When does a website AI assistant make sense?

An assistant makes sense when it answers recurring questions or guides a visit using information the business can keep up to date. If the problem is that essential information is hard to find, first improve how the page explains it.

Pick the questions the business actually receives and define what a good answer or next step looks like in each case. A normal month of calls and emails is a good starting point for that list.

Website wireframe on paper with content blocks and an orange pen
A paper wireframe makes the visitor journey easier to review before deciding how the website will look.Photograph by Compagnons / Unsplash

Criteria for deciding whether you need one

Before commissioning an assistant, go through these points with your team. If two or more fail, it is probably better to fix the website or customer service first. An assistant amplifies what you already have, whether it is in order or not.

  • Some questions come up again and again, and the answer can be kept up to date.
  • Someone in the business will review the sources and answers regularly.
  • There is a route for handing the conversation to a person, with known hours.
  • The assistant can do its job without asking for sensitive data.
  • The chat window can say it is an AI system, as Article 50 of the EU AI Act requires.

Can the assistant replace website content?

No. Services, limits and contact routes should be understandable without a conversation with the system. The page is what people who do not want to chat read, what Google indexes and what remains if the system fails. The assistant adds one more way to find it.

Define what it can use to answer

Information needs a source and someone responsible for updating it. When the system does not have an answer, it should say so and offer another route. If the assistant gives old opening hours or conditions, check who is meant to review that information.

  • Authorised sources and information outside the scope.
  • Responsibility for updating services, conditions and content.
  • The data needed for each function.
  • Requests that should be referred to a person.

Does it have to say it is an AI?

Yes. Article 50 of the EU AI Act says people must be informed that they are interacting with an AI system, unless this is obvious, and that the information is given clearly and at the latest at the first interaction. A name that does not mislead and a notice in the chat window take care of it.

Want us to look at your case? Leave us your details and we’ll tell you where we would start.

Let's talk

A worked example with a clinic in Girona

Take a physiotherapy clinic in Girona, a made-up example with no client data. The recurring questions are clear. Opening hours, how to book, which insurers are accepted, what to bring to the first appointment and how to get there. The assistant sticks to those answers and takes them from the pages the reception team maintains.

When someone describes pain or asks for advice about a treatment, the assistant does not assess it. It explains that the physiotherapist will do so at the first appointment and offers to book or to connect them with reception. It does not collect health data either. To book, it sends the person to the booking system, which asks for a name and phone number.

Test the difficult questions too

Do not test only the questions that already have a prepared answer. Try what happens when a fact is missing, when someone asks for something outside the scope or when they ask it to do a different job. OpenAI recommends testing an application with input designed to break it.

Review the journey as well as the content. People need to understand what happened, how to correct something and how to continue if the assistant did not help. OpenAI also recommends having a person review outputs wherever possible.

Coloured sticky notes arranged on a paper planning board
Making decisions visible helps separate ideas from priorities and share what needs to be reviewed.Photograph by Walls.io / Unsplash

What data can the assistant ask for?

The minimum needed for the task. The Spanish data protection authority (AEPD) advises users to check that a chatbot offers a privacy policy identifying the data controller, clear data protection information and whether it keeps learning from conversations. Do not collect health or banking data in the chat.

Checklist before going live

Once the assistant answers the test questions well, the part usually left until last remains. Who monitors it, how errors are reported and when the sources are reviewed. Go through it with your team before switching the assistant on.

  • The chat window says it is an AI assistant before the first reply.
  • The privacy policy explains who is responsible and what happens to the conversations.
  • Handover to a person has been tested from a phone and reaches the right people.
  • Unanswerable questions, out-of-scope requests and attempts to change its function have been tested.
  • There is a simple way to report a wrong answer and someone reads it.

How will we know it works once it is live?

By reading conversations. Set aside time each week to review a sample, note the unanswered questions and check that handovers reach the team. OpenAI recommends keeping human review and a way for users to report problems. Those notes decide what to update on the website.

Look at a demonstration critically

Ona on the crida studio website is our own assistant demonstration. It shows one way of interacting and a tone of voice, but each business needs its own sources and use cases defined before copying anything.

Before launch, settle maintenance, usage costs, responsibilities and review criteria. Usefulness shows in specific tasks. Ask whoever gives the demonstration to try one of your real questions and one that has no answer on your website.

Add crida studio as a preferred source on Google