AI chat
Answering at two in the morning, within limits.
An AI assistant on a website is worth having when it captures enquiries outside working hours and answers the same handful of questions your team fields all day. It is a liability when it is allowed to answer anything.
Should a local business put an AI chatbot on its website?
It is worth it if you receive a meaningful number of enquiries outside working hours, or if the same handful of questions consumes your team's time all day. It is not worth it if enquiry volume is low and somebody can answer everything within the hour.
Where it is worth it, three things are non-negotiable: it answers only from a set of answers you approved, it is clearly labelled as automated, and it has a route to a person that somebody actually monitors.
The test for whether something should be automated at all is what a wrong answer costs. Unhelpful is an acceptable risk with a human escalation path. Harmful is not, which is why anything resembling clinical, legal or financial advice stays with a person.
What it is genuinely good at
Capturing an enquiry at eleven at night from somebody who would otherwise have moved on. Answering where you are, what you offer, whether you take a particular insurance, and what the parking is like. Routing a message to the right person. That is the realistic scope and it is worth having.
The value is mostly about timing rather than intelligence. A person who has a question at an inconvenient hour either gets an answer or contacts somebody else, and for many local businesses a meaningful share of enquiries arrive outside the hours anybody is watching.
The secondary value is removing repetition. If your front desk answers the same six questions forty times a day, moving even half of those saves real time, and the questions people ask are logged, which turns out to be more useful than the automation itself.
What it is not good at is anything requiring a judgement about a specific person's situation. That is not a limitation of current technology to be worked around, it is a category of thing that should not be automated.
People are already comfortable asking machines
BrightLocal's 2026 consumer survey, base 1,002 United States adults, measured how people now use generative tools when choosing a local business. It is a vendor study with a stated method.
SourceLocal Consumer Review Survey, BrightLocal, 2026, base 1,002 US consumers
Which changes the reception a labelled assistant gets. The objection that people will be put off by talking to a machine is weaker than it was, provided the machine is honest about being one and can hand over.
How to build one that does not embarrass you
The constraint is the product. An assistant that will answer anything is a demonstration, not a deployment.
Write the approved answers first
A list of the questions your team actually gets, with the answer your business wants given to each. This is the real work, it has to come from you, and the assistant can only be as accurate as this document is.
Define the refusal boundary explicitly
What it must not attempt, and what it says instead. In healthcare that includes anything about symptoms, conditions, medication or suitability for treatment. In legal it includes anything that could be read as advice. The refusal has to be graceful and has to lead somewhere useful.
- A written list of topics it declines
- A specific phrase for declining that does not sound like a fault
- An immediate route to a person or to a booking
Label it clearly
People should know they are talking to an automated assistant. Beyond being the honest thing to do, a labelled assistant that hands over cleanly produces less frustration than one pretending to be a person and failing.
See the remaining steps: How to build one that does not embarrass youHide the remaining steps: How to build one that does not embarrass you
Capture the enquiry, always
The point is not conversation, it is capture. Name, contact details and what they wanted, into the same place every other enquiry lands. A brilliant conversation that captured nothing is a failure.
Put a person on the other end
An escalation that routes to an unmonitored inbox is worse than no automation, because it looks like a response and is not. Somebody has to own it, with a defined window.
Read the transcripts monthly
This is where most of the value actually is. What people ask tells you what your website fails to explain, which service to promote, and what your pricing page should say. It is free research from people who were interested enough to ask.
Everything runs on accounts in your name, and can be switched off without anything else breaking.
What it must never do
In regulated categories these are not preferences. They are the reason the assistant is constrained in the first place.
- Answer a question about symptoms, a condition, medication, or whether somebody is a suitable candidate for a treatment.
- Give anything that could reasonably be read as legal advice, or imply that a relationship has been created.
- Quote a price that depends on specifics it cannot assess.
- Confirm or discuss an existing customer's or patient's details, which is a confidentiality breach regardless of how the question was phrased.
- Respond to a complaint with anything other than an acknowledgement and an immediate handover.
See the full checklist: What it must never doHide the full checklist: What it must never do
- Make a claim about results or outcomes, which is subject to the same substantiation rules as any advertisement.
- Improvise an answer about a service you do not offer, an hour you do not keep or a location you do not have.
- Continue when somebody has clearly asked for a person.
Every one of these is a boundary written into the configuration rather than a hope about behaviour, and every one is agreed with you before anything goes live.
Where the conversation goes
A chat transcript is a record of what a customer told you, sometimes including things they would not have put in a form. Three questions should be answered before a widget goes on a site rather than after.
Where is the data stored, who can access it, and what is it used for beyond serving you. If a vendor cannot answer those plainly, that is the answer.
