Case Study
Permission to Send
An email assistant should earn its access the way a new hire does, one kind of message at a time. Here is what that looks like when you actually build it.

By Colin P. Highland
The most important word in an email assistant is SEND, typed by a human, in capitals, with nothing else around it.
Picture a day spa. A customer writes in on a Thursday afternoon asking whether the couples package can move from Saturday to Sunday and whether her gift card covers the hot stone upgrade. The assistant reads the message, checks the package terms and the gift card policy, and writes a reply. Then it holds the reply. It sends the owner a short private note with the draft inside and waits.
If the owner answers with exactly SEND, the reply goes out once, from the spa's real address, inside the customer's original thread. If she answers with anything else ("Sunday is fine but the upgrade is extra, and be nice about it"), the assistant treats that as an edit, revises the draft and asks again. "Looks good, send it Monday" does not count as SEND. Software people call that a strict parser. Everyone else calls it not guessing.
I build these systems, so read what follows with that in mind. I'll say up front that most of the engineering is spent on the things the assistant is not allowed to do yet.
Reading comes first
Before an assistant writes a single word, it should read. On a new inbox ours does a read-only pass over recent mail and reports back what it found. Which senders never get a reply (the linen service's invoices, the software newsletter, the landlord's automated receipts). Which kinds of messages arrive most. Which few it would suggest handling first.
Nothing in that report takes effect on its own. The owner answers the report, and only the parts she confirms get applied. This sounds like a formality until you remember how many people have handed a new tool their inbox and watched it start "helping" before anyone told it what help looked like.
Writing like the person who signs it
An owner's email voice is a real thing and it's usually nothing like a template. Some people open with "Hi Dana!" and three exclamation points. Some open with the answer and sign with an initial.
So the assistant reads the owner's own sent mail and builds a voice profile from it. There are presets and a few dials on top of that (warmer, shorter, more formal), and the owner can see every style rule the system has adopted and revoke any of them. One small check I'm fond of watches for pleasantries. If the owner has never once written "I hope this email finds you well," the draft doesn't get to either.
Owners can also answer by voice memo. Plenty of them will dictate an answer from the car before they'll type one on a phone. The memo is transcribed, the transcript is used to write the draft and shown back in the note so the owner can see what was heard, and then the transcript is treated as temporary. It is a means of writing the reply, and nobody needs it hanging around afterward as a record of the owner talking to herself in traffic.
Every claim needs a source
The failure people fear most from an AI writer is the confident wrong answer, the reply that promises a price or a policy that doesn't exist. The defense is dull and it works. Before a delegated reply makes a claim, the claim gets checked against something the business actually wrote down: the knowledge base, the price list, a document in the library, an earlier message in the thread. Each task in the reply gets a recorded verdict on whether it was grounded. If it can't be grounded, the assistant asks instead of composing something plausible.
Quote follow-ups show why this matters. A customer thread can mention a dozen dollar amounts. One of them is a quote the business owes a follow-up on. The others are parking, a competitor's price the customer mentioned, the cost of a gift card from last year. Telling those apart is harder than it sounds, and following up on the wrong one makes a business look like it isn't reading its own mail.
The same discipline applies to attachments. When a customer asks for the intake form, the assistant picks the one document in the library that matches, decides whether to attach it or link it, and the send log shows whether the link was opened. That last part saves a surprising number of "did you get my email" calls.
When the answer is in somebody's head
Plenty of questions can't be answered from any document. Does the Sunday couples room fit two tables or three? Is Maria back from leave next week? In a real office the person at the front desk walks down the hall and asks.
The assistant does the same thing by email. It sends a private note to the teammate who handles that topic, waits for the answer and captures it when it arrives. Then it asks whether it should remember the answer for next time, and the learning only sticks if the owner confirms. Proposals that sit unconfirmed expire, so the knowledge base doesn't fill up with half-decided facts. If nobody answers, the assistant will chase the question politely. It won't chase on threads that involve health information, because nobody needs a reminder email about a patient's question bouncing around a team inbox.
Earning the right to send alone
Eventually an owner gets tired of typing SEND on replies she never changes, and that's the moment autonomy becomes worth discussing.
Autonomy here is granted per kind of message. Rescheduling requests might earn the right to go out unattended after a clean streak of drafts approved without edits. Refund disputes may never earn it, and the owner can pin any category to always-draft no matter how good the record looks. The assistant proposes the promotion and the owner says yes. It doesn't promote itself.
The part I care about most is that the permission can be lost. If an owner edits an auto-sent category heavily, marks a reply as a complaint, or takes over a thread to correct it, that kind of message drops back to draft. The owner gets a private note saying it happened and why. A permission that can only ever be granted works like a ratchet, and nobody should want a ratchet wired to their outbox.
Timing gets the same care. Replies wait for business hours and go out after a short delay, because an answer that lands nine seconds after a customer hits send at 2:14 in the morning reads exactly like what it is. Owners also get plain commands they can reply with at any point. TAKE OVER hands the thread back to a human, UNDO reverses the last thing the assistant did, and MUTE silences a sender.
Who is allowed to know what
An assistant with memory raises an obvious question about who that memory is for. What the assistant knows about a customer is used only when writing to that customer. A forwarded email from a stranger claiming to be the same person, or quoting the same phone number, gets nothing from the history. The context it pulls from older correspondence is deliberately bounded, so a reply about Sunday's appointment isn't built on three years of unrelated threads.
For health-related businesses the rules are stricter again. A staff member can add a fact to the knowledge base by emailing it in, which is one of the most useful features in the whole system, and the same channel refuses anything that looks like patient information.
When email is the wrong channel
Some threads should stop being email. A customer is confused, or upset, or the answer takes four back-and-forths that a two-minute call would settle. The assistant can offer a call or a text from inside the thread. The text names the email subject so the customer knows why her phone just buzzed. If the call doesn't connect or nobody picks up, the team gets one note about it and the customer gets a follow-up. The worst outcome in customer service is a promise of a call that nobody tracks, and it is very easy for software to produce.
Who this is for
If your business gets twenty emails a day and you like answering them, keep answering them. You will be faster and warmer than anything I can build, and your customers can tell. These assistants earn their keep for the owner who is clearing the inbox at 10 p.m. after a full day on the floor, and for the team where the right answer usually lives with someone who isn't at the desk.
Whoever's assistant you end up evaluating, mine included, find the setting that lets it send on its own and ask what takes that permission away. If the answer is nothing, keep looking.
Photographs: Jan van der Wolf, Pexels. Lettering and crops added. Scenes and names are illustrative.
Related Case Studies
The Opening Line · Full Episode
The AI that greets every visitor at workforcewave.com walks Viva through 72,798 business phone greetings, then the two of them try to write better ones.
Glad You Knocked · Scene 10
A great opening line sets the promise. So Viva asked for the closing line.
What Can We Conjure for You? · Scene 9
An elevator music company, a bakery run by retired magicians and a submarine-themed dentist. The lightning round.