Automate the frequent, low-risk jobs first. For most small businesses that means sorting and tagging incoming messages, then AI-drafted replies to routine questions that a person approves, then proactive messages that stop questions arriving. Let AI answer customers directly only after its drafts have proved reliable. Complaints and refunds stay human.
Most owners start at the wrong end. The obvious first move looks like a chatbot on the website, because it's the most visible. But a chatbot answers customers without anyone checking, so it needs the most preparation and does the most damage when it's wrong. Drafted replies give you most of the time saving in the first month, and they teach you which answers the AI gets right before you let it speak for you.
Sort a month of messages before choosing anything
Every recommendation below depends on what your customers actually write about. Export or scroll back through one month of email, website chat, social messages and booking-platform messages, and count them by type. Add a rough time per message. An hour with a spreadsheet is enough.
Here is the result for an illustrative 22-room boutique hotel with one guest-relations person and a hands-on owner:
| Message type | Per month | Minutes each | Hours a month |
|---|---|---|---|
| Pre-arrival questions (parking, check-in, dogs, dinner times) | 210 | 4 | 14.0 |
| Availability and booking requests | 120 | 7 | 14.0 |
| Changes and cancellations | 70 | 6 | 7.0 |
| Invoice and receipt requests | 45 | 3 | 2.3 |
| After the stay (lost property, feedback) | 40 | 5 | 3.3 |
| Complaints | 25 | 20 | 8.3 |
| Wedding and event enquiries | 30 | 15 | 7.5 |
| Spam and sales pitches | 100 | 0.5 | 0.8 |
| Total | 640 | 57.3 |
Two things stand out that the owner hadn't expected. Pre-arrival questions take as much time as booking requests, even though each one is quick. And complaints are only 4% of messages but take 15% of the time. The first is a prime target for AI; the second is where AI should help the person handling it, not replace them.
Score each message type on volume, stability and risk
Give each type a score from 1 to 3 on three questions, then multiply. The highest products are where to start.
- Volume: 3 if it's in your top three by hours, 1 if it's rare.
- Answer stability: 3 if the answer is the same for everyone and written down somewhere; 1 if it needs a judgement each time.
- Low risk: 3 if a slightly wrong answer costs a follow-up message; 1 if it costs money, safety or a customer.
For the hotel: pre-arrival questions scored 3 × 3 × 3 = 27. Invoice requests 2 × 3 × 3 = 18. Booking requests 3 × 2 × 2 = 12, because availability changes and a wrong confirmation is costly. Complaints 2 × 1 × 1 = 2. Event enquiries 2 × 1 × 2 = 4, since every wedding is different. The ranking that falls out is close to the order below, and it will be for most businesses, but check yours: a campsite may find "directions and arrival times" at the top, and a members' club "membership renewal and card problems".
Same scoring, different first job: a members' club and a guest house
The order above is typical, not fixed. Running the same count and scoring at two other businesses shows how the first job can change.
A hypothetical members' club with 900 members and a part-time secretary counted 380 messages in a month. The biggest group, at 110, was "my card won't open the door" or "I can't log in to book a court". Those answers aren't stable at all: each needs someone to look up the member's record and reset something. The top scorer instead was subscription questions ("when is my renewal due?", "can I pay monthly?"), at 3 × 3 × 3. So the club's first automation was drafted replies for subscription questions, and its second was a fix to the booking app's password reset, which removed most of the login messages without any AI.
A six-room guest house run by a couple, in a second invented example, had only 150 messages a month, a third of them via a booking platform's messaging. With so few messages, sorting added little. Their first job was drafted replies for the platform messages, where the same six questions made up almost everything, and a saved project in their chat assistant holding the house rules and local recommendations. At 150 messages, drafting alone saved them about four hours a month, which they judged worth the $20 personal plan and not worth a help desk.
The lesson from both: count first, and let the numbers pick the job. If your biggest message type is caused by a broken process, fix the process before automating the replies to it.
First: sorting and tagging what comes in
The lowest-risk automation is one customers never see. AI reads each incoming message and labels it: booking request, pre-arrival question, complaint, invoice, spam. It can also set a priority and send the message to the right person or folder. Nobody gets a wrong answer, because the AI isn't answering anyone.
Why do this before anything else? Because every later step depends on it. You can't send pre-arrival questions to drafted replies, or complaints straight to the owner, until messages are labelled. And the labels give you a running count of each type, which is how you'll measure every later change.
For the hotel, the direct saving was small: spam cleared automatically saved under an hour a month. The real gain was that the 25 complaints now reached the owner within minutes instead of sitting under a pile of parking questions. Most help desks and several email tools can do this; the method is covered in AI ticket triage, and for a shared inbox handling sales and invoices too, in AI email triage for shared inboxes.
