You already know the twenty questions. Where is my order. Can you do it cheaper. Do you deliver to my area. Are you open Saturday. Is the deposit refundable. They arrive by email, Instagram DM, WhatsApp, contact form and phone, and each one interrupts whatever you were doing. Answering them is not hard work, which is why it eats the day.
AI is useful here because the job is bounded: read a short message, find the relevant fact, write four sentences in a recognisable voice. The failure mode is predictable. The draft comes back fluent, polite and forgettable, and the customer can tell within a line that no person wrote it. This guide builds a reply system that avoids that: facts from your records, voice from a brief you wrote, a human reading everything first. Our standards set out what we will not publish.
Why the robot problem is usually a context problem
Owners assume the uncanny feeling comes from the writing. It rarely does. The tell is that the reply contains nothing the customer did not already know: it restates the question, apologises for “any inconvenience” and promises to “get back to you”. Missing information reads as evasion.
Three inputs fix most of it, none of them prompt tricks: the facts of this case, from your records; a written voice brief rather than the word “friendly”; and escalation rules that sit outside the model.
Step one: collect the questions you actually get
Start from the messages, not a list invented at the desk. Open your sent folder, WhatsApp history, Instagram inbox and contact form submissions for the last sixty days, and copy the customer’s own wording into one document, a line per message, unedited, with names stripped.
Log the question, not the topic
“Delivery” is a topic. “Do you deliver outside the ring road and is it the same price” is a question, and that is the unit you build a template from. Expect a long list with heavy duplication.
Sort by frequency, then by cost
Group the duplicates and count them, then mark each group with how long a good answer takes and how badly a wrong one would hurt. Parking is frequent and cheap; refunds are rarer and expensive. The cheap group pays for automation immediately; the expensive group keeps a human.
Keep both vocabularies
Your site says “bespoke joinery”; customers write “can you make me a shelf”. Note both: the model mirrors whichever you give it, and the customer’s words make a reply feel like an answer.
Step two: write a house-voice brief
A voice brief is one page, written once, pasted into every drafting session. Keep it concrete: “warm” and “professional” tell a model almost nothing, because every business claims both.
What belongs in the brief
- Six real sentences you have written. Three yes, three no, copied from sent mail. These do more work than any description of tone.
- Formality settings. Contractions or not, first name or “Hi”, your name or the business at sign-off.
- Sentence length. If you write short, say so and give a maximum. Left alone, a model writes longer sentences than most owners.
- A banned-phrase list. “Valued customer”, “apologise for any inconvenience”, “reach out”, “circle back”, “we strive”, “your satisfaction is our top priority”, “at your earliest convenience”. Each has a plainer replacement.
- How you say no. The shape of a good refusal: what you cannot do, why in one clause, what you offer instead.
- Formatting rules. Em dashes and bullet points, and how many exclamation marks (usually none in anything apologetic).
What to leave out
Leave out your brand story. Models love to wedge it in, and someone asking where their sofa is does not want to hear about craft. Leave out price lists too unless the list is the single dated source of truth; a stale price in a prompt is worse than none.
Step three: build templates with placeholders
A template is the stable half of a reply: structure, tone, the shape of the answer. The model fills the variable half, so it has far less room to invent, because most of each sentence is already decided.
The four parts of a usable template
- An opening line showing the message was read, usually by naming the specific question.
- The answer, from placeholders for facts in your records:
{{first_name}},{{order_ref}},{{ready_date}},{{price}},{{collection_window}}. - The next step, as one question or action, so the customer does not guess what happens now.
- The signature: a real name and a direct number.
Placeholder discipline
Set one rule and enforce it: a placeholder may only be filled from a field that exists in your order system, booking calendar or price list, and if the field is empty the draft fails loudly and goes to a human. Never ask the model to estimate a date or price to fill a gap. The worst replies are not rude; they are confidently wrong about a date.
Store templates where you already work
Templates need no new software. Saved replies in your email client, quick replies in your messaging app, or a text file with one template per section all work. The drafting step is the only place you need a tool, and a general chat assistant is enough as long as the output lands in a draft box, not a send button.
Step four: make human review fast
Every reply passes a human before it leaves. That is not a compromise on automation; it is what makes it safe. The goal is a thirty-second review rather than a five-minute rewrite, checking four things in a fixed order.
- Facts. Name, order reference, date, price, address. Compare to the record, not the draft.
- Commitments. Anything promised: a time, refund, discount, callback. Can you deliver it this week?
- Emotion. Does the temperature match? A light reply to an angry customer reads as mockery; a heavy apology for a trivial question reads as bad news.
- Read aloud. Say the first two sentences out loud. If you would not say them to a person in front of you, edit.
Before and after: one reply, five edits
The example below is invented to show the mechanics, not a real customer. A customer writes: “I ordered the oak shelf on the 12th and was told it would be ready last Friday. Still nothing, and this is the second time I’ve had to chase you.”
The weak draft
“Dear valued customer, thank you for reaching out to us regarding your order. We sincerely apologise for any inconvenience this may have caused. Your satisfaction is our top priority and we understand your frustration. We are currently looking into the matter and will get back to you as soon as possible. Thank you for your patience and understanding.”
The rewrite
“You’re right, it was promised for Friday and it wasn’t ready, and I’m sorry you had to chase us twice. The shelf is built and waiting on its final coat; we’re two days behind in the workshop. It will be ready to collect Thursday after 2pm, or I can bring it to you Friday morning. Tell me which suits and I’ll book it in. — Sam, 07xxx xxx xxx”
Why each edit was made
- “Dear valued customer” became “You”. The greeting is the loudest signal of a mail merge, and a customer who has emailed twice is not a category.
