Concept case studyHealthcare · Outpatient clinic
From repetitive questions to an AI-assisted front desk, with people in charge.
How NairobiX would use AI carefully in a clinic: answering routine questions from approved information, preparing work for staff, and keeping every clinical and sensitive decision with people.
Concept engagement — illustrative scenario, not historical client results.
Patient asks
Do you accept my insurance, and is a GP available tomorrow morning?
Automated reply · labelled as automated
Yes, we accept this insurer for outpatient consultations. GP consultations run 8am–5pm. I can pass a booking request to our front desk — would you like me to?
Answered from: Insurance list (updated 2 Sep) · Clinic hours
Escalated to staff · urgent wording
21:14 · Fixed guidance sent · On-duty staff alerted
The assistant did not answer. The patient received the clinic's standard emergency guidance, and the on-duty nurse was notified to call back.
- Industry
- Healthcare · Outpatient clinic
- Focus
- AI with human oversight
- Solution areas
- AI SolutionsBusiness AutomationCRM & Sales Systems
- Reading time
- 9 min
02The context
A business that works — up to a point.
Consider a private outpatient clinic in Nairobi with several doctors across general practice and a few specialities, a pharmacy counter and a laboratory. The front desk handles walk-ins, phone calls and a busy WhatsApp number.
A large share of messages ask the same things: opening hours, whether a particular doctor is in today, consultation fees, which insurers are accepted, how to prepare for a test, where to park, whether results are ready.
The staff are good with patients. But answering routine questions competes with the work that genuinely needs them — booking, reminders, following up missed appointments, and patients who are anxious or unwell.
What already works
- Trusted doctors and a loyal patient base
- A clinic management system for records and billing
- Friendly, experienced front-desk staff
- Clinic policies that are mostly written down somewhere
What the team manages by hand
- Answering the same questions dozens of times a day
- Taking bookings by phone and WhatsApp, then re-entering them
- Reminder calls the day before appointments
- Chasing patients who missed appointments
03The problem
Not a lead problem. A system problem.
The obvious answer is another receptionist. The underlying problem is that routine, verifiable information takes the same human attention as the moments that genuinely need a person — and the urgent message sits in the same queue as the question about parking.
Access
Patients ask routine questions on WhatsApp and phone at all hours.
Staff time is consumed by repetition; after-hours questions wait until morning.
Knowledge
Correct answers exist — in staff heads, on notices and in old messages.
Answers vary depending on who replies.
Booking
Booking requests arrive in chat and are re-entered into the clinic system.
Double handling and occasional errors.
Reminders
Reminder calls happen when there is time.
Missed appointments leave unused slots.
Escalation
Urgent or sensitive messages sit in the same queue as routine ones.
The messages that matter most can wait the longest.
Reporting
There is no view of what patients ask, or which needs go unmet.
Improvements rely on anecdote.
Recognise some of this in your own business?
04The opportunity
The opportunity is not to replace the front desk with a chatbot. It is to let AI handle what is routine and verifiable, route everything else to people faster, and give staff their time back for patients.
Generative AI can produce confident answers that are wrong — the risk that the US National Institute of Standards and Technology calls confabulation. In a clinic that risk is unacceptable for anything clinical. So the design starts from boundaries: an approved knowledge base the assistant may answer from, escalation rules that take precedence over it, and people who own both.
Health information is also among the categories Kenya's Data Protection Act treats as sensitive personal data. The assistant is designed to collect as little as possible — it books; it does not ask about symptoms.
01
Attention
02
Capture (focus of this engagement)
03
Qualify (focus of this engagement)
04
Nurture
05
Convert (focus of this engagement)
06
Deliver (focus of this engagement)
07
Retain (focus of this engagement)
08
Measure (focus of this engagement)
Where this engagement concentrates first, within the NairobiX growth system.
05The system design
The architecture, layer by layer.
The proposed architecture has eight layers. Three of them — knowledge, escalation and governance — exist to constrain the AI, not to extend it. Select a layer to explore it.
- Why it exists
- The assistant may only answer from content the clinic has written and checked. No approved answer, no automated answer.
- What moves through it
- Approved answers, their source and when they were last reviewed.
- What happens next
- The assistant retrieves from it when a patient asks.
- Why it exists
- Routine questions get consistent, immediate answers — including in the evening.
- What moves through it
- The question, the answer given and the knowledge items used.
- What happens next
- Anything outside the knowledge base, or any sign of urgency, goes to escalation.
- Why it exists
- Safety comes first, so the rules sit in front of the model — not behind it.
- What moves through it
- The reason for escalation, and the conversation so far.
- What happens next
- Urgent messages receive fixed, pre-approved guidance and alert on-duty staff; others enter the staff inbox.
- Why it exists
- Removes re-typing, and keeps sensitive information out of the chat.
- What moves through it
- The booking request only.
- What happens next
- Front-desk staff confirm the booking in the clinic system.
- Why it exists
- Staff see what needs them, in priority order, with context already summarised.