Second: AI-drafted replies a person approves
This is where most of the early time saving comes from. The AI writes a reply to each routine message using your FAQs and policies; a person reads it, fixes anything wrong and sends it. It's much faster than writing from scratch, and every draft is checked.
An illustrative guest email and the draft that came back:
Guest: "Hi, we're arriving around 11pm on Friday, is that OK? Also can we bring our dog, and is there parking?"
AI draft: "Thank you for your message. Our 24-hour reception will be happy to welcome you at 11pm. Dogs are welcome in our garden rooms for $25 per night, and we have free on-site parking for all guests. We look forward to seeing you on Friday."
Two errors. The hotel's reception closes at 10pm; late arrivals collect keys from a code box, with the code sent by text on the day. And parking is free but limited to eight spaces, which matters on a full Friday. The dog charge was right because it was in the FAQ. The fixed version took 40 seconds to edit.
The prompt behind drafts like this is short. The important parts are the source rules and the flag for missing information:
Draft a reply to this guest message.
Use ONLY the facts in our FAQ and policies below. If the guest asks something
they don't cover, write [CHECK: question] instead of answering it.
Tone: warm, brief, no exclamation marks, sign off as "The team at [hotel]".
Guest message: [paste]
FAQ and policies: [paste or attach]For the hotel, the arithmetic is simple: pre-arrival replies dropped from about 4 minutes to about 1.5, saving 210 × 2.5 minutes, or nearly 9 hours a month. Most help desks now have a draft-reply button; a chat assistant on a business plan with your FAQ saved in a project works too.
Third: messages that stop the questions arriving
The cheapest customer message to answer is the one that never arrives. Once your sorting shows which questions come in most, send the answers before people ask. For the hotel, that meant a pre-arrival email three days before each stay covering parking, late arrival and the key box, dogs, and dinner times.
AI's role here is small but useful: writing a clear version of the message, and adding one line personalised to the booking, such as "We see you're arriving with a dog; the garden rooms have a door straight onto the lawn." The sending itself is usually a scheduled message in the booking or property-management system, not AI at all.
The effect: pre-arrival questions fell by about 30%, from 210 to about 147 a month. Combined with drafted replies, the job went from 14 hours a month to under 4. Invoice requests went the same way once the booking system emailed a receipt automatically at check-out: they fell from 45 a month to about 10, with no AI involved at all. Some of the best "AI customer service" wins turn out to be a setting you already had.
A holiday-let manager gets even more from this step, because guests arrive at unstaffed properties. Take a manager with 15 cottages as an illustration: one arrival-day message with the key-safe steps, wifi name and heating instructions could cut arrival-evening calls by about half. The key-safe code goes in that private message, never in anything public. If your repeat questions need fuller answers, turning support tickets into help centre articles shows how to write them up from your own message history.
Fourth: letting AI answer routine questions on its own
Only now does a website chatbot, or an AI agent answering email directly, make sense. You have labelled messages, a month of checked drafts showing which answers the AI gets right, and FAQs corrected along the way. The rule for this step: let the AI answer, unsupervised, only the question types where its drafts needed no fixes for several weeks in a row.
Take a hypothetical campsite: its chatbot answers questions about pitch sizes, electric hook-up, dog rules, shower block hours and the nearest shop, and hands everything else to the office. It doesn't touch bookings or complaints. The owner set it live only after four weeks of drafted replies showed those five topics needed no corrections.
Pricing for AI that answers on its own varies by how it counts:
- Intercom Fin charges $0.99 per resolved outcome.
- HubSpot's Customer Agent needs a seat on any Professional or Enterprise hub (usually Service Hub) and uses 50 HubSpot Credits per resolved conversation, about $0.50; Starter plans aren't eligible.
- Shopify Inbox now includes a free AI agent for stores on Basic or above using new customer accounts. It can use web search as a secondary source, so test it on your own policy questions before trusting it.
Two rules protect you here. Tell customers they're talking to AI at the start of the chat; if you sell to customers in the EU, the AI Act's transparency duties require it. And set clear points where the AI stops and passes to a person, which when a chatbot should hand over to a human covers in detail.
Fifth: actions such as bookings, changes and refunds
An AI that can change things, such as creating a booking, moving dates or issuing a refund, needs a connection to your systems and much tighter limits. Treat it as a separate project, after the first four steps are working.
A sensible order within this step is: sending links that let customers act themselves (a "change your booking" link), then read-only lookups ("your booking is for two nights from the 14th"), then simple changes within fixed rules, and last of all anything involving money. Many small businesses never need the last stage. A members' club might happily let AI answer "when does my membership renew?" from the membership system while every refund or fee waiver still goes to the secretary.
