B2B technology20253–5 months
Client: A B2B software company
One of the project’s screens, in a desktop frame. It is shown again with its caption under “On screen”.
At a glance
Retrieval assistant, evals and guardrails
Product experience and hand-off to sales
AI infrastructure and observability
From brief to result
The brief in full, how we approached it, what we built and what it set out to change.
Sales engineers answered the same product questions again and again, and the answers drifted: an old limit here, a retired feature there. Prospects wanted answers in the moment, but a wrong answer about security or pricing was worse than none.
The brief: an assistant that answers from approved sources, says when it does not know, and hands over to a person when the question needs one.
We built the eval set before the assistant. Real questions from sales calls, each with its approved answer and the source it must cite, became the gate every model and prompt change has to pass.
Guardrails were written as policy first: what the assistant may answer, what it must refuse, and when it routes to sales. Prompt-injection tests (OWASP LLM01) run with every release.
Questions collected from call notes, tagged by topic and by risk.
The eval set and its rubric: grounded, correct and cited — or a clean refusal.
Retrieval over approved documentation, the assistant interface and the hand-off to sales.
Released only when the eval set passes; injection tests on every change.
A retrieval assistant grounded in approved product documentation, gated by an eval set, with prompt-injection tests and a hand-off to a person when confidence is low.
The assistant is built to answer only from approved documentation, show the page it used, and pass pricing and security questions to a person; every conversation is logged for review.
Accuracy is read against the eval set, not anecdotes, and the figures are the client’s to publish.
Picture story
The pictures on this page are illustrative: licensed stock photography and code-built mock-ups chosen to show the setting, not the client’s own work. Credits are at the foot of the page.
On screen
Illustrative screens, built in code to show the kind of screen the programme produced — not the client’s own. Open any of them full screen.
Field notes
Outcomes
What each measure is compared with, when it is read and who reports it — and the results, once the client approves them.
Answer accuracy on the eval set
Escalation rate to sales
Demo-to-opportunity rate
Figures pending client approval
We publish results only with the client’s written approval. Until then, the plan beside this is what the programme is judged on, and we can walk you through it under NDA.
Credits
People are named only with their agreement.
Licensed stock photography and code-built mock-ups, chosen to show the setting — not the client’s own work.
Keep going
Next case · 04 · Retail & commerce
Be cited in AI answers, not only ranked in search results.
Read the next case
Tell us what you’re building. We’ll answer straight.
01About 2 minutes
You get A reply from a lead, not a sales queue
02About 8 minutesMost useful
You get Options and a first scope after one call
03About 15 minutes
You get Receipt confirmed and a named bid lead
Every engagement starts with a written scope and a quote agreed before work begins. How each package is priced
