Member-ops AI inside an existing health fund stack
CBHS · Embedded AI · AU health fund
Member services still lived across claims, policy, and inbox tools. Straightforward questions bounced between teams, and AI sat in a slide deck — not in the work.
We wired an agent into the systems they already used: retrieval over approved policy, suggested replies for first-line staff, and a human approval step before anything left the desk.
Frontline keeps the same consoles. AI drafts. People decide. Repeat contacts drop.
31% faster first-response on routine member queries.
Context
Mid-size health funds do not have a transformation budget that can replace core claims, policy, and CRM systems. They also cannot wait for a greenfield “AI platform.” Member questions still arrive in the same inboxes, and staff still work in the same consoles.
The gap was not a lack of chatbots on a marketing site. It was that first-line teams were answering the same policy and claims questions by hunting across documents and previous tickets — while leadership had seen AI demos that never touched those tools.
What we focused on first
We mapped the member journeys that already existed: eligibility, extras, waiting periods, and “where is my claim.” For each, we listed the source of truth the fund already trusted — PDFs, knowledge articles, and structured policy fields — and the humans who must still sign off.
The insertion was deliberately narrow: draft a first response, cite the source, and leave send in human hands. No new member portal. No replacement of the claims engine.
Delivery approach
We connected retrieval to the approved knowledge the fund already maintained, then surfaced suggested replies inside the existing ticketing view. Guardrails included allow-listed sources, a visible citation, and a hard stop on clinical or financial advice the model is not permitted to give.
Staff trained on when to accept, edit, or ignore the draft. Exceptions — anything the retrieval could not ground — stayed on the manual path. That kept risk inside the operating model they already ran.
Outcomes and how we measured them
Time-to-first-response on the targeted routine categories fell by about 31% after the insertion stabilised, measured on the same queues and staffing levels. Repeat contacts for “I already asked this” dropped as answers became more consistent with the written policy.
The qualitative shift mattered as much: member ops stopped treating AI as a side project. It sat in the work they already did.
What we would emphasise for similar programmes
Insert AI into the queue you already have. Do not stand up a parallel chatbot that staff will not use. Ground every draft in sources the business already owns, keep send with a human, and measure first-response and repeat contact — not demo wow.
“We did not rip out our member stack. AI just started sitting in the queue with us — and the easy tickets stopped eating the day.”
Liked this insight?
See how these concepts apply to your business. Book a short fit call and we'll map a practical next step.