Should a Small Business Let AI Answer Customer Messages?

Yes, for routine messages with one right answer, such as opening hours, prices, directions, booking changes and order status, provided anything sensitive goes to a person and customers can always reach one. No for complaints, refunds outside your policy, health or safety questions and anything emotional. Most small businesses should start with AI drafting replies for approval, not sending them.

The deciding factor is the cost of a wrong answer, not the number of messages. An AI that gives the wrong parking directions costs you a late customer. One that tells a pet owner a swollen face can wait until Monday, or confirms a refund your policy doesn't allow, costs far more. In a widely reported 2024 tribunal case, an airline was ordered to compensate a customer after its website chatbot made up a refund policy. You own what your AI says.

Follow me on Instagram@sagnikteaches

Sort your last 100 messages before deciding anything

The quickest way to answer the question for your business is to look at what customers actually send. Scroll back through your inbox, social messages and website form until you have 100 messages, and tag each one with a type. It takes about an hour, and you can ask AI to do the first pass if you paste in the messages with names removed.

Connect on LinkedInSagnik Bhattacharya

An illustrative tally for a four-surgery dental practice:

Subscribe on YouTube@codingliquids

For this practice, 32 messages could be answered by AI outright, 40 could be drafted by AI and checked, and 22 must go to a person every time. That split is typical of health businesses: plenty of routine traffic, but a sizeable red group that no amount of tuning should hand over. A shop or café usually has a larger green share.

Three levels of AI involvement, from safest to riskiest

Move a message type up a level only when the evidence says it's ready: for example, fewer than one draft in twenty needing a factual correction over two weeks. Many businesses stay at levels 1 and 2 permanently and still save most of the time, because typing the reply was the slow part, not deciding what to say. What to automate first in AI customer service covers the ordering in more depth.

The red lines, by business type

Every business needs a written list of message types the AI must never answer on its own. The list depends on what can hurt someone in your line of work:

• Dental practice: pain, swelling, bleeding, trauma, anything after treatment that worries the patient.
• Veterinary practice: any symptom, suspected poisoning, breathing difficulty, labour, injury. These need a call-us-now answer, not advice.
• Pharmacy: any question about a medicine, dose, interaction or side effect. Those are for the pharmacist.
• Osteopath or podiatrist: new injuries, numbness, sudden changes, wounds that aren't healing, anything that might need urgent medical attention.
• Hearing-aid shop: sudden hearing loss, pain or discharge from the ear.
• Every business: complaints, refunds or credit outside the written policy, legal threats, requests to see or delete personal data, messages about another person, and anyone who asks for a human.

When a red-line message arrives at night or at the weekend, the AI shouldn't stay silent either. It should send one holding reply, written by you in advance, and alert a person. For example:

Thanks for your message. This needs one of our team, so I've passed
it on and someone will reply by [time] on [next working day].

If your pet is having trouble breathing, has eaten something
poisonous, is bleeding heavily or has collapsed, please call our
emergency line now on [number]. Please don't wait for a reply here.

Handover design is its own topic, covered in when an AI chatbot should hand over to a human.

The knowledge document the AI answers from

Whichever level you choose, the AI's answers are only as good as the written facts it has. Most tools let you upload or paste a document; some read your website instead, which is riskier because old pages linger. Write it plainly, one fact per line, with a date at the top. An illustrative extract from the veterinary practice in the example below:

• Updated: 22 September. Owner: practice manager. Review every month and after any price change.
• Opening hours: Monday to Friday 8.30am to 7pm, Saturday 9am to 1pm, closed Sunday and public holidays.
• Emergencies outside these hours: call the out-of-hours partner on [number]. Never tell a client to wait for us to open.
• Nurse clinics: Tuesday and Thursday afternoons, $32 per appointment (weight checks, nail clips, anal glands). Book online or by phone.
• First vaccinations: puppies from 8 weeks, second dose 2 to 4 weeks later. Link: [booking page]. Do not advise on vaccination timing for adult or overdue pets; pass to a nurse.
• Payment plan: $18 a month for dogs, $15 for cats; includes annual vaccination, flea and worm treatment, two health checks. Full terms: [link].
• Never answer: symptoms, medication doses, whether an animal needs to be seen, complaints, anything about a bill already issued.

