Service 02 of 04
AI development,evidence first
We build agents that answer from content you own and say so. The interesting engineering is not the answer — it is knowing when there isn’t one. That is the part most demos skip and every real deployment needs.
Who this is for
Organisations where getting it wrong has a cost.
Customer service teams with predictable, high-volume questions your pages already answer — and where being wrong is not acceptable.
Organisations that want their content to be answerable and cited, not interpreted or improvised.
Teams where the answer to 'why did it say that?' needs to trace to a specific page, not a model's generalisation.
Not for organisations wanting an agent to improvise, speculate, or answer questions their pages do not address.
The evidence contract
Five properties every agent we ship has to hold.
Grounded
Retrieval runs over sources you control — your pages, your policies, your documents. No open-web fallback, so an answer can never come from somewhere you have not approved.
Cited
Every reply carries the page it came from, as a link the reader can check. Citation is part of the answer's layout, not a tooltip.
Refusing
With no supporting source the agent declines and routes to a human. We design that state first, because it is the one that protects you.
Evaluating
A question set drawn from your real support history, run on every change, with refusal rate and citation accuracy tracked as first-class numbers.
Logged
Every question, retrieval and refusal is inspectable by your team. You can read what people asked and what the system did about it.
The refusal, in practice
“I can’t answer that
from our pages.”
What we will not build
Agents that answer from the open web while wearing your brand.
Anything that impersonates a named member of your staff.
Automated decisions on people — hiring, credit, admissions, grades.
Training on your customers' data without their explicit consent.
Engagement
What a build looks like.
Four stages, in order. The refusal case before the happy path, every time.
Read your pages and support history
Before any code, we read what your visitors actually ask. We identify the questions your pages cover, the ones they almost cover, and the ones they cannot answer at all.
Design the refusal case
The first thing we write is what the agent says when it does not know. That state is designed before we write a single answer. It is the one that protects your brand.
Build the question set
A representative set drawn from your real traffic, not made-up examples. Every answer and refusal is reviewable before go-live.
You review before it ships
You see every answer and every refusal in the question set. One you are not satisfied with means we do not ship. No exceptions.
What we build
Grounded chat & support agents
Public-facing or internal. Agent AI Chat is our own, in production.
Retrieval over your content
Ingest, chunking and freshness for sites, handbooks and PDFs.
Model integration into existing tools
Inside the CRM or admin your team already opens each morning.
Evaluation & guardrail work
Often on someone else's build, as a day-rate engagement.