We use AI in this business and we think it earns its place. That is exactly why the limits are worth stating plainly. Most disappointment with AI in local marketing comes from expecting it to replace judgement rather than to remove delay. The useful question is not whether AI is impressive — it is which specific job you are handing it.
What it is genuinely good at
Replying immediately, at any hour. Speed is the thing most local businesses lose work on, and it is the thing software is best at. An enquiry answered in seconds at nine at night is worth more than a perfect reply on Tuesday.
Asking routine questions consistently. What is the job, where, roughly when, is it a rental or your own home. The same questions every time, in the same order, without getting bored on the fortieth enquiry.
Handling the mechanics of booking. Offering real times, writing the appointment in, sending a confirmation, sending a reminder, handling a reschedule.
Working across channels at once. Text, website, WhatsApp, email and social messages all land in one place and all get answered.
Keeping a record. Every conversation stored and searchable, so the person who turns up has read what was said.
What it cannot do
It cannot judge a job it has never seen. Whether that crack matters, whether the access is workable, whether the customer’s description matches reality — that is a trade judgement made by someone with eyes on it.
It cannot price properly. A price depends on what you find, your current workload, your margin and what you are willing to take on. An AI can repeat a rate you gave it. It should not be inventing figures.
It cannot make a promise on your behalf. Anything about timescales, warranties, regulations or what you will do for free must come from you and be written into what it is allowed to say.
It cannot read a person. A customer who is distressed, angry, vulnerable or confused needs a human being quickly. Detecting that reliably is not a solved problem.
It cannot fix a weak offer. If people are not enquiring because your prices, reviews or service are the problem, answering faster just means finding out faster.
It cannot be accountable. Responsibility for what a system says on your behalf remains yours.
How it goes wrong
The failures have a pattern. The most common is confident invention — a model producing a plausible answer to a question it was never given the facts for. It does not sound uncertain when it does this, which is what makes it dangerous.
The second is scope creep: an assistant set up to book appointments gradually being asked to quote, advise and negotiate because it seemed to cope.
The third is silent drift: prices change, you stop covering an area, a service ends, and nobody updates what the assistant knows. It keeps confidently booking work you no longer do.
The fourth is tone. Relentlessly cheerful replies to a customer whose boiler has failed in February read as insulting.
Example: the reply that should not have been sent
An illustrative scenario. A customer messages a heating business at eleven at night: no hot water, a baby in the house, how much to fix it tonight.
A badly configured assistant answers instantly with a cheerful message about booking a free survey next Thursday and a note that prices start from a figure it half-remembered. Technically it replied fast. It also failed to notice an emergency, quoted something it should not have, and left someone worse off than if nobody had answered.
A properly configured one recognises the words that mean urgency, says plainly that this needs a person tonight, gives the out-of-hours number and flags it to a human immediately. It books nothing and quotes nothing.
The difference is not the technology. It is that someone decided in advance which situations the system must hand over, and wrote that rule down.
Setting the boundary before you switch it on
Decide these before anything goes live, and keep the answers somewhere you can revise.
What it may say. Your services, areas, general process, availability — the facts you have approved.
What it must never say. Prices it was not given, timescales, guarantees, anything about safety or regulation.
When it hands over. Emergencies, complaints, anything about a previous job, anything about vulnerability, and simply being asked to speak to a person.
Who watches it. Someone should read a sample of real conversations every week at first. Not the summary — the actual messages.
How customers know. Being open that an assistant is handling the first reply is both fairer and less awkward than being caught. UK and EU regulators are moving towards expecting disclosure; the ICO’s guidance on AI is the place to start for data protection duties.
Deciding whether AI fits a particular job
Run any proposed use through these questions. If you are answering no to the first three, the job should stay with a person.
- Is the task repetitive and defined, rather than a judgement call?
- Can you write down every fact it needs to do it correctly?
- Would a wrong answer be easy to spot and cheap to correct?
- Is speed genuinely the thing that matters here?
- Have you decided which situations force an immediate handover to a person?
- Is someone reading real conversations regularly, not just the totals?
- Is what it knows — prices, areas, services, availability — kept up to date when things change?
- Are you comfortable with a customer knowing exactly how their first reply was written?
Used within those limits, AI removes delay from a business without removing the person from it. Used outside them, it makes mistakes faster than you can.