Read the full breakdown: Where the conversation goesHide the full breakdown: Where the conversation goes
In healthcare this is a formal requirement rather than a preference. A vendor handling patient information generally needs a business associate agreement, and tools that do not offer one do not belong on a page where patients identify themselves. The Department of Health and Human Services has also published guidance on online tracking technologies that affects what may run on those pages at all.
Retention is the part nobody decides. Transcripts accumulate indefinitely by default. A deliberate retention period, and a decision about what gets copied into your CRM and what does not, is worth making at setup.
We build on accounts in your name, document what each connection can access, and set retention rather than leaving it at the vendor's default.
We do not publish a price for this piece of work on its own, because the right scope depends on what already exists. What is published is the bundle pricing: 2,400, 3,600 or 4,800 dollars a month depending on which channels are running. You can read the full breakdown on the pricing page, and you will get an exact number in writing before anything starts.
Chat, forms and a phone number
Chat is an addition rather than a replacement, and adding it badly can reduce enquiries rather than increase them.
| Item | Best for | Watch out for |
|---|---|---|
| Phone number | Urgent and high value enquiries | Nobody answering, which wastes everything upstream |
| Form | Considered enquiries, detail, out of hours | Too many fields, and silent delivery failures |
| AI chat | Quick questions, out of hours capture | Covering the call button on mobile |
| Human chat | High value categories during business hours | Cost, and inconsistency when it is unstaffed |
The mobile point is not trivial. A chat bubble that sits over the call button on a small screen can cost more enquiries than the chat produces, and it is the first thing we check.
How to tell whether it earns its place
Enquiries captured outside working hours that would otherwise have been lost. This is the primary measure and it is the reason most of these are worth having.
Questions answered without escalation, as a share of conversations, which tells you whether the approved answers cover what people actually ask.
Escalation response time, because a handover nobody answers is worse than no handover.
Conversations that produced a booking or a real enquiry, tracked in your CRM rather than in the chat vendor's dashboard.
And the qualitative one that matters most: read the transcripts. If people are repeatedly asking something your website should have answered, the chat has just told you what to fix.
If none of the numeric measures moved after a month, we would recommend turning it off rather than defending it.
How this connects to the rest
This sits inside AI implementation and depends entirely on CRM implementation. A captured enquiry with nowhere to land and nobody following up has made the failure faster rather than fixed it.
It overlaps with marketing automation, and the same principle applies to both: automate timing, not judgement.
It affects site speed, because a chat widget is a third party script. Loading it on interaction rather than on page load is usually the right compromise, and it is covered in website speed optimisation.
The transcripts feed content strategy and conversion rate optimisation, since what people ask a chat widget is a direct list of what the website failed to say.
Count your out of hours enquiries first.
If a meaningful number of people contact you when nobody is watching, this is worth building. If not, we will tell you to spend the money elsewhere.
Straight answers.
Will customers know it is a bot?
They should, and we label it. An assistant pretending to be a person and then failing produces far more frustration than one that is honest and hands over cleanly.
Can it book appointments?
It can offer available times and hand over to your booking system, which is usually the right scope.
Where scheduling depends on a clinical or operational judgement, the assistant should capture the request and let a person place it.
What happens when it does not know the answer?
It says so and routes the person to a human, and somebody has to be on the other end of that route.
An assistant that improvises rather than declining is the single biggest risk in this category, which is why it answers from an approved set rather than from general knowledge.
Is it safe for a medical practice?
Only with a tightly defined boundary and a vendor who will sign the appropriate agreement. Anything about symptoms, conditions, medication or suitability for treatment stays with a person.
A well built assistant in healthcare answers logistics: where you are, what you offer, which insurance you take, and how to request an appointment.
Will it slow down our website?
A chat widget is a third party script and it has a cost. Loading it on interaction rather than on page load removes most of that.
The larger risk is layout: a bubble that covers the call button on a phone can cost more enquiries than the chat produces.
How do we know if it is working?
Enquiries captured outside working hours, questions resolved without escalation, and conversations that became real enquiries in your CRM.
If none of those moved after a month, we would recommend switching it off. A tool that cannot show what it removed is a subscription rather than an improvement.
Where this comes from.
Primary documentation and published research behind the guidance on this page.
- HHS guidance on HIPAA and online tracking technologies (opens in a new tab)
- HIPAA and marketing, U.S. Department of Health and Human Services (opens in a new tab)
- Artificial intelligence, Federal Trade Commission (opens in a new tab)
- NIST AI Risk Management Framework (opens in a new tab)
- WCAG 2 checklist, WebAIM (opens in a new tab)A practical summary of the Level A and AA success criteria.
Talk to the team
A short call, a look at how the business currently shows up, and a straight answer on what we would do first.