AI chatbots that answer from your documents and hand off to a human
Website and WhatsApp assistants trained on your docs, prices and policies
Look at a week of your inbox and the same eight questions come back. Hours, price ranges, whether you cover a postcode, where an order is. An assistant that answers those from your own documents, and hands anything else to a person with the conversation attached, takes a real load off a small team without pretending to be a colleague.
Assistants that know your business and admit what they do not
The assistants that fail are the ones allowed to improvise. Ours answer from a private index of your own documents, prices, policies and FAQs, and cite where the answer came from. When a question falls outside what it has been given, it says so and passes the conversation to a person along with everything said so far, so nobody has to start again.
We build on Claude or GPT, depending on the job and where your data is allowed to live. Guardrails are agreed in writing: what it may quote, what it must never promise, and how it behaves when somebody is clearly annoyed.
Before anything goes live, we assemble an evaluation set from real questions out of your inbox and chat logs, and score the assistant against it. That score is repeated monthly, because the content behind it changes and accuracy drifts quietly.
Where it runs
Website chat, the WhatsApp Business API, Instagram, or inside your existing help desk. Analytics show what people ask, what got answered, and which questions it could not handle. That last list is usually the most valuable output in the first quarter: it tells you exactly which page on your site is missing. If the answers need to trigger something in a system, that is workflow automation.
What an assistant build includes
Knowledge base & retrieval
Your documents, FAQs, price lists and policies indexed privately, with a sync so an updated PDF reaches the assistant the same day.
Website & WhatsApp channels
One assistant across web chat, the WhatsApp Business API and Instagram, with the tone adjusted per channel.
Human handoff
Escalation to your team or help desk carrying the full transcript, so the customer never repeats themselves.
Guardrails
Scope, tone and the list of things it must never promise, written down, enforced in the prompt and tested adversarially.
Evaluation
A scored test set built from your real questions, run before launch and again every month against the live assistant.
Analytics
Top questions, resolution rate, handoff rate, and the log of questions it could not answer, which doubles as a content plan.
From audit to a live assistant
Question inventoryweeks 1–2
We read a month of your chats and emails and count what actually repeats, rather than guessing at the scope.
Scope & guardrailsweek 3
What it will answer, what it will refuse, when it hands over, and which documents count as the source of truth.
Build & indexweeks 3–6
Retrieval index built from your content, channels connected, and the escalation path tested with your team.
Pilot behind a flagweeks 6–8
Live for a share of real traffic, with every conversation reviewed and the eval set scored before it goes to everyone.
Runongoing
Monitoring, a monthly accuracy review, and the unanswered-question log turned into new content.
From $6,000/project
Builds from $6,000; hosting and model usage from $60/month.
AI assistant, answered
What can an AI chatbot actually answer?
Whatever is documented: hours, pricing, availability, policies, how-to questions, and order status where we can connect to your system. Anything outside that gets handed to a person, which is a design decision rather than a limitation we are apologising for.
How do you stop it making things up?
Answers come only from retrieved passages of your own content, with citations, and the model is instructed to say it does not know rather than fill a gap. Then we test that with an evaluation set, including deliberately awkward questions.
Which AI models do you use?
Claude or GPT for most work, chosen on the task and on where your data is permitted to be processed. Gemini when Google Workspace is already the system of record. We are not tied to a vendor and will move you if the economics change.
Is our data safe?
Your documents sit in a private retrieval index that we host for you, and nothing you provide is used to train a third-party model. The data flows are written down before the build starts, and we sign an NDA before seeing a single file.
How much does an AI chatbot cost?
Builds start at $6,000, which covers the knowledge base, the channels, guardrails, the evaluation set and a pilot. Running costs are roughly $60 a month for hosting and model usage at typical volumes, and you see the actual figures monthly.
How long does it take to build?
Three to six weeks in most cases, plus the two-week readiness audit if you want the wider picture first. The pilot period at the end is not padding; it is where the awkward real questions show up.
Answer the repeat questions automatically.
Send us a month of chat or email history. We will count what repeats and tell you how much of it an assistant could take.