Health · AI agents
Sheba Scribe
An AI copilot that lets one community health worker do the work of three.
Community health workers speak naturally during a household visit. Sheba Scribe turns the conversation into structured records, flags danger signs against national protocols, schedules follow-ups and syncs to the ministry's health information system — even with no signal in the village.
Launch market: Bangladesh · 4 cited sources · Facts reviewed 14 October 2026
- Cash to MVP (est.)Our estimate of the cash needed to reach a working MVP: one-time setup plus monthly running costs for the build period. It excludes founder salaries.
- $7,550
- MVP timelineMinimum viable product: the smallest version real users will use and pay for, built to learn fast.
- 12 weeks
- Margin after AI (est.)Of every 100 you charge, how much is left after paying for the AI models and infrastructure behind it. Higher means a healthier business.
- ~70%
- Editorial scorePlankito's own judgement: the average of six 1–5 ratings (AI leverage, moat, competition gap, capital efficiency, timing, scalability). It is an opinion to compare ideas, not a forecast.
- 4.2 / 5
At a glance
- Who uses it
- Community health workers and their supervisors.
- Who pays
- NGO and ministry programmes, per worker per month.
- Who should build it
- A mobile engineer, a clinician, and someone who knows public procurement.
- The hardest part
- Clinical safety, and slow grant-driven procurement.
01
Opportunity
Frontline health workers spend a large share of their day on paperwork and duplicate registers instead of patients. Danger signs get missed because protocols are long and memory is short. Data reaches district managers late and incomplete, so outbreaks and gaps in care are spotted after the damage is done.
The evidence
5.85M
Regionalhealth workers short in WHO's African Region by 2030 — work that copilots must help existing staff absorb.
16% vs 29%
Globalgenerative AI use in the Global South versus the Global North, while the need for health workers is greatest in the South.
67%
Globalsuccess rate for AI bought from specialised vendors, versus about a third as often for internal builds — the case for a focused health product.
280×
Globalcheaper to run a GPT-3.5-level model than two years earlier, putting a copilot within a public-health budget.
Why now
- Speech models can now transcribe and structure a conversation in local languages, offline-first on a phone.
- Protocol checking — matching symptoms to national guidelines — is exactly what AI reasoning models do well with grounding.
- Health ministries are digitising registers onto open platforms, creating systems an AI copilot can plug into.
- Ambient clinical documentation is proven in rich-world hospitals; the frontline in low-income countries is untouched.
02
The AI case
What AI makes possible here, why the big labs won't absorb it, and whether the margins hold.
AI leverage — what is newly possible
- Listens to a household visit (with consent) and fills the register: vitals, symptoms, pregnancy stage, immunisations.
- Checks findings against national protocols and flags danger signs with a clear referral instruction.
- Plans the next week's visits by risk, not by habit.
- Gives district managers live dashboards instead of month-old paper summaries.
Lab-proof moatsA reason competitors, including big AI companies, can't easily copy you: data, distribution, trust or local knowledge.
Protocol and system integration
Encoded national guidelines and certified connectors to the ministry's health information system. Each country is a long, careful project.
Government and NGO contracts
Public-health procurement is slow to win and rarely switched once a system is embedded in daily work.
Clinical safety record
An audited track record of flagged danger signs and outcomes is a trust asset no new entrant has.
Inference economicsThe cost of running AI models each time a customer uses the product. It decides whether an AI business makes money. (estimates)
- Unit
- Health worker per month
- AI cost per unit
- ~$2–3
- Price per unit
- ~$10–12 (paid by NGOs, ministries or private clinic networks)
- Gross margin
- ~70%
How the margin is protected
- →Run speech recognition on-device where possible; sync text, not audio.
- →Use a small model for form-filling and a large model only for protocol reasoning.
- →Price per worker-month in multi-year programme contracts.
03
Market & expansion
Launch in Bangladesh. Then expand.
- Buyers are NGOs running community health programmes, ministries of health and private clinic networks — not the workers themselves.
- Bangladesh has large, well-known community health worker programmes, making it a strong pilot market.
- Safety, consent and data protection must be designed in from the first pilot, not added later.
- Validate: hours per day each worker spends on paperwork, and how many danger signs are currently missed or referred late.
Editorial scorecardPlankito's own judgement: the average of six 1–5 ratings (AI leverage, moat, competition gap, capital efficiency, timing, scalability). It is an opinion to compare ideas, not a forecast.
- AI leverage
- 5/5
- Moat durability
- 5/5
- Competition gap
- 4/5
- Capital efficiency
- 3/5
- Timing
- 4/5
- Scalability
- 4/5
Where it travels next
Sub-Saharan Africa
The largest health-worker shortfall and extensive community health worker programmes.
India & Pakistan
Huge frontline workforces with heavy reporting burdens.
Southeast Asia
Village health volunteer networks that ministries are digitising.
04
Competition & edge
Who already plays here, what they leave open, and what could go wrong.
Ambient clinical scribes (e.g. Abridge, Microsoft Dragon Copilot)
AI documentation for doctors in hospitals.
