Service 07 — Chatbots & AI

Custom agents,
trained on
your business.

Custom personalised chatbot development — you are not getting a second-hand template with your logo on it. Agents grounded in your own content, with honest limits and a human escape hatch.

What is included

  • Content grounding
  • Guard rails
  • Human handoff
  • Evaluation suite
  • Privacy posture
  • Cost and latency control

Most chatbots fail for the same reason

They are deployed to deflect support tickets, they are not grounded in anything specific, and within a week customers learn that the fastest route to a human is to type nonsense at the bot. The deflection number looks good and the satisfaction number quietly collapses.

A useful agent needs three things: real grounding in your own content, an honest admission when it does not know, and a fast handoff to a person. We build all three, and we measure resolution rather than deflection — the two are not the same number.

Grounded, not improvised

We index your documentation, product data, policies and past tickets, and the agent answers from retrieved passages with the source attached. When the retrieval finds nothing relevant, it says so and hands over. That single behaviour is the difference between a tool people trust and a novelty they route around.

Deliverables

What you
actually get.

01

Content grounding

Your documentation, product data and policies indexed and retrievable, with citations shown to the user.

02

Guard rails

Topic boundaries, refusal behaviour and tone, tested against adversarial prompts before anything goes live.

03

Human handoff

A clean escalation with the full conversation attached, into the helpdesk your team already uses.

04

Evaluation suite

A fixed set of real questions with expected answers, run on every change, so an improvement is provable.

05

Privacy posture

What is sent to a model, what is retained, and what never leaves your systems — written down before launch.

06

Cost and latency control

Caching, routing between models and token budgets, so the bill scales with value rather than with traffic.

Shapes of the work

Where this
usually lands.

Use caseWhat good looks likeWhat we measure
Support deflectionAnswers with sources, escalates cleanlyResolution rate, not deflection rate
Internal knowledgeFinds the policy nobody can locateTime to answer, staff satisfaction
Lead qualificationAsks the three questions sales needsQualified conversations per week
Product guidanceNarrows a big catalogue to a shortlistAdd-to-cart and return rate

Stack

What we build
this with.

We work in your stack when you have one. These are our defaults when the choice is ours.

PythonLLM APIsRAG pipelinesVector storesNode.jsWebSocketsPostgreSQLRedisPlaywright

Questions

Before you
commit.

Will it make things up?
Any language model can. We reduce it with retrieval-grounded answers, visible citations, refusal behaviour when nothing relevant is found, and an evaluation suite that runs on every change. We do not promise it is impossible, because that promise would be false.
Where does our data go?
Wherever you decide, and it is written down before launch. We can keep retrieval and storage entirely inside your infrastructure and send only the minimum to a model provider — or run a smaller model on hardware you control.
Can it act, not just answer?
Yes, with confirmation steps for anything consequential. Booking, refunds or account changes get an explicit user confirmation and a full audit trail.
How much does it cost to run?
We model it per conversation before you commit, including caching and model routing, so the monthly number is not a surprise.

Ignition

Need chatbots & ai?

Tell us the problem, the users and the systems already in place. We will identify the right discovery or build step.