Unloq has been a great partner to work with. Their approach helped us identify where AI should be applied and the AI solution they developed is now delivering on the needs of our client and their users.
The results
30,236
recorded coaching calls distilled into 44,000+ expert interventions
364
people supported since Joules went live in February 2026
90%
first-time pass on the response evaluation suite, 125 of 139 questions
ICE Creates had years of coaching expertise locked in recorded calls, and no way to reach anyone at the moment they needed it. Phone coaching happens on the coach’s schedule, and cravings do not keep one.
The raw material sat as audio files and Word documents in SharePoint with no ingestion pipeline, no transformation logic and no quality controls. The service had to work in Welsh and English from the first day, and it had to be safe enough for people in active withdrawal. In a consumer product a poor answer costs someone a bad afternoon. In a stop smoking service commissioned by a health board, a response that invents coping advice or mishandles a disclosure of risk is a safeguarding incident.
The coaches knew what to say at that moment, because they had said it on thousands of calls, but none of it was in a form a machine could use.
We built the data platform and the AI application across five workstreams, and we did not build a search over health content. The corpus had to be interventions, because someone fighting an urge already knows the benefits of stopping and needs one concrete thing to do in the next five minutes.
Each workstream was data engineering first. Claude comes in where reasoning and language change the quality of the reply, and we chose different models for coaching, for content generation and for data extraction, matching each model to the job in front of it.
Joules is now being extended from smoking to vaping.
How we did it
- 01
Turn the calls into interventions
We transcribed 30,236 recorded coaching calls and pulled out the expert interventions in them, each tagged with what triggered it, the coaching strategy and the technique. That gave Joules more than 44,000 interventions drawn from real coaching, instead of a library of health facts.
- 02
Build the platform around them
Ingestion, hybrid search, notifications, the data model and telemetry went in as five workstreams. Claude writes each reply under one instruction: keep the clinical substance of the retrieved intervention exactly, say it in the coach’s voice and invent nothing.
- 03
Put safety in front of the answer
Risk detection runs before the knowledge base is searched, so when it flags a risk the system stops before it generates a word. A person is alerted, the event is logged inside the client’s own environment, and a Welsh speaker gets the safety response in Welsh.