Small businesses tend to land in one of two ditches with AI messaging. Either they automate nothing, and the owner answers "what time do you open" at 11pm on a Sunday for the eight hundredth time. Or they bolt a chatbot onto the site, it confidently tells a customer something untrue, and they rip it out and never try again.
The line between those two is not where most people assume. It has very little to do with how clever the AI is, and everything to do with what it costs you when it is wrong.
The only rule that matters
Before automating any message, ask: *if this response is wrong, what happens?*
- Wrong opening hours → mild embarrassment, customer checks elsewhere. Automate it.
- Wrong price quoted → you either eat the difference or start the relationship with an argument. Automate carefully, with the price coming from your actual data.
- Wrong availability → you double-book, someone drives an hour for nothing. Only automate if it reads live availability.
- Wrong advice on a medical, legal or safety question → serious. Never automate.
- Wrong tone on a complaint → you lose the customer and possibly get a public review about it. Never automate.
This maps cleanly onto what to build, in order.
Tier 1: automate immediately
Questions with one true answer that does not change:
Opening hours, location and parking, what to bring, what is included, prices for standard items, cancellation policy, whether you take dogs, whether beginners are welcome, how to find you.
For most small businesses this is genuinely 60–80% of inbound messages, and answering them has zero value beyond being answered fast. An assistant grounded in your own written policies handles this reliably.
The trick is grounding: the model must answer only from your documents, and say "let me get someone to confirm" when the answer is not in them. A general-purpose model left to improvise about your business will improvise about your business.
Tier 2: automate with a live data connection
Availability, booking status, order status, "did my payment go through", "can I move my booking to Thursday".
These are safe to automate only when the assistant is reading your real system in real time. An assistant guessing at availability is worse than no assistant. One that checks the actual calendar and offers real slots is the single highest-value thing you can build.
That means the messaging layer needs a connection into your booking or CRM system. That connection, rather than the AI, is the actual project.
Tier 3: qualify, then hand over
This is where the money is for service businesses, and it is the piece most people skip.
You do not need AI to close the sale. You need it to work out, before a human spends time, whether this is even a customer: what they want, when, roughly what budget, and — often decisive — whether they are in your service area.
We ran exactly this problem on a solar installation client. Enquiries arrive from all over, but the business only covers certain regions. Every out-of-area enquiry consumed the same human attention as a real one. The fix was a WhatsApp flow that captures the postcode first, checks it against the coverage area, logs it against the contact in the CRM, and routes accordingly — out-of-area gets a polite, immediate answer, in-area gets a human quickly.
Nothing about that requires a clever model. It requires the message to reach the right system. But it removes a large slice of wasted human time, which is what the owner actually wanted.
Never automate these
Complaints. Someone angry needs a human, quickly, and needs to know they got one. An AI apology reads as contempt.
Anything with a safety or medical dimension. "Is it safe for my seven-year-old to surf tomorrow?" is a judgement call about conditions, ability and liability. Do not let a model answer it.
Negotiation and bespoke quoting. Anything where the number depends on judgement.
The moment of delight. The repeat customer coming back for the fourth summer, the couple booking their anniversary, the customer who just left a lovely review. These are the messages that build a business. Automating them is a false economy, and people can tell.
WhatsApp specifically
A few practical realities, because WhatsApp is where most of this happens in Portugal and southern Europe.
Personal WhatsApp cannot be automated properly. You need WhatsApp Business, and for real automation the WhatsApp Business API through a provider like Twilio, 360dialog or similar. This means business verification with Meta, which takes days rather than minutes, and it needs doing before anything else.
Message templates. Business-initiated messages outside a 24-hour window must use pre-approved templates. Reply-within-24-hours is free-form. This shapes what you can build: automated reminders need templates approved in advance.
Costs are per conversation, not per message, typically a few cents. Real, but small.
Handover has to be seamless. The customer must never repeat themselves to the human. If they do, you have made things worse than having no bot.
Say it is a bot. Not legally required everywhere, but it sets expectations and people forgive a bot for things they would not forgive a person. "Hi, I'm the automated assistant — I can answer questions and check availability, and I'll pass you to the team for anything else."
What it costs
- Website assistant grounded in your own content: €1,500–€4,000, plus €20–€100/month running
- WhatsApp automation with API setup and routing: €2,500–€6,000, plus provider and message costs
- Qualification into your CRM with live availability: €4,000–€10,000
The build is not the expensive part any more. Deciding what it must never say, testing it against real awkward messages, and connecting it to your systems — that is what you are paying for. See what we do on custom AI.
The four-week version
Do it in this order.
Week 1: read your inbox. Take your last 200 customer messages and sort them into piles. Most businesses discover that four questions make up half of everything. That is your scope, and it is nothing like what you would have guessed.
Week 2: answer the top pile automatically. Grounded in your own text, with a clear handover for anything else. Keep a human watching.
Week 3: read the logs. Every time it said "let me get someone", ask whether it should have been able to answer. Feed those back in.
Week 4: connect one live system, usually availability. That is the jump from novelty to useful.
Then stop and see what the messages look like. The businesses that succeed at this automate one tier at a time and keep a human in the loop far longer than they think is necessary. The ones that fail try to automate the whole conversation on day one, get burned publicly, and conclude the technology does not work.
It works. It just should not be the one talking to your angriest customer.
Ready to take action?
Let's discuss how we can help your business grow with strategic web solutions and automation.
Start a Conversation