Software publishing2026

An AI support assistant for a software publisher

A support assistant grounded in the product documentation: it answers agents while citing the exact passage the answer came from, and stays silent rather than inventing.

A support conversation: an agent reports a licence error message, and the assistant splits the answer into two cases (the folder’s licence is out of date, or too many users are ticked), giving the menu path to follow for each, step by step. Every step carries in brackets the number of the source it came from, and the “Sources (4)” block, left open under the answer, names the four documentation articles cited. — Sourced answerDemonstration data

The same conversation when the documentation does not hold the answer: the assistant says the search found nothing relevant, asks which screen the agent is working from rather than inventing one, and offers to be taught the answer so it can cite it next time. Further down it does answer, but under an “Outside the documentation” heading that separates what is sourced from what is only a lead. — Outside the documentationDemonstration data

The learning screen, filtered to one of the supported applications: one conversation in three left unanswered, no sourced answer waiting to be checked, six taught answers live, and 86% of the questions raised here dealt with. Below, the questions still unanswered, sorted by frequency, each with the date it was seen and a button to reopen it. — LearningDemonstration data

Where it started

Agents searched a large documentation set by hand. Answer quality depended on how long the agent had been there, and the same customer could get two different answers depending on who picked up.

What we built

A documentation base chunked, vectorised and kept current automatically, queried through a conversational assistant. The design constraint was not model power but traceability: every answer carries its sources, and the assistant explicitly says it does not know rather than producing something plausible. Processing in Europe, with no data retention by the provider.

What changed

A brand-new agent answers from the same documentation as a veteran. And because the answer is cited, it is verifiable in one click: trust does not rest on the model, it rests on the source.

What changed, in numbers

of answers cite their sources: 100%of answers cite their sources
data retained by the provider: 0data retained by the provider
support domains, one assistant: 6support domains, one assistant

Let’s talk about your project

Describe what you are trying to build. We answer with an honest first read on feasibility and scope, even when the answer is that you do not need us.

Talk about your project