Let your agent draft client replies, then send them yourself
Can AI draft emails to clients safely? Yes, if your agent drafts and you send. AI agents for client work: what to give yours, what to check, what to write.
A yoga studio owner emails her web designer on a Sunday night. Can the timetable page show the new Saturday class, and while he's at it, could the site take bookings? The designer asks his AI agent to draft a reply. Thirty seconds later it's ready: friendly, clear, and it says the booking system "can be set up this week". It can't. It isn't in her plan, he hasn't priced it, and this week is already full.
The draft was fluent. It was also wrong in the one sentence that mattered. That sentence is why, with AI agents for client work, the agent drafts and you send.
Can AI draft client replies safely? Yes, as long as the agent drafts and you send. Give it the facts it needs, read every draft for facts, promises, prices, dates and tone, and keep the send button as a step only you take. Most of the time you'll change a word or two. Now and then you'll catch the sentence that would have cost you a weekend.
Why let an agent draft replies at all?
Because replies are where a one-person agency's time leaks. Each one is short, but there are dozens a week, and each needs you to remember what the client pays for, what you last said and what's already on your list.
An agent is good at exactly that kind of gathering. With access to the client's notes, it can find the plan, the last three messages and the request, and turn them into a tidy first draft. You go from writing to editing, which is faster and usually better.
What an agent can't do is know what you haven't told it, or decide what you're willing to promise. That's the part you keep.
What does your agent need to write a good draft?
A draft is only as good as the facts behind it. Before you ask for one, make sure the agent can see:
- What the client pays for. Their care plan or project scope, including what's not included.
- The conversation so far. The last few messages, so it doesn't repeat a question you've answered.
- Your prices. So an extra gets a quote, not a guess.
- Your rules. How you sign off, how fast you reply, words you never use, and what you never promise in writing.
The simplest way to give an agent this is to keep each client's notes in one place it can read. The guide to organising client files for an AI agent shows a folder layout that works.
Tools can also connect your agent to the places those facts live. MCP (the Model Context Protocol) is an open standard for connecting AI apps to outside data and tools, which is how an agent can read a client's records without you pasting them in. The short explanation of what MCP means covers it in plain words.
Then ask for the draft in one clear sentence: "Draft a reply to the yoga studio. Timetable change is included in her plan. Online booking isn't: say I'll send a price by Wednesday."
Direction like that matters more than any clever prompt. It tells the agent what's decided, so it doesn't decide for you. If you haven't decided yet, say that too: "Say I'll check and reply by Tuesday." A draft that buys you time honestly is better than one that answers too soon.
It also helps to keep a few of your own past replies in the client's notes. An agent that has seen how you write to the café owner (short, warm, first names) will draft closer to your voice than one working from a blank page.
What should you check before you send?
Read every draft as if a client will forward it to their business partner, because sometimes they will. Five checks cover most mistakes:
| Check | What to look for | Example of a miss |
|---|---|---|
| Facts | Names, prices, dates, page names, what's included | "Your plan includes three changes a month" when it's two |
| Promises | Any date, price or piece of work you haven't agreed | "I can set up bookings this week" |
| The question | Does it answer what the client actually asked? | Replies to the timetable, ignores bookings |
| Tone | Does it sound like you, at the right warmth? | Overly formal to a client you text daily |
| Extras | Anything that isn't included gets a quote, not a yes | "Happy to add that!" for a new page |
If you edit the draft, read the final version again. Edits are where new mistakes slip in: a changed date in one sentence and the old one left in another.
Which replies should you write yourself?
Some messages need your own words, even if the agent helps you gather the facts first:
- Bad news. A missed date, a broken form, a price going up.
- Apologies. An apology in borrowed words reads that way.
- Money disagreements. A disputed invoice or an unpaid month.
- Anything with legal weight. Ending an agreement, ownership of the site, refunds.
- A client who is upset. They want to hear from you, not a well-structured paragraph.
For these, ask the agent for the facts ("list what we agreed about the gallery page and when") and write the reply yourself.
How do you approve a draft properly?
Approval should be a real step, not a glance. Three habits make it one:
- Approve the exact words. The version you read is the version that goes. If anything changes after you've approved it, including a small fix, read it again. It's the same idea as approving one exact version of a website.
- Keep sending separate from drafting. The agent prepares; you press send. If your setup lets an agent send messages directly, turn that off for client replies.
- Keep a record. Save the reply you sent with the client's details. When a question comes up in three months, you'll know exactly what you said.
The wider rule, and the list of things an agent should never send on its own, is in what your AI agent should never send without you.
Should you tell clients an agent drafts your replies?
That's your call, and it depends on the client. Many won't mind as long as the reply is accurate and sounds like you, because you read and sent it. If a client asks, answer honestly. The guide on whether to tell clients you use AI covers how to say it in a way that builds trust.
What matters more to clients than how a reply was drafted is whether it was right. The yoga studio owner would never have known the first draft existed. She would have remembered a booking system promised for a week that never came.
A quick routine for drafted replies
- Sweep client messages at set times, not all day.
- For each request, ask the agent for a draft with one sentence of direction.
- Run the five checks.
- Rewrite anything in the "write it yourself" list.
- Send, and save what you sent.
You still read every word. You just stop writing every word from scratch, and the mistakes you catch are often the ones you would have made yourself, tired, late in the evening.
Where Volant fits
Volant lets your agent draft client replies with the client's plan and messages in view. Anything Volant sends or publishes waits for your approval, so a draft stays a draft until you've read it. The page on approving what your agent sends shows how that works day to day.