Case study · Website customer service

Redita

The AI assistant Alpha Impact built and uses on its own website. It answers with the information the company wrote, collects requests and hands the conversation to the team when needed.

Product
AI assistant for websites, built by Alpha Impact
Where it runs
On alphaimpact.pt, on desktop and mobile
Who uses it
Website visitors, and the team in a back office
What we did
Design, development and hosting
The alphaimpact.pt home page with the assistant open. The visitor asks whether invoices received by email can go into the ERP without being copied by hand; the assistant explains invoice and document automation and leaves links to the solution and to the call. The assistant on a phone. The visitor asks whether it works with the invoicing software they already use; the assistant replies that it depends on the software, lists the integrations already built and suggests a call.
Screens with demo data: the visitors, contacts and conversations are fictitious. The answers are from the Alpha Impact assistant.

The challenge

The same questions every day, many after hours.

An SME's website gets the same questions every day, and many arrive after hours. Generic assistants answer quickly, but they make things up when they do not know and never tell the team who was left without an answer.

Repeated questions
How much it costs, what it includes, whether it works with the software the company already uses, how to book a call.
After hours
Many arrive at night or at the weekend and sit in a form waiting for an answer.
Visitors who give up
Some leave the website before getting an answer, and the team never finds out.
Assistants that make things up
Generic ones answer quickly, but they make things up when they do not know and never tell the team who was left without an answer.

How it works

From the question on the website to the request with the team.

The visitor talks to the assistant on the website. The team sees everything in a back office and joins the conversation when needed.

  1. The company writes down what it knows

    Company details, opening hours, prices and frequently asked questions sit in one place. Each item shows whether the assistant is using it, and what is missing appears on Home.

    The Company tab in Knowledge, with Alpha Impact's name, description and services. Each field is tagged Used by the assistant, and the top summarises how many details and questions the assistant is using.
  2. The assistant answers on the website

    In seconds, in European Portuguese, with links to the right pages. In the conversation, the team sees why it answered that way: what it used, what it cost and how long it took.

    A conversation in the inbox: the owner of a hair salon asks which plan to choose for online bookings. The assistant recommends the Negócio plan and explains why; below the answer are the cost, the model and the response time.
  3. When it does not know, it does not make things up

    It says it does not have that information and offers two ways forward. The contact reaches the team as a request to follow up, with the original question, and the question goes on the list of things to teach the assistant.

    • Leave my contact details
    • Talk to a person
    The list of open requests, filterable by type: requests to talk to a person, a contact to follow up with the question the assistant could not answer, and a new quote request. The assistant on a phone. The visitor asks whether they shoot drone videos for weddings; the assistant explains Alpha Drone's services, says it has no information about weddings and shows the Leave my contact details and Talk to a person buttons.
  4. Guided conversations for quotes and bookings

    Asking for a quote follows set steps: subject, contact details, budget, deadline and consent. Booking a call opens the website's calendar. The team designs each conversation in a visual editor.

    The editor for the Ask for a quote guided conversation: a trigger, the subject buttons, the Ask for a quote action with the Sent, Cancelled and Failed outputs, and the end.
  5. The team joins the conversation

    Whoever is available takes over the live conversation, and the visitor sees who is answering. On the team's side, the customer is on the left and the replies on the right, as in a help desk.

    The inbox with a conversation taken over by the team: a customer says the website's contact form stopped sending emails, the team finds the problem and the customer confirms the test email arrived.
  6. Test before changing

    In the lab, the team talks to the assistant as a visitor would and sees each step of the answer. Changes in behaviour only reach the website once they are published.

    The lab: on the left, the scenario (after hours, nobody available); in the middle, a question about plans for a physiotherapy clinic and the answer; on the right, the diagnostics with each step, the model, the cost and the response time.

Screens with demo data: the visitors, contacts and conversations are fictitious. The answers are from the Alpha Impact assistant.

The back office

The team sees on one screen what needs its attention.

Home in the Redita back office. On the left, what needs the team: a person waiting, a contact to follow up, a new quote and an unanswered question, each with a button to deal with it. On the right, the checklist of what must be ready before going public.
Home
Who is waiting, the contacts to follow up, new quotes and what is missing before going public.
Conversations
Live conversations and history, with who answered: the assistant or the team.
Requests
Contacts to follow up, quotes and requests to talk to a person, each with its status.
Knowledge
Company details, opening hours, prices, frequently asked questions and documents.
Statistics
What visitors ask and how conversations end.
Privacy
Explicit consent in forms, retention periods and requests to access or erase data.

The assistant's rule

It answers only with what the company wrote.

Every answer starts from the details, frequently asked questions and documents the team wrote. Frequently asked questions have a fixed answer, with no AI cost, and every answer is corrected into European Portuguese.

When it does not know, it says so. It offers to take contact details or put the visitor through to a person, and the question reaches the team so it gets an answer from then on.

Technology

Light on the website, protected behind it.

Widget
Light and isolated from the website, with a notice that you are talking to an AI and a link to the privacy policy.
AI
Anthropic's Claude models, answering only from the company's knowledge.
Safeguards
Against manipulation attempts, tested with simulated attacks.
Hosting
Node.js, React and MySQL on a dedicated server with Docker and HTTPS, with backups.

Result

The website answers the usual questions at any hour, with the information the company itself wrote. What the assistant does not know reaches the team with the question and the contact details, and then gets an answer.

Solution: AI assistants

Start with a 30-minute call.

Tell us which task takes up most of your team's time. We reply with a concrete idea and, if it makes sense, a fixed-price proposal.

Book a 30-minute call