Notice the "never answer" lines inside the facts. Putting instructions next to the information they limit is more reliable than a separate rules list the AI may weigh less heavily. Building the FAQ your AI needs covers how to collect these facts from your team.

What AI replies cost on each channel

The price depends heavily on where your messages arrive. Figures as of September 2026:

A quick sum for the dental practice above, if it received 400 messages a month and let AI auto-answer its 32% green share through a per-resolution tool at $0.99: about 128 resolutions, roughly $127 a month. Drafting in the email tools it already licenses would cost nothing extra but still needs a person to send. Which is cheaper overall depends on what staff time is worth to you and how many of those messages arrive out of hours.

A veterinary practice's inbox over six weeks

An illustration with round numbers. A three-vet practice receives about 420 messages a month across email, Facebook, Instagram and a website form. Receptionists spend around two minutes per message, about 14 hours a month, and replies often take three hours or more on busy days.

Weeks 1-2, suggest mode. The practice writes a two-page knowledge document: hours, emergency arrangements, prices for routine services, vaccination schedules, payment plan details, parking, and the red-line list. AI suggests replies in the shared inbox. At first staff edit about 40% of drafts, mostly because the document was missing things (the nurse clinic days, the out-of-hours provider's number). Each gap gets added. By the end of week 2, edits are down to about 15%.

Weeks 3-4, auto-send for five green types. Opening hours, directions, repeat-prescription request acknowledgements, the vaccination booking link and payment plan information go on auto-send, about 38% of volume. Anything containing a symptom word, and anything the AI isn't confident about, goes to a person with the holding reply above.

Weeks 5-6, review. A receptionist samples 20 automated replies a week. Two errors turn up: the AI quoted last year's nurse clinic price (the price list hadn't been updated in the knowledge document) and told a client the practice opened at 8am on a public holiday. Both are fixes to the document, not to the AI.

Illustrative outcome after six weeks: receptionist time on messages down from about 14 to about 7 hours a month, green replies in under two minutes instead of hours, and no red-line message answered by the AI. The bigger lesson is where the errors came from: the knowledge document, every time. Stopping an AI chatbot giving wrong answers is mostly about keeping that document current.

A reply the AI got wrong, and the fixed version

Before the red-line rules were in place, an illustrative pharmacy tested AI drafting on real messages. One customer asked: "Can I take ibuprofen with my blood pressure tablets?" The AI's draft:

Ibuprofen is generally safe for most adults when taken as directed. However, it's always best to check with your doctor if you have any health conditions.

It sounds careful, but it answers a medicine interaction question with a generalisation, which is exactly what a pharmacy must not do. Painkillers of that type can interact with some blood pressure medicines, and whether it matters depends on the person's actual prescription. The fixed rule sends every medicine question to the pharmacist, with this automatic reply:

That's a question for our pharmacist, who can check it against the medicines you take. You can call us on [number] or ask at the counter, no appointment needed. We're open until 6pm today.

A second illustrative case, from a podiatrist's inbox: a patient wrote "I've been waiting three weeks for my insoles and nobody has called me back, this is ridiculous." The AI drafted a cheerful reply with a link to the booking page. Complaints read as routine to a model trained to be helpful. The fix was a rule that any message expressing frustration, however mild, goes to the practice manager, and the AI's only job is to tag it.

AI isn't only a risk in the inbox; for routine messages it often writes a better reply than a rushed human does. An illustrative hearing-aid shop's receptionist, between customers, had answered "do you do free hearing tests and how long does it take" with: "yes we do, book online." The AI draft for the same message, from the shop's knowledge document:

Yes, hearing tests are free and take about 45 minutes, including time to talk through the results. You're welcome to bring someone with you. You can book online at [link] or call us on [number]; we usually have appointments within the week.

Every fact in it came from the document, it answers the time question the receptionist skipped, and it anticipates the next question (can I bring someone?). This is the realistic prize: consistent, complete answers to the questions you get fifty times a month.

Tell customers when they're talking to AI

Meta's messaging policy requires automated chats on Messenger and Instagram to disclose that they're automated where the law requires it, and recommends doing so everywhere. If you sell to customers in the EU, the AI Act's transparency duty to tell people they're dealing with a chatbot has applied since 2 August 2026. It also helps: customers forgive a bot for not knowing something; they don't forgive being misled about who they were talking to.

