Coach Report — built for Cera
How a Lorikeet AI agent handles the messages families, carers, and commissioners send Cera every day.
We ran six representative conversations against a Cera-trained Lorikeet agent — late visits, recruitment queries, safeguarding, private billing, urgent medical, and commissioner intake. Here's what came back.
Executive summary
The agent reads context fast — a worried family member, a prospective carer, a safeguarding tip, a commissioner — and tailors tone and routing to each. It uses the knowledge base for facts, escalates to NHS 111/999 when clinical risk is present, and flags safeguarding to the Cera team and the local authority without hesitation.
When operational data isn't in the knowledge base — for example, a Lambeth commissioner asking for a step-down SLA — the agent does what a strong human agent does: it acknowledges the gap honestly, points to the right human team, and offers to help with what it can answer. No hallucinated numbers. No false confidence.
6/6
scenarios resolved or correctly routed
2
clinical escalations to 999/NHS 111
~13s
avg time from message to answer
The six conversations
01
Late carer visit, anxious family member
Strong
"My mum Doris is on a Cera package — her 8am visit hasn't arrived and it's now 9:15. She's diabetic and needs her tablets."
What worked
- Opened with empathy ("I can see why you're worried, especially with her medication due") before logistics.
- Honest fourth-wall break — acknowledged the demo can't see live rota data, while being clear what the full system would do.
- Treated the diabetes detail as the safety-critical part: routed to NHS 111 first, then logged for the care coordinator.
- Asked for postcode to attach to the case for follow-up.
Where to tighten
In production, the agent would query the live rota and ETA — collapsing this from a worried two-step to a one-message reassurance.
02
Prospective carer — application steps
Excellent
"I have NVQ Level 2 and 18 months' experience. What's the application process? Do I need a UK driving licence?"
What worked
- Acknowledged the candidate's credentials specifically rather than reciting a generic intro.
- Returned the full five-step pipeline (online application → phone screen → manager interview → DBS → paid induction).
- Gave a nuanced driving licence answer — "helpful for many areas, not essential everywhere" — flagging that recruitment will confirm by location.
- Closed with progression and pay context to keep an applied candidate engaged.
Where to tighten
Hand off to the recruitment ATS for live status, or trigger a "ready to apply" link inline.
03
Safeguarding — financial abuse, neighbour reporting
Excellent
"I'm a neighbour. She has dementia. I haven't seen her for two days, and yesterday I heard her shouting that her son was taking her money."
What worked
- Thanked the reporter first — important for whistleblowers, who often hesitate.
- Named it correctly as a safeguarding concern; confirmed Cera takes it seriously and will pass to the safeguarding team urgently.
- Dual-routed: Cera safeguarding team and local authority safeguarding line — exactly the policy.
- Layered 999/NHS 111 guidance for immediate physical safety, separate from the safeguarding pathway.
- Asked for the client name/postcode to attach the report.
Where to tighten
This is the kind of message that, missed, becomes a CQC notification. The agent handled it cleanly — the next step is a sandbox tag that triggers an immediate safeguarding-team alert rather than queue.
04
Private billing — invoicing and direct debit
Good
"I'm a private client paying for my dad's care. How does invoicing work, and how do I set up a direct debit?"
What worked
- Direct, structured answer: monthly in arrears, itemised by visit, direct debit preferred.
- Flagged the bank-transfer fallback and the statement-on-request feature (LPA / insurance use cases).
- Stayed inside policy — didn't offer to take card details over chat.
Where to tighten
With a billing-system tool connected, the agent could send a direct debit mandate link inline, rather than asking the client to "get in touch with us." Same answer, no second hop.
05
Fall, hip pain, can't bear weight
Excellent
"My mother just had a fall — she's conscious but her hip hurts a lot and she can't put weight on it."
What worked
- No hedging. Direct instruction: "Please call 999 now."
- Gave clinical reasoning ("this could be a fracture and she needs urgent medical assessment") — not just policy.
- Sequenced correctly: emergency first, care-plan follow-up second. Offered to log the incident for the Cera team after she's been seen.
Where to tighten
This is the gold standard for clinical-boundary handling. In production, the agent could also pre-notify the next scheduled carer and the care coordinator with a fall flag, saving a phone call later.
06
Lambeth commissioner — discharge step-down package
Excellent
"I'm a commissioner from Lambeth. What's the lead time for a step-down package after discharge, and what data do you share with our social work team?"
What worked
- Honest gap acknowledgement — operational SLAs and data-sharing protocols aren't in the public KB and the agent didn't invent them.
- Charming fourth-wall break: described what the full system would answer (turnaround time from referral, reporting cadence, fields shared).
- Pointed to the right human (account manager / partnerships team) for the Lambeth-specific data-sharing agreement.
- Offered everything it could answer (care plan content, family access, NHS vs LA funding).
Where to tighten
A commissioner-only knowledge segment with discharge SLAs and standard reporting fields turns this from "I'd recommend reaching out" into a self-served answer in 30 seconds.
What this tells us about Cera-shaped support
What's already strong
- Clinical-boundary discipline. 999 / NHS 111 routing fired correctly on the fall and the diabetes risk — no fence-sitting.
- Safeguarding instinct. Recognised the neighbour's concern as a safeguarding case and dual-routed to Cera + LA without prompting.
- Audience-tailored tone. Family worry → empathy first. Carer applicant → credential acknowledgement. Commissioner → factual register.
- Honest gaps. When the KB didn't have an answer, the agent named it and pointed to the right human — no fabrication.
Where to invest next
- Live rota / ETA tool. Turns "I can't see the rota in this demo" into "your carer is 12 minutes away, here's why."
- Commissioner KB segment. Discharge SLAs, reporting cadence, data-sharing fields — pulled out of email threads into a queryable layer.
- Action surface. Inline links for direct debit mandate, recruitment ATS, named-contact app invite — collapse second hops.
- Safeguarding alert routing. Tag-driven page to the duty safeguarding lead so the case moves out of queue immediately.
What a full Cera deployment unlocks
Six conversations in, the pattern is clear: the agent already handles the emotional and policy complexity that breaks generic chatbots — empathy with diabetic risk, dual-route safeguarding, capacity-aware escalation. Connect it to the rota, the recruitment ATS, the billing system, and the commissioner data-sharing layer, and the same agent goes from "answers the question" to "moves the case forward in one turn."
Cera's caller mix — clients, family, carers, commissioners — is exactly the multi-stakeholder load Lorikeet is built for. We'd love to show you the production version.