What stays with people, and how AI still helps there
Some messages should never be answered by AI on its own, however good it gets: complaints, refund and compensation requests, anything involving safety or health, sensitive personal situations, and high-value enquiries such as weddings. AI still helps the person handling them:
- Summarising a long complaint thread into three lines before the owner reads it in full.
- Pulling the facts together: booking details, previous messages, what was promised.
- Drafting a first reply that the owner rewrites in their own words.
For the hotel, summaries and first drafts brought the average complaint from 20 minutes to about 12, without the guest ever reading an unedited AI reply. The detail of doing that without making things worse is in whether AI can handle complaints.
A 90-day order of work
For the hotel, the sequence ran like this. It's a reasonable template for most small teams:
- Weeks 1-2: count a month of messages by type; write or tidy the FAQ and policies (parking, arrival, pets, cancellations, invoices).
- Weeks 3-4: switch on sorting and tagging; check the labels daily for a week, then weekly.
- Weeks 5-8: drafted replies for the top two message types; keep a note of every fix made to a draft.
- Weeks 7-10: write and schedule the pre-arrival message; watch whether the matching question type falls.
- Weeks 11-13: if any question type needed no draft fixes for three weeks, trial AI answering that type directly, on one channel, with a clear handover.
At the end of 90 days, the hotel had cut about 16 hours a month from its 57, mostly from pre-arrival questions, complaint handling time, invoices and spam. Booking requests, events and complaints were largely unchanged in time but faster to reach the right person. That's a realistic result: the savings come from the dull, repetitive messages, which is exactly why they're first.
Misreadings that send small businesses the wrong way
- "AI customer service means a chatbot." A chatbot is one channel and usually the fourth step. Sorting and drafting save more time sooner.
- "We'll need fewer staff." In a small team, the more likely result is that the same people spend their time on bookings, events and guests in the building instead of typing parking directions. Plan for that rather than for a cut.
- "The AI learns our policies by itself." It only knows what you give it. If your cancellation policy lives in one person's head, write it down first.
- "Once it's set up, it runs itself." Prices, times and policies change, and the AI will cheerfully quote the old ones. Somebody needs to update the FAQ and check a sample of AI replies every week; the weekly sampling routine takes about 20 minutes.
One number per step, checked before the next begins
Pick one number per step and compare four weeks before with four weeks after:
| Step | Measure | Hotel's result (illustrative) |
|---|---|---|
| Sorting | Time from complaint arriving to owner reading it | From about a day to under an hour |
| Drafted replies | Minutes per routine reply; share of drafts needing fixes | 4 min to 1.5 min; fixes fell from 40% to 12% |
| Proactive messages | Count of the question type they target | 210 to about 147 a month |
| AI answering alone | Resolved without handover; customer complaints about the bot | Trial only; zero complaints in 3 weeks |
If a step doesn't move its number within a month, fix it or drop it before adding the next. Stacking a chatbot on top of a messy FAQ and unlabelled inbox is how small businesses end up with AI customer service their customers dislike.
Further reads
- AI Chatbot vs Live Chat vs Help Desk: What a Small Team Needs — Choose the channel before choosing the AI.
- Your First 30 Days of AI Customer Support for an Online Shop — A month-one plan if you run an online shop.
- Should You Build or Buy an AI Chatbot for Customer Service? — Decide whether a chatbot should be bought or built.
- How Much Does Customer Service Software With AI Cost per Agent? — What support software with AI costs per person.
- How to Hire Your First Customer Service Person Alongside AI — When AI isn't enough and it's time to hire.
- How to Offer Multilingual Customer Support With AI Translation — Add translation once the basics are working.
- AI Use Cases by Department for Small Businesses (With Examples) — 21 practical AI use cases across seven departments, each shown in a small insurance brokerage, with how to start and what to watch.
- Should a Small Business Let AI Answer Customer Messages? — Sort your last 100 messages, pick the right level of AI involvement, set red lines by business type, and know what each channel costs per reply.
- How to Write Reply Templates That Keep AI Replies On-Script — Rebuild your canned responses so AI fills them in without drifting: locked lines, marked slots, never-say lists and a clear rule for when to hand over.
- Outcome-Based AI Pricing: Paying per Resolution, Task or Result — How per-resolution and per-task AI pricing works, why each vendor's definition of resolved changes the bill, and how to check what you're paying for.
- AI or Outsourced Customer Service: Which Costs Less? — A bike shop's 380 monthly calls and messages, costed with a human answering service, an AI first line and a cheaper third route most owners miss.
- What Can AI Realistically Do for a Small Business? 15 Tasks — Fifteen jobs AI can realistically take on in a small business, rated by how much you can hand over, with tools, costs and the catch for each.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: facts on Intercom Fin, HubSpot Customer Agent and Shopify Inbox pricing and eligibility from the vendors' pricing and knowledge-base pages, checked September 2026.