AI agents · support practice

Chatbots that answer from your docs,
not from guesswork.

Support and sales bots grounded in your own content, with citations, human escalation and an audit trail. We measure answer quality before launch, so you can prove it works instead of hoping.

Scope a chatbot See RAG development

What it is

What chatbot development actually involves

A chatbot is easy to demo and hard to trust. The demo answers three questions well. Production has to handle the messy ones: the customer who asks about a policy you changed last month, the one who pastes an order number, the one asking something the bot should refuse.

The work is the grounding and the guardrails, not the chat window. We connect the bot to your real content, restrict what each user can retrieve, define what it must refuse, and build an evaluation set so answer quality is a number you can watch rather than a feeling.

What we build

Bots that hold up in production.

Every build ships with retrieval you can inspect, evals you can rerun and a switch that stops it.

Support bots

Grounded in your help centre, policies and past tickets, with a citation on every answer and a clean handoff to a human when confidence drops.

Cited answersHuman escalationTicket deflection reporting

Sales and onboarding bots

Qualify, answer product questions and book the call. Connected to your CRM so the conversation does not die in the widget.

CRM write-backLead qualificationBooking handoff

Internal knowledge bots

For your own team, over docs, runbooks and wikis, with per-user permissions so nobody retrieves a document they could not already open.

Per-user access controlSlack and TeamsRunbook retrieval

Guardrails and refusals

Written rules for what the bot must not answer, plus prompt-injection handling on any tool it can reach.

Refusal policyInjection handlingPII redaction

Evals and monitoring

A golden question set agreed with you up front, scored on every change, with traces for every conversation.

Golden setRegression scoringFull traces

Channels

Web widget, WhatsApp, Slack, Teams or your own app. Same brain, different front doors.

Web widgetWhatsAppSlack and Teams

Built on

The stack we reach for.

Chosen per brief. If your team already runs something else, we work in it.

ClaudeGPTOpen-weight modelspgvectorQdrantLangGraphTemporalLangfusePostgresNext.js

The process

How a chatbot build runs.

Six to ten weeks for most builds. You see a working bot in week two, not at the end.

Corpus audit

Golden set

Retrieval build

Guardrails

Channel wiring

Evals and handover

Common questions

Chatbot questions we get first.

Not answered here? Email us. A real reply from an engineer, no form.

How is this different from a plain ChatGPT wrapper?

A wrapper answers from the model's memory and invents the rest. We retrieve from your content first and hand the model only what it needs, with a citation on every claim, so a wrong answer is visible instead of confident and silent.

Will it make things up?

Not from nothing. It answers from your corpus and cites the source. Where retrieval finds nothing relevant, we make it say so and offer a human, rather than filling the gap.

Can it see data a user should not see?

No. Permissions are enforced at retrieval time, not in the prompt, so the bot can never surface a document the asking user could not open themselves.

How do you prove it works before we launch?

We agree a golden question set with you at the start and score every change against it. You get the numbers, and the set stays yours to rerun long after we hand over.

What does it cost to run?

Model spend depends on traffic and how much context each answer needs. We size it during the corpus audit and usually route planning to a strong model and extraction to a cheap one to keep it predictable.

From the blog

Field notes

RAG development

Retrieval that is measured, not guessed.

Read it

AI agent cost in 2026

What drives the quote, and the ops nobody budgets for.