- What moves through it
- Conversations, summaries, priorities and actions taken.
- What happens next
- Staff reply, book or call back.
- Why it exists
- Fewer missed appointments, and a person — not an automated message — reaching out when someone doesn't come.
- What moves through it
- Appointment, reminder status and attendance.
- What happens next
- Missed appointments create a task to rebook.
- Why it exists
- Using AI responsibly with patients is an ongoing practice, not a setting.
- What moves through it
- Review findings, corrections and policy changes.
- What happens next
- Corrections update the knowledge base and escalation rules.
- Why it exists
- It shows what patients need — and where the clinic's information has gaps.
- What moves through it
- Everything above, summarised weekly.
- What happens next
- Knowledge gaps are filled; recurring questions improve the website and notices.
Layer 01 of 08
Approved knowledge base
Hours, doctors' schedules, fees, accepted insurers, test preparation, directions — each item with an owner and a review date.
- Why it exists
- The assistant may only answer from content the clinic has written and checked. No approved answer, no automated answer.
- What moves through it
- Approved answers, their source and when they were last reviewed.
- What happens next
- The assistant retrieves from it when a patient asks.
Every business needs a different version of this architecture.
06Before / after
From a disconnected pattern to a connected one.
The same people and much of the same technology — organised as one system instead of several.
Before · current operating pattern
- WhatsApp on the front-desk phone
- Phone calls
- Notices on the wall
- The clinic system
- A paper reminder list
- Staff memory
Every question — routine or urgent — waits in the same line for the same people.
After · designed future-state system
- 01Approved knowledge
- 02AI assistant
- 03Escalation rules
- 04Booking requests
- 05Staff inbox
- 06Reminders
- 07Governance
- 08Reporting
Routine answered, urgent escalated, people focused where they matter.
An architectural comparison of operating patterns — not a record of measured results.
07The customer journey
One customer, through the designed system.
What the customer experiences at each step, and what the system does behind it. Names and details are illustrative.
What the customer experiences
What happens in the system
01
Ask
Customer
At 9pm, a patient asks whether the clinic accepts her insurer and whether a GP is available tomorrow.
System
Escalation rules check the message first: routine, nothing urgent.
02
Answer
Customer
She receives a clear answer within seconds, and the message says it is automated.
System
The assistant answers only from the approved insurance list and clinic hours, and shows what it used.
Patient asks
Do you accept my insurance, and is a GP available tomorrow morning?
Automated reply
Yes, we accept this insurer for outpatient consultations. GPs are available 8am–5pm tomorrow. Would you like me to pass a booking request to our front desk?
Answered from: Insurance list (reviewed 2 Sep) · Clinic hours
03
Request
Customer
She says yes and gives her name and a preferred time. She isn't asked about symptoms.
System
A booking request is created — no clinical detail collected.
Booking request · GPAwaiting confirmation- Preferred time
- Tomorrow, morning
- Visit type
- General consultation
- Channel
- WhatsApp · after hours
- Next
- Front desk confirms at 8:00
04
Confirm
Customer
At 8am, she receives a confirmation from the front desk with the time and the doctor's name.
System
Staff confirm the slot in the clinic system; the confirmation goes out from the record.
05
Remind
Customer
The day before her follow-up test, she gets a reminder with preparation instructions.
System
The reminder uses approved wording from the knowledge base.
06
Attend
Customer
She attends. At the desk, staff have time to talk to her.
System
Front-desk time previously spent on routine questions goes to patients in the room.
07
Escalate
Exception path
Customer
Another patient, late at night, writes about chest pain.
System
The escalation rule fires before the AI: fixed emergency guidance is sent, and the on-duty nurse is alerted to call.
Urgent · call back now
22:41 · Escalated by rule · Assistant did not answer
Standard emergency guidance sent. Conversation summary attached. On-duty nurse: please call the patient.
08
Review
Customer
Nothing changes for patients — this happens behind the scenes.
System
Each week, the practice manager reviews a sample of AI answers and all escalations, and corrects what was wrong.
09
Measure
Customer
Patients get routine answers faster, and urgent ones reach a person sooner.
System
Reporting shows question topics, escalations and the questions the knowledge base couldn't answer.
Questions by topic · layout
Hours & availability
Fees & insurance
Test preparation
Not in knowledge base
Illustrative layout — not data
08The technology layer
Technology, chosen last.
The AI model is the least important decision here. The knowledge base, the escalation rules and the review process decide whether the system is safe; the model is replaceable. Choices depend on data-processing terms, cost and how each option performs in evaluation on the clinic's own questions.
Proposed stack · example architecture
Each technology fills a capability the system design called for. Remove the requirement, and the tool goes with it.
AI layer: OpenAI or Anthropic models
Messaging: WhatsApp Business Platform
Database: PostgreSQL
Staff tools: Next.js
- Business requirement
- System design
- Technology
AI layer
OpenAI or Anthropic models
AI assistant · Staff inbox
Role
The language model that answers routine questions from approved content and summarises conversations for staff.
