Skip to main content

Answers from your own facts

AI Chatbots & Agents

A chatbot is only as useful as what it is allowed to say. We build assistants that answer from your own documents and data and hand over to a person when they should, and agents that can take a simple action such as looking up an order or booking a slot.

What we mean by a chatbot or an agent

Most chatbots that disappoint fail in one of two ways. Some follow a script, so the first question that is off the script ends in “Sorry, I did not understand.” Others are a language model with no limits, and sooner or later they state a price, a policy or a delivery date that nobody at the company ever agreed to.

The useful version sits between those. The assistant answers in its own words, but only from material you have approved: your FAQs, product pages, policy documents and order data. When the answer is not in that material, it says so and offers a person. An agent goes one step further and is allowed to act, for example to check a stock level or create a support ticket, through a narrow connection to your system that logs every call.

The chat in the corner of this page is a small example of the first kind. Its topic buttons are written by hand in our admin, and the free-text box is answered by Gemini using only a page of facts we typed ourselves. It will not quote a price, and when it has nothing grounded to say it points you to a person.

Where it pays for itself is volume. A shop answering “where is my order” all day, or a clinic fielding the same questions about timings and fees, gets hours back each week. A business that gets five enquiries a week does not need one, and we will say so.

Is this the right fit?

Worth building when

  • The same questions reach your team every day by email, phone or WhatsApp
  • The answers already exist in documents, product pages or a database
  • Customers want to check an order, a booking or a status outside office hours
  • Your staff hunt through shared drives for policies they half remember

Not worth it when

  • You get a handful of enquiries a week, where a good FAQ page does the job for free
  • The answer depends on a judgement a qualified person has to make, such as a diagnosis or legal advice
  • There is no approved source material yet. That has to be written first
  • You want a bot to replace your support team outright

What goes into one

Answers from approved sources

Your documents, pages and data are prepared for the assistant, and it answers from them. A question outside that material gets a polite handover to a person.

Handover to a person

On WhatsApp, email or your helpdesk, with the conversation so far attached so the customer does not have to repeat it.

Actions through your systems

For agents: looking up an order, creating a ticket or booking a slot, each through an API with its own permissions and a log of every call.

Guardrails you can read

Topics it must refuse, and a rule that it never states a price or a promise unless that exact line is in its sources.

A review log

Every conversation is kept, so your team can see what people ask, where the assistant fell short, and what to add to its sources next.

How a build runs

  1. Pick the questions

    We start from your inbox and call notes and agree which questions are worth automating. It is usually a short list.

  2. Prepare the sources

    Clean, current answers to those questions, in a form the assistant can use. This step decides the quality more than the choice of model does.

  3. Build and test

    A working assistant on a staging page, tested against real past questions, including the ones it should refuse.

  4. Launch and tune

    Live on your site or on WhatsApp. For the first month we read the review log every week and fix what it shows.

Built with

We pick the model for the job and your data rules. The site chat on this page runs on Gemini.

Part of AI Development

  • Gemini API
  • OpenAI API
  • Claude API
  • PHP
  • Node.js
  • Python
  • PostgreSQL with pgvector
  • WhatsApp Business API

Questions we get asked

Which AI model do you use?

Whichever suits the job and your data rules. The chat on this site runs on Gemini. For a client we try two or three models on your real questions before choosing, because cost and accuracy vary more by task than by brand.

Will it make things up?

It can, if it is allowed to. We restrict it to approved sources, tell it to say when it does not know, and test it with questions designed to tempt it. The review log catches what testing misses.

Can it work on WhatsApp?

Yes, through the WhatsApp Business API. Meta charges for some message types on that channel, separately from our work, and we explain those charges before you commit.

Where does our data go?

Your documents stay in your own database or cloud account. Each question is sent to the model provider to get an answer, so we choose a provider whose data terms you are comfortable with, and switch off training on your data where the provider allows it.

Have you built this for clients before?

Client work in this area is new for us, and we would rather you heard that here than on a call. We built the assistant on this site, and the APIs, databases and integrations underneath are work we do every week. If you would rather hire a team with a portfolio in exactly this, that is a fair choice.

What does it cost?

It depends on how many questions it covers, how many sources it reads and whether it takes actions. After a short scoping call you get a fixed quote for the build and an estimate of the monthly model cost.

Send us the questions your team answers most.

A list copied from your inbox is enough. We will tell you which ones an assistant can handle well, and which are better left to a person.

No sales sequence. One person reads this and replies. Rather give more detail?