AI Assistant With a Human Coach as the Safety Net
A client-facing AI assistant for OME Health, designed to hand the conversation to a human coach the moment it is not sure.
OME Health is a UK digital health and nutrition coaching platform, headquartered in London. Each client is paired with a coach who guides a 12-week personalised programme, using tracker data, biological test results, and lifestyle questionnaires to shape week-by-week guidance.
Visit website ↗Put an AI assistant in front of clients without taking the coach out of the loop.
GreenM built a client-facing AI assistant, deployed inside OME Health's own cloud and integrated into the client-messaging app the platform already used. The assistant answers routine client questions and hands anything unsafe or uncertain to the client's human coach.
Rigid rules, and a coach who could only be in one conversation at a time
OME Health's guidance ran on a fixed rules engine: predictable, but rigid. Clients had a growing pile of personal health data and a coach who could only be in one conversation at a time. Routine questions queued and waited for a human, and a health platform cannot let an AI answer freely on medication or symptoms.
- Rules-based logic that answered only what it was built for
- Coaches stretched across one-to-one conversations
- Client questions waiting in a queue for a human reply
- Hard safety line around medication, supplements, and symptoms
- Rigid, rules-based recommendations
- Every routine question routed to a human coach
- Personal data underused in day-to-day guidance
- No safe way to automate client answers
GreenM delivered an AI assistant grounded strictly in OME Health's own content, set up so that every message is risk-checked before a reply is sent. Confident answers go straight to the client. Anything uncertain is escalated to the assigned coach, with full context.
- Answers drawn only from OME Health's own guidance, not the open web
- A confidence and risk check on every message before it is sent
- Confident answers to the client; uncertain ones handed to the coach
- Runs inside OME Health's own cloud, inside their existing client-messaging app
The safety core of the system. When the assistant is not confident it can answer safely, it does not guess. It hands the conversation to the client's coach and tells the client it is doing so. Escalation covers the four cases where a wrong answer does real harm:
- Medications, supplements, and dosing
- Emotional support and signs of distress
- Urgent or acute symptoms
- Ambiguous health signals
On escalation, the client gets a calm handover message, and the coach gets the client's identity, the escalation reason, a suggested response, and a deep link straight into the chat. No medical advice is given during handover. The client is never left waiting, and the coach never starts blind.
- Confidence-scored responses
- Human-in-the-loop escalation
- Answers grounded in OME Health's content
- White-label, multi-region ready
- Client-owned deployment
The assistant reached real client users and left OME Health able to keep building on their own.
Live in client pilots
Running with real client users across pilots in Lithuania and Singapore.
Owned by the client's team
Architecture clear enough that OME Health's own engineer continues the work post-handover.
Follow-on engagement
Pencilled in for OME Health's next initiative, the Horizons data and AI project.
Cost fit to the model
Matching the model to the task keeps per-client AI cost inside subscription economics.
"On top of what you delivered, a good reflection of the architecture, it's clear to know what to do and how to proceed." — Aidas · CEO, OME Health
Client-facing AI has to fit the business model underneath it, so cost was a design input from the start.
- Started with a lighter model, then moved to a more capable one for better answer quality; the two swap in about five minutes with prompts and tuning preserved
- The assistant pulls in only the context it needs for each answer, holding usage down
- Measured cost per question: about $0.085 on the lighter model and $0.166 on the stronger one, against a subscription near €8 per client per month
- Matching the model to the task keeps the unit economics workable
DELIVERY
Delivered in a roughly eight-week build across three stages, from a secure environment to a production assistant handed to OME Health's own team.
Phase 1
Set up & secure
- Secure cloud environment inside OME Health's own account
- Networking, resource controls, and budget alerts
- Automated, credential-free deployments
Phase 2
Build & integrate
- Connected the AI models and message-routing logic
- Confidence and risk scoring that drives escalation
- Grounded answers in OME content; real-time messaging
- Coach escalation with full context and deep links
Phase 3
Test & hand over
- Guardrail tuning, tone adjustments, safety testing
- Localisation: emergency numbers, neutral coach refs
- Production rollout, docs, handover to OME engineer
The stack behind the assistant: Anthropic Claude (Haiku and Opus) via AWS Bedrock, orchestrated by CrewAI with a confidence and risk check on every message; RAG grounding on a Postgres/Aurora vector store; CometChat for real-time coach handover; AWS Lambda behind API Gateway; deployed to OME Health's own AWS account via Bitbucket pipelines with OIDC.
- Claude Haiku
- Claude Opus
- AWS Bedrock
- CrewAI
- RAG
- Postgres / Aurora
- CometChat
- AWS Lambda
- API Gateway
- Bitbucket OIDC
Healthcare-grade handling of personal health data, kept inside the client's own environment.
Client-owned infrastructure
Runs inside OME Health's own cloud account. No client data leaves their environment, and ownership transferred at handover.
UK GDPR alignment
A UK platform handling special-category health data: design follows UK GDPR and the Data Protection Act 2018, data kept in the client's cloud.
Human oversight by design
The escalation flow keeps a human coach in control of every unsafe or uncertain answer, in line with ICO and WHO guidance.