Why it fits
Either could fill this role. The choice is made through evaluation on anonymised historical questions, and on data-handling terms.
Messaging
WhatsApp Business Platform
Access · Reminders
Role
The patient-facing channel, with utility templates for confirmations and reminders.
Why it fits
Where patients already message the clinic; the Platform allows the assistant and staff inbox to share one number.
Database
PostgreSQL
Knowledge · Governance · Reporting
Role
The approved knowledge base, conversation logs and review records, with access controls.
Why it fits
A reliable, well-understood database the clinic's data can be kept in, with clear retention rules.
Staff tools
Next.js
Staff inbox · Knowledge
Role
The staff inbox and the knowledge-base editor used by the practice manager.
Why it fits
A focused internal tool is simpler for staff than a general-purpose platform.
Illustrative technology choices for this concept scenario. They are not a record of a client implementation, imply no partnership with any vendor, and are not the only suitable stack.
09The automation layer
Rules first. Then workflows. Then automation.
Three workflows, one principle: rules run before AI, and people are the last step for anything that matters.
- Process
- The sequence of work the business agrees to follow.
- Rule
- A condition written down — “if no reply in three days”.
- Workflow
- A rule connected to the actions it should cause.
- Automation
- The workflow running without anyone starting it.
- Technology
- The tools that carry it out.
Routine question
- Trigger
A patient sends a message.
- Rule
Escalation rules check for urgency, clinical content or complaints first. If none match…
- Action
…the assistant answers only if the approved knowledge base covers the question, and labels the reply as automated.
- Notification
If the knowledge base doesn't cover it, the conversation goes to the staff inbox with a summary.
- Follow-up
The assistant offers a booking request or a person.
- Human
Staff handle everything outside the approved knowledge.
- Measure
Record the topic, whether it was answered, and knowledge gaps.
10Designed outcomes
What the system is designed to improve.
01
Routine questions answered consistently
The same approved answer, whoever asks and whenever they ask.
02
After-hours coverage for routine needs
Evening questions about hours, fees and insurance don't wait until morning.
03
Urgent messages reach people faster
Escalation rules pull them out of the queue instead of leaving them in it.
04
Less re-entry
Booking requests arrive structured, ready to confirm.
05
Fewer missed appointments
Reminders with preparation instructions, and a person following up when someone doesn't come.
06
Visibility of patient needs
What patients ask, and where the clinic's information falls short.
07
Knowledge kept current
Every approved answer has an owner and a review date.
08
Safer use of AI
Boundaries, disclosure, review and data minimisation designed in from the start.
Design objectives the system is built around — not measured client results.
11Measurement framework
What we would measure.
No invented numbers — a plan for knowing whether the system is working, set before it is built.
Access
- Messages by topic and hour
- What patients need, and when.
- Answered from approved knowledge
- Share of routine questions the assistant could answer.
Quality & safety
- Answer accuracy
- From the weekly human review of sampled answers.
- Escalation accuracy
- Messages escalated that should have been — and any that should have been but weren't.
- Corrections
- Knowledge-base changes made after review.
Operations
- Time to human contact
- For escalated and urgent messages.
- Booking request → confirmed
- Time taken, including after-hours requests.
- Staff time on routine messages
- Sampled periodically.
Patient
- Missed appointments
- And rebooking after a missed appointment.
- Patient feedback
- On the messaging experience, including the assistant.
System
- Knowledge gaps
- Topics the assistant couldn't answer.
- Retention compliance
- Conversations deleted on schedule.
12The human layer
Where people stay responsible.
Technology supports people who make decisions.
In healthcare, the human layer is the system's core, not its fallback. AI assists with routine information and summaries; people own every clinical, sensitive and exceptional decision — and they own the AI's knowledge and boundaries too.
What the system does
- AI answers routine questions from approved content and summarises conversations for staff.
- Automation routes, reminds and records.
- The knowledge base keeps one approved, reviewed version of the truth.
What the team decides
- Anything clinical — symptoms, advice, results
- Urgent or distressing situations
- Exceptions to fees or insurance
- Complaints
- What goes into the approved knowledge base, and when it changes
13What we would build
One system, built in stages.
01
Strategy
Use-case selection, escalation policy and governance, agreed with clinical leadership.
02
Knowledge
An approved knowledge base with owners and review dates.
03
AI
An assistant with boundaries and disclosure, evaluated on anonymised historical questions before any patient sees it.
04
Automation
Routing, reminders and missed-appointment follow-up.
05
Digital
A staff inbox and knowledge-base editor.
06
Analytics
Topic, escalation and knowledge-gap reporting, plus the weekly review.
Components of one system, built in stages and connected to each other — not a list of separate services.
In this scenario, the value of AI comes as much from what it is not allowed to do as from what it does.
14 · The next step
Your business probably has a different version of this problem.
Every business has a different starting point. The right system depends on your goals, requirements, customers, processes and existing technology.