By KYN AI Advisory Team — AI implementation specialists, Singapore
An AI reply desk reads every reply to your outbound email or WhatsApp message, gives it one label, drafts a response where one is safe to draft, and routes the rest to a named person. The work is a pipeline of small steps, not one clever prompt. The label decides what happens next. Nothing is sent without a person approving it, until you have measured the desk and chosen which labels can go unattended.
What happens to a reply, step by step?
Seven steps, in order. Each step has one job and can be tested alone.
- Receive. The reply arrives from the mailbox or messaging channel and is stored with the original message and the thread.
- Clean. Signatures, quoted history and legal footers are stripped so the model reads only the new text.
- Check for stops. Code, not a model, looks for unsubscribe words, complaint words, legal terms and regulator names. A hit sends the reply straight to a person and halts the thread.
- Classify. The model assigns one label from a fixed list.
- Draft. For labels that allow it, the model writes a short answer from approved material: your offer, your price list, your calendar.
- Route. The label and a confidence flag choose the queue: auto-log, review, or escalate.
- Log. The result is written to the CRM with the label, the draft, the reviewer and the time.
Step 3 comes before the model on purpose. A rule that stops on "unsubscribe" cannot be talked out of it. See our note on prompt injection in business AI agents for why a reply is untrusted input.
Which labels should a reply desk use?
Use a short list with no overlaps. Six to eight labels is enough for most SMEs.
| Label | Draft allowed? | Next step | |---|---|---| | Interested | Yes, to a person for review | Book a call or send a quote request | | Not now | Yes, short and polite | Set a follow-up date | | Wrong person | Yes | Ask for the right contact, log it | | Question | Draft from approved answers | Review, then send | | Unsubscribe | No draft | Suppress in every channel | | Out of office | No draft | Pause until the return date | | Complaint or legal | No draft | Escalate to a named person |
A reply that fits no label goes to a person. An "other" label that auto-replies is a trap. Our walk-through of reply classification covers each label in detail.
How does the desk draft an answer?
The draft comes from material you approved, not from the model's memory. The model receives the reply, the thread, your offer text, your price list and a style sheet. It returns a short answer plus the list of sources it used.
Three rules keep drafts safe.
- Answer only what was asked. A question about delivery time gets a delivery answer, not a pitch.
- Quote only from the approved list. If the reply asks for a price not on the list, the draft says a person will confirm, and the item goes to review.
- One next step. Each draft ends with one action: a time to call, a form, a question.
The reviewer sees the reply, the draft and the sources side by side. Approve, edit or reject takes seconds when the page is laid out well.
What should the desk never do alone?
Match the rule to what is at stake: outbound messages, money and anything irreversible wait for a person.
- Sending to a contact who has not yet replied to you.
- Quoting a price, term or date not on the approved list.
- Replying to a complaint, a legal threat or a regulator.
- Agreeing to a change of terms.
- Deleting or merging a contact.
Our guide to approval gates for high-risk agent actions sets out how to build the gate. The related question of sending at all is covered in is it safe to let an AI agent send emails to customers.
How do you know the desk is accurate?
Test before you trust it. Take 100 real replies from the last quarter, have two people label them, and compare the desk's labels with theirs.
- Overall match rate. Share of replies where the desk and the people agree.
- Costly misses. Count separately an "interested" labelled as "not now", and any unsubscribe missed. One missed unsubscribe matters more than ten wrong "question" labels.
- Edit rate. Share of drafts a reviewer changes before sending.
Re-run the test whenever you change the model, the label list or the prompt. We cover the numbers in how to measure an AI reply desk.
What does a reply desk cost per reply?
In KYN's live systems for an insurance brokerage, sorting one reply costs about S$0.01. That covers the classification call only and depends on the model and the length of the text. Drafting adds a second call. At 500 replies a month, the model cost is a few dollars. The larger cost is the review time, which is why the edit rate matters.
What does a worked example look like?
Take an illustrative reply to a first-touch email: "Thanks, but we renewed our broker last month. Maybe look at us again in the new year. Also, how did you get my address?"
| Step | Result | |---|---| | Clean | Signature and quoted email removed; two sentences remain | | Stop check | "how did you get my address" matches a data-source question; the rule fires | | Classify | Primary label: not now. Flag: data-source question | | Draft | None. The routing rule overrides drafting | | Route | Named person, with the original email and the recorded source of the address | | Log | Label, flag, rule that fired, reviewer and time |
Two things sit in one message, and the cautious one wins. The person answers the data-source question, sets the follow-up date, and the thread pauses. A desk that drafted a cheerful "see you in January" would have skipped the question that matters.
What are the common failure points?
- Quoted history leaks in. The model labels the old email instead of the new reply. Test with long threads.
- Auto-replies read as interest. "Thank you for your message" is not a buying signal. Detect auto-reply headers in code.
- Mixed languages. A reply that opens in English and ends in Bahasa Indonesia needs a test case of its own.
- Drafts that promise. A draft says "we can deliver by Friday" because the thread mentioned Friday. Block dates and prices that are not in the approved list.
- Silent failures. A step errors and the reply sits unread. Alert on any reply older than your response limit with no label.
Where KYN fits
KYN built a reply desk for an insurance brokerage as part of an outreach system: one checkable first line per email, then a desk that sorts each reply. It runs in the client's own cloud account. The agent has its own restricted account, every action is logged with the actor, and outbound messages wait for a named person. See sales systems and security.
The first 14 days are a review of your stack and interviews with your team. A full build ships in under 90 days. Pricing is a one-off build fee and a monthly retainer of at least twelve months, quoted after the review.
Frequently asked questions
Can the desk reply on its own?
It can for low-risk labels once you have measured accuracy, such as acknowledging an out-of-office or confirming a booked slot. Anything that quotes terms or answers a complaint should keep a person in the loop.
What happens when the model is unsure?
The reply goes to the review queue with a flag. Set a rule that low confidence never auto-sends. A reviewer's decision is logged and can be used as a test case later.
Does the desk work on WhatsApp as well as email?
Yes. The pipeline is the same: receive, clean, check, classify, draft, route, log. Only the intake and the sending step differ by channel.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
