know to be useful?
Four things, and they’re the same four behind any build worth doing:
- The problem. Which emails, specifically? Not “email” as a category. The five question types that eat the most time.
- The data. What does a good reply need to see? Your calendar, your price list, your booking system, the previous thread, your policies.
- The constraints. What must it never do? Never quote a price it wasn’t given. Never confirm a date without checking the calendar. Never send.
- Success. What does your inbox look like on a good Tuesday? Usually: everything routine already drafted, and the four emails that need you sitting at the top.
Skip the second one and you get fluent replies with wrong facts in them, which is worse than no reply at all.
What goes wrong?
Three failures, and they’re all avoidable.
It answers from nothing. If the system can’t see your calendar, it will still confidently write about your availability. Connect the data or restrict the scope.
It sounds like a brochure. Almost always because it was set up from a tone preset instead of your own sent mail.
Nobody reads the drafts after month two. The approval step becomes a rubber stamp. Keep it honest by checking a sample properly once a week rather than glancing at all of them.
Where should you start?
Pick one email type. The most repetitive one, the one you could write in your sleep.
Get that working properly, live with it for two weeks, and only then add the next. Doing one type well teaches you more about what your setup needs than launching six at once, and it’s a small enough thing that if it’s wrong, nothing is lost.
Almost nobody regrets starting narrow. Plenty of people regret pointing something at their whole inbox on day one.
If you want a straight answer on which of your emails are worth handing over and which ones aren’t, book a free readiness call. Twenty minutes, no pitch, and if it’s not a fit you still leave with clarity.