AI & Automation

AI chatbots that answer from your own documents

Retrieval-augmented assistants with citations, confidence thresholds and clean human handover.

A chatbot is only useful if you can trust what it says. The scripted kind cannot answer anything unanticipated; the ungrounded kind answers everything, including things that are not true — and an invented refund policy is a liability, not a feature. We build the version in between: retrieval-augmented assistants that answer only from your own material, cite the source on every answer, decline when confidence is low, and hand over to a person with the full conversation attached. The result is an assistant safe enough to put in front of customers, and a useful side effect — a running report of every question it could not answer, which is precisely the list of what your documentation is missing.

The problem

What this solves

Support teams answer the same forty questions endlessly while genuinely difficult cases wait. Scripted bots deflect without resolving, which customers correctly experience as being blocked rather than helped. And the fear that stops most businesses adopting AI support is entirely rational: a system that speaks confidently about your pricing, policies or capabilities without being grounded in them will eventually say something you have to honour.

Bots that invent answers

Ungrounded models stating policy your business then has to honour.
Scripted trees that deflect

Menus that block customers rather than resolving anything.
The same forty questions, forever

Skilled support time consumed by what a document already answers.
Knowledge that drifts

An assistant confidently quoting last year's pricing.
No route to a human

Customers trapped with a machine that cannot help.

Our approach

How we go about it

Retrieval before generation. Your documents are chunked, embedded and indexed; every question retrieves the relevant passages, and the model answers only from them with the source attached. Below a confidence threshold it declines and escalates rather than guessing. Re-indexing runs when documents change, so the assistant and your actual policies never drift apart — the failure mode that turns a helpful bot into a liability six months after launch. We measure on resolution and satisfaction, never on deflection rate. Optimising for deflection is how businesses build phone trees that customers hate.

Outcomes

What changes for you

Grounded in your content

Answers generated only from your documents, with the source cited.
Knows when to stop

Below its confidence threshold it declines and escalates.
Handover with context

The agent receives the whole conversation; the customer never repeats themselves.
Web, app and WhatsApp

One assistant, reachable wherever your customers are.
A documentation gap report

Unanswered questions become a prioritised list of what to write next.

Frequently asked questions

Common questions about ai chatbot development.

How do you stop it making things up?

By construction, not by prompting. The model may only answer from passages retrieved out of your own documents, every answer carries its source, and anything below a confidence threshold is declined and escalated instead of guessed.

What do we need to provide?

Whatever you already have in writing — policies, product pages, a support inbox export, FAQs. We audit that material first, because the assistant can only ever be as good as what it is grounded in.

Can it run on WhatsApp?

Yes, through the official WhatsApp Business API, alongside web and in-app deployment. The same grounded assistant, reachable wherever your customers already are.

Which model do you use?

Whichever fits the accuracy, latency, privacy and cost requirements of your case — and we will explain the trade-off rather than defaulting to whichever is fashionable. For sensitive data we will discuss options that keep processing under your control.

What does it cost to run?

There is a build cost and a running cost — the running cost scales with conversation volume and the model chosen. We size both before you commit, because an assistant with a surprising monthly bill gets switched off.

Related services

Want an AI assistant you can actually trust?

Send us your documentation. The quality of the assistant is decided by that material — we will audit it honestly before quoting.

Start a project Talk to our team