Only deploy one that answers from your own documents and refuses everything else. An unconstrained model speaking for your business will make commitments you have to honour.

The obvious idea

You have a tool that writes fluent answers and customers who ask questions at all hours.

Connecting the two is the first thing almost everybody thinks of, and several suppliers are now selling exactly that.

It is a genuinely appealing proposition for a small business that cannot staff a chat window.

It is also the highest-risk use of these tools available to you, for a specific reason.

What it will say

None of these is a malfunction. The tool produces plausible answers, and a plausible answer to what does this cost is a number, so it supplies one. It has no way to distinguish between knowing and sounding right.

The difference from everything else

Which is why this deserves separate treatment.

Everywhere else these tools are used, a person reads the output before anybody else sees it.

A chat widget removes that step entirely: the output goes straight to a customer, in your name, without review.

So every checking habit that makes the other uses safe is unavailable here by design.

That is the whole of the problem, and no amount of careful prompting resolves it.

The only workable answer is constraining what it can say rather than reviewing what it did say.

What you are bound by

A question worth asking before installing anything.

A customer who is quoted a price by something on your website will reasonably expect that price.

Whether you are legally bound depends on the circumstances and is not a comfortable thing to be arguing about.

Practically, you either honour it or you have a dispute with somebody who has a screenshot.

The same applies to statements about warranties, timescales, and what is included.

Assume you will honour whatever it says, and design accordingly.

The version that works

Since this is not an argument against all of it.

Constrain it to answer only from documents you supply: your pages, your prices, your frequently asked questions.

Instruct it to say it does not know and offer contact details rather than answering from general knowledge.

Give it no ability to make commitments, quote outside a published range, or discuss anything regulated.

And put a visible line saying it is automated, with a route to a person.

That is a considerably less impressive product and it is the one that will not cost you.

Ask any supplier specifically whether their product answers from your documents only, since the answer separates the safe products from the rest.

A worked example

A business installed a chat tool that answered from general knowledge as well as their site.

Within a fortnight it had told a customer that they offered emergency call-outs, which they had stopped two years earlier, and quoted an hourly rate roughly a third below what they charge.

The customer arrived expecting both.

They honoured the rate on that job and removed the widget the same day.

The replacement answers only from a document of eleven questions and hands everything else to a form.

It is far less clever and has produced no incidents in the months since, which was the whole objective.

Read the transcripts weekly

Which is the ongoing obligation people skip.

Whatever you install, somebody needs to read what it actually said, at least weekly at first.

That is where you find the invented answers, the questions it handled badly, and the enquiries it lost.

It is also the most direct record of what customers actually ask, which is worth reading regardless.

Set aside twenty minutes a week for the first two months, and reduce it once you know how it behaves.

A widget nobody has read the transcripts of is a widget nobody knows the behaviour of.

Consider what it competes with

The comparison that decides whether to bother.

An instant, slightly limited answer at ten at night beats a form nobody replies to until Tuesday.

It does not beat a telephone number somebody answers, which for most trades is what customers actually want.

So the case is strongest where enquiries arrive outside hours, where the questions are repetitive, and where the alternative is silence.

And weakest where a call would have happened anyway, since the widget can absorb somebody who was about to phone.

Check your call volumes for a month after installing one, since a fall there against no rise in enquiries is the answer.

Watch what it does with personal information

A second exposure that sits alongside the accuracy problem.

Customers type things into chat windows they would not put in a form: their address, their situation, occasionally their health or their finances.

That text is being sent to whichever service powers the widget, under terms you agreed to and they did not, which is the same disclosure question as pasting anything else in.

Check where the transcripts are stored, how long for, and whether your privacy notice mentions any of it.

And ask the supplier directly whether conversations are used to improve their product, since the answer belongs in that notice.

The counter-case

Constrained versions genuinely work.

Answering opening hours, areas covered, and what a service includes, from your own text, handles a large share of enquiries reliably.

The technology is also improving quickly, and products designed for this are considerably safer than connecting a general tool to your site.

And a business losing enquiries overnight has a real problem that this addresses.

Constrain it to your own documents, forbid commitments, label it clearly, and read the transcripts weekly.

Before installing one

  1. Ask whether it answers from your documents only.
  2. Forbid prices outside a published range.
  3. Forbid commitments and guarantees.
  4. Require it to say it does not know.
  5. Label it as automated.
  6. Give a route to a person.
  7. Read the transcripts weekly.

Step one is the question that separates the products, since anything answering from general knowledge will eventually describe a business that is not yours.

Whether to have chat at all is covered in live chat on a one-person business.


Frequently asked questions

Why is this riskier than other uses?

Because everywhere else a person reads the output first. A chat widget sends it straight to a customer in your name, so every checking habit is unavailable by design.

What will it get wrong?

A price you do not charge, a timescale you cannot meet, an area you do not cover, a service you dropped, or a guarantee nobody authorised.

Why does it do that?

It produces plausible answers, and a plausible answer to what does this cost is a number. It cannot distinguish between knowing and sounding right.

Am I bound by what it says?

Whether legally depends on circumstances. Practically you either honour it or dispute it with somebody holding a screenshot, so assume you will honour it.

What version is safe?

One constrained to answer only from documents you supply, instructed to say it does not know rather than answer generally, with no ability to make commitments.

What is the ongoing obligation?

Reading the transcripts weekly at first. That is where you find invented answers and lost enquiries, and it is the most direct record of what customers ask.

West Coast Media Solutions Inc. provides web design, web development, hosting, digital marketing, and business consulting to organisations across Canada, drawing on more than twenty-five years in the field.

Considering a chat widget that answers questions?

Ask the supplier whether it answers from your documents only. That single question separates the safe products.

Start a Conversation