AI chatbot development

A support assistant that answers from your content, and knows when to stop

We build chatbots that resolve the repetitive questions, cite the page they answered from, and hand anything sensitive to a person with the conversation attached.

  • Answers limited to approved sources
  • Refusal instead of invention
  • Transcripts in your account
  • Human handoff built in

Written scope and a fixed estimate within two business days. NDA on request.

Support assistant
grounded · 3 sources
Can I change my plan mid-month?
Yes — changes take effect from the next billing date. Anything already invoiced is not re-rated.

Billing policy · section 3.2

Plan changes · help centre

A refund for last month needs a person — I have passed this on with your account details attached.

Channels

Where the assistant lives

One knowledge base, delivered on the channel your customers or staff already use. Start with one channel; add the rest once the answers hold up.

On your website

A support or sales assistant that answers from your docs, captures the details you need, and hands off with the full conversation attached.

Inside your product

An in-app assistant that helps users complete tasks — with awareness of their account, plan and current screen.

In your team tools

A Slack or Teams assistant that answers internal policy and process questions from approved sources.

On WhatsApp and messaging

The same knowledge base behind a messaging channel, with templates and opt-out handled properly.

Voice and phone

An assistant that answers common calls, books appointments and routes anything complex to a person.

Handoff to humans

A clean transfer to email, ticket or live chat when confidence drops or the customer asks for a person.

Behaviour

The six rules that keep a support bot out of trouble

  • Answers restricted to your approved sources, with links shown
  • “I do not know” as a first-class answer instead of a fabricated one
  • Prompt-injection and off-topic handling tested before launch
  • Conversation transcripts stored in your account, not in a black box
  • PII redaction where the transcript leaves your systems
  • A visible route to a human with the context already collected

Delivery

Four steps, with a measured pass rate before launch

The test set is agreed up front. If the assistant fails it, we do not ship it and call it success.

  1. 1

    Pick the job

    Support deflection, lead qualification or internal answers — one job per launch.

  2. 2

    Prepare the knowledge

    We clean and chunk the documents or help-centre content the assistant is allowed to answer from.

  3. 3

    Write the behaviour

    Tone, refusals, escalation rules and what the assistant must never claim.

  4. 4

    Test with real questions

    A set of 50–100 real customer questions, scored for correct answer, grounded answer and safe refusal.

Stack

What we build it with

Models

  • OpenAI
  • Anthropic
  • Self-hosted options

Retrieval

  • pgvector
  • Help-centre sync
  • Doc ingestion

Channels

  • Web widget
  • Slack
  • WhatsApp

Operations

  • Transcripts
  • Analytics
  • Escalation rules

Proof

Live products, real users

Install one of our apps and see how the support flows and edge cases are handled.

FAQ

Chatbot questions, answered plainly

Will it replace our support team?

It should remove the repetitive questions, not the team. In practice that means handling volume around FAQs, order status and policy questions, then handing the rest to people with the context already gathered.

How accurate is it on our content?

Accuracy is measured against your content, not promised in advance. We build a test set from real questions and report pass, partial and unsafe-refusal rates before launch. If the content is out of date, the assistant will be too — so content cleanup is part of the work.

What if it gives a wrong answer?

Answers are limited to retrieved passages and shown with their source, and the assistant is instructed to decline when nothing relevant is found. Anything commercial or irreversible is routed to a person. We also log every answer so wrong ones can be traced to a gap in the content.

Can it take actions, like booking or refunds?

Yes, through scoped API calls with confirmation steps. We treat anything that changes a record or moves money as a separate decision with its own approval rules.

Do we need a new help centre?

No. We can work with your existing help centre, PDFs or internal wiki, provided the content is accurate. If it is not, we will tell you before writing any code.

Request a quote

What should the assistant be able to resolve on its own?

Send us your help centre, your most common questions, and the limits you want enforced. We will tell you what is realistic, what needs content work, and what it costs.

No obligation. We will tell you if we are not the right team for the job.

  1. 1

    Tell us the goal, platforms and must-have features.

  2. 2

    We reply within two business days with scope, approach and a fixed estimate.

  3. 3

    Sign an NDA first if you prefer — nothing about your idea is public.