Wording that works at the start of an automated reply:

Hi, this is [Business]'s automated assistant. I can help with opening
times, prices and bookings. For anything else, or to talk to the
team, just type "person" and we'll pick it up.

For the detail on placement and wording, see what to tell customers at the start of a chat.

Setting it up in the right order

1. Sort 100 messages into green, amber and red (an hour).
2. Write the knowledge document, including the never-answer lines (two to three hours, spread over a week as questions come up).
3. Write the holding reply for red-line messages and the disclosure line (20 minutes).
4. Switch on suggest mode in the inbox you already use, for all channels you can connect (an hour, depending on the tool).
5. Run two weeks of suggestions, adding every missing fact to the document the same day.
6. Move one green type to auto-send, then another a week later, never all at once.
7. Book the weekly sample in someone's diary before you switch anything to auto-send.

Owners who skip to step 6 usually find the problems in step 5 anyway, only in front of customers.

Signs it's working, and signs to pull back

Check these every week for the first two months, then monthly:

• Edit rate on drafts. Falling towards one in ten or fewer means the knowledge document is doing its job.
• Handover rate. A stable share of messages going to people is healthy. A sudden drop may mean the AI has started answering things it shouldn't.
• Reopen rate. Customers replying "that doesn't answer my question" or asking again by phone.
• A weekly sample. Read 20 automated replies against your knowledge document; the routine is laid out in a weekly sampling routine for AI support replies.

Pull a message type back to draft mode if you find a factual error in it twice in a month, if any red-line message gets an AI answer, or if a customer complains about the bot. None of those means AI was the wrong choice; each means a rule or a document needs fixing before you let it send again.

Letting AI reply: common worries

Will customers mind getting a reply from AI?

Most don't mind for simple questions if the answer is quick, correct and clearly labelled, and if a person is easy to reach. What annoys people is an AI that pretends to be human, loops without answering, or blocks them from a real person. Tell them at the start that replies are automated and how to reach the team.

Does the AI need access to my booking or order system?

Only for jobs that depend on live information, such as confirming an appointment time or an order's status. General questions about hours, prices and policies can be answered from a written knowledge document. Start without system access, and connect it later for one job at a time, with read-only access where the tool allows it.

What happens if the AI answers at 2am and gets it wrong?

You are generally responsible for what your business tells customers, whoever or whatever sent it. Limit overnight auto-replies to green-category questions, have every red-line message receive a holding reply with emergency guidance, and check the overnight replies first thing each morning so you can correct any error before the customer acts on it.

Further reads

• Who Is Liable When Your AI Chatbot Gets It Wrong? — Who is responsible when an automated reply gets it wrong.
• Can AI Handle Customer Complaints Without Making Them Worse? — Why complaints need a different approach from routine questions.
• Is It Safe to Let AI Reply to Customers on WhatsApp? — The WhatsApp-specific rules and risks of AI replies.
• Is Instagram DM Automation Safe? What Gets Accounts Banned — What keeps automated Instagram replies within Meta's rules.
• One AI Inbox for WhatsApp, Instagram and Facebook Enquiries — Bring every channel's messages into one AI-assisted inbox.
• How to Train an AI Chatbot on Your FAQs, Policies, and Prices — Feed the AI your prices and policies so answers stay current.
• How Florists Can Prepare for Valentine's and Mother's Day With AI — An eight-week countdown for florists: forecast, pre-order menu, stem maths, cut-off messages, card-message rules and delivery routes.
• WhatsApp Business App vs API: Which Do You Need for AI Replies? — The free app now has Meta's own AI agent; the API lets you bring your own AI and team inbox. How to choose, with a cost sum and the October 2026 changes.
• Can AI Run Your Social Media on Autopilot? What Breaks — Nine things that break when social media runs unattended, an illustrative month's failure log, and the 15-minute daily check that keeps automation safe.
• How Much Does a WhatsApp Chatbot Cost? Fees, Tools and Setup — Meta's WhatsApp fees after the 1 October 2026 change, what platforms and AI replies add, three monthly budgets, and the setup steps that take longest.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Meta Messenger Platform and Instagram Messaging API policy; Meta Business Agent pricing; WhatsApp Business Platform pricing changes from 1 October 2026; Shopify Inbox, Intercom Fin and HubSpot credit pricing pages.