GapBuilt for rich-world hospitals, English and electronic health records — not village visits offline.
Health information platforms (e.g. DHIS2)
Open-source data systems used by many ministries.
GapThey store data; they do not capture it from conversation or check it against protocols. A partner, not a rival.
Digital health NGOs (e.g. Jacaranda Health)
Targeted digital services such as maternal health messaging.
GapFocused on specific programmes; not a general copilot for frontline workers.
Where the edge comes from
- Offline-first and voice-first — designed for the village, not the hospital.
- Integrates with the ministry's existing system instead of replacing it.
- Gives time back to workers, which makes adoption pull rather than push.
Risks and how to handle them
A missed or wrong flag harms a patient.
The copilot assists, never decides; conservative thresholds; human sign-off on referrals; clinical audit of every release.
Health data privacy and regulation.
Explicit consent, encryption, in-country hosting and alignment with national digital health policy.
Slow, grant-dependent procurement.
Start with an NGO partner with existing funding; publish outcome evidence to win government contracts.
05
Business model
Programme licences
Per worker per month in multi-year NGO and ministry contracts.
Private clinic networks
Documentation and follow-up for paid community health services.
Analytics
District and national dashboards for managers and funders.
06
Stack & data
Modern, proven tools a small team can run without a DevOps hire.
Worker app
Offline-first Android app (Expo / React Native) with local encrypted storage
Works with no signal; syncs when connectivity returns.
Speech
On-device speech-to-text in local languages, with cloud fallback
Privacy and offline use.
Reasoning
Claude API grounded in encoded national protocols
Careful, explainable protocol checks with citations to the guideline.
Integration
FHIR and DHIS2 connectors
Fit the ministry's existing systems instead of creating a new silo.
Dashboard
Next.js 16 + Tailwind CSS v4
Supervisor and district manager views.
Database
PostgreSQL + Drizzle ORM, in-country hosting
Relational health records with audit trails and data residency.
Evals
Clinician-reviewed case library for every protocol rule
Measure flag sensitivity and false alarms before each release.
Ops
Sentry + PostHog (self-hosted where required)
Observability without sending health data abroad.
Data model6 core tables, for the engineer starting the first migration.
workers
Health workers and supervisors.
- id
- name
- phone
- area_id
- supervisor_id
- role
households
Registered households.
- id
- area_id
- head_name
- location
- members_count
visits
A household visit.
- id
- worker_id
- household_id
- visited_at
- consent
- synced_at
observations
Structured findings.
- id
- visit_id
- person_id
- code
- value
- source (voice/manual)
flags
Protocol-based danger signs.
- id
- visit_id
- rule_id
- severity
- action
- acknowledged_by
protocol_rules
Encoded national guidelines.
- id
- programme
- condition
- logic (json)
- guideline_ref
- version
07
Costing
Directional monthly estimates in USD. Your numbers will differ — the structure is what matters.
| Item | MVP / monthWhat the product costs to run each month (hosting, AI models, tools) at MVP stage. | Scale / month |
|---|---|---|
| Hosting & database (in-country) | $100 | $1,200 |
| Model inference | $80 | $2,000 |
| Speech (cloud fallback) | $40 | $800 |
| Clinical reviewers (part-time) | $500 | $3,000 |
| Field support & training | $300 | $2,500 |
| Monitoring & security | $30 | $400 |
| Total | $1,050 | $9,900 |
One-time setup
08
MVP & roadmap
Shippable in about 12 weeks, then a path to production.
Build
- One programme (maternal and newborn health) with its national protocol.
- Voice-to-form visit capture, offline, with consent prompt.
- Danger-sign flags with referral guidance.
- Supervisor dashboard with sync status and flagged cases.
Not yet
- Other programmes
- DHIS2 live sync
- Predictive visit planning
- Private clinic plan
Roadmap to production
Validate
Weeks 0–6- Shadow 10 health workers for a day each
- Map the paper registers to a digital form
- Secure an NGO pilot partner and ethics approval
MVP
Weeks 7–18- Pilot with 50 workers
- Measure time saved per visit
- Flag sensitivity validated by clinicians
Traction
Months 5–12- DHIS2 integration
- Second programme
- 500 workers
Production
Months 13–24- Government contract
- Second country
- Published outcome study
09
Start this week
Your first steps. Tick them off as you go.
Metrics that prove it is working
- Minutes of paperwork saved per visit
- Danger-sign detection sensitivity and false-alarm rate
- Visits synced within 24 hours
- Cost per worker per month
10
Sources
Every figure on this page links to its source. Facts reviewed 14 October 2026. Costs and market notes without a citation are our estimates.
- 1.WHO Africa — Africa's health workforce expands but shortages, unemployment and migration intensify (2025)
- 2.Microsoft — The continued state of global AI diffusion in 2026 (2026)
- 3.Fortune (reporting MIT NANDA) — MIT report: 95% of generative AI pilots at companies are failing (2025)
- 4.Stanford HAI — AI Index Report 2025 (2025)