- “Any inconvenience” became “it wasn’t ready”. The first hedges that there may have been none. Naming the failure plainly is what an apology sounds like when a person means it.
- The company’s feelings became the customer’s. “Your satisfaction is our top priority” describes the business; “I’m sorry you had to chase us twice” refers to something that happened.
- “Looking into the matter” became the actual status. Built, needs a final coat. That sentence turns a holding pattern into an answer.
- “As soon as possible” became two named options. The original pushed the next move onto the customer while pretending not to. A specific day plus a choice leaves one decision, not another email.
Note what the model got right: a grammatical, non-defensive reply in seconds, with usable structure. The failure was entirely about content only the owner had, which is why the fix is a system rather than a better prompt. The dates came from the workshop schedule, not the model; an unbuilt shelf would have left the placeholder empty and sent the draft to a human.
Step five: decide the route before you draft
The most useful document you will make is a table saying, per question type, who handles it and how. Decide once and the daily decisions get smaller.
| Question type | Example | Handling | Why |
|---|---|---|---|
| Order or booking status | “Is my order ready?” | Auto-draft, reviewed | Answer is a record field; review catches a wrong date. |
| Hours, address, parking | “Are you open Saturday?” | Auto-draft, reviewed | Static facts you state once. |
| Listed item price | “How much is the small one?” | Auto-draft, reviewed | Price comes from the price list, not the model. |
| Custom quote or trade rate | “Best price for ten?” | Human only | Margin is an owner call; a written number commits you. |
| Complaint or service failure | “Second time I’ve chased you.” | Human only | Needs accountability and a remedy, not a draft. |
| Refunds, credits, terms | “Can I get my deposit back?” | Human only | Money and policy, quotable later. |
| Invoice disputes | “I’ve already paid that.” | Human only | Requires reading the ledger. |
| Legal, safety, health | “Safe for a child’s room?” | Human only, often phone | Compliance statements should not be generated. |
| Delay needing a new date | “When will it arrive?” | Human only | Only a person can commit the workshop or courier. |
| Urgent problem | “Water is coming through.” | Phone | Typing is wrong for anything time-critical. |
| High-value enquiry | “Fitting out three sites.” | Phone, same day | Large jobs are won on a conversation. |
| Bereavement or illness | “We lost my dad, we need to cancel.” | Human only, often phone | No template should reach this message. |
| Supplier pitches | “Quick intro…” | One-line template | No relationship risk; skim weekly. |
The pattern matters more than the rows. Facts about an existing transaction can be drafted. Anything creating a commitment, touching money, or landing on someone’s worst week stays with a human; urgency moves to phone.
What AI should never send unreviewed
Some categories are not “review carefully” but “keep a model away from the send button”.
- Money in any form. Refunds, credits, discounts, payment plans, late fees, debt collection. A wrong figure is a contract you did not mean to sign.
- Admissions of fault or liability. “We damaged it”, written by a model in your name, is a statement you may have to stand behind.
- Promises about timing. Dates, delivery windows, callbacks. The model cannot see the workshop, the rota or the courier.
- Legal, safety, medical or insurance content. Anything reading as advice or a compliance claim.
- Emotionally loaded messages. Complaints, bereavement, illness, disputes: being slightly wrong there is unforgivable.
- Another customer’s information. Name, address, order, complaint. A model working from a messy inbox can paste the wrong one in.
- Public replies. Review responses, social comments, forum threads: read by all, indexed forever.
- Anything you would not want screenshotted. If seeing it on a customer’s feed would embarrass you, it is not ready.
Step six: review the system, not just the replies
Once a week, read ten replies that went out. Not for errors, which the daily review covers, but for patterns: sentences you always delete, placeholders you always correct, questions that keep arriving because a page on your site answers them badly.
Three things worth counting
- Replies sent unedited. Near zero after a month means a vague brief or generic templates. Very high, with replies that feel wrong, means you approve too quickly.
- Customers who wrote back asking the same thing. The honest test of whether a reply answered.
- Drafts thrown away entirely. A rising count usually means that question type belongs in the human-only list.
A tick in a notebook is enough; the point is the trend, not a dashboard.
Your first week
Monday, build the question log from sixty days of messages. Tuesday, write the voice brief. Wednesday, build three templates for the most frequent questions. Thursday, write the handling table and the never-send list and stick them where you work. Friday, run it on one real inbox for an hour: draft every reply, send none unedited.
By the second week you stop writing replies and approve them instead: thirty seconds of judgement rather than five minutes of composition. The customer still gets an answer that sounds like a person, because a person read it before it went.
If you are choosing tools, the constraint to hold onto is that the draft must land somewhere a human approves, not in a queue that sends on a timer. Everything else is preference. More guides on the practical side of small business AI are in the articles index, and the about page explains how corrections and tool limits are handled. If this feels like more overhead than it saves, three templates and a phone may be all you need.
Related: if the phone is the leak rather than the inbox, the AI receptionist guide covers what an answering service can and cannot absorb, and the questions to ask before paying for one.
Next step: once replies are handled, the bigger margin leak is usually quoting. Writing quotes and estimates with AI covers the document itself; the tools directory covers the software around it.
Last reviewed and updated on November 4, 2026.