Gates Foundation Grand Challenge

AI That Helps Donors Give More and Give Sooner to Maternal Health

MamaMatch AI removes the friction between donor intent and maternal health impact — connecting everyday givers to the right partner faster, and bringing them back with personalized context that makes every next gift count.

$91M+
Raised, Ramadan 2026
50+
Partner organizations
192K+
Donors across 3 campaigns
136
Countries represented

One integrated system addressing all three areas

Most proposals address 1-2 challenge areas. This addresses all three as interdependent layers — each strengthens the others.

🔍
Area 1 — Understand

AI-Powered Donor Matching

The gap: Donors want to give but feel paralyzed by 50+ organizations — most leave without giving.Helps donors navigate a complex, multi-layered cause. AI identifies donor preferences — country, type of support, and urgency — and matches them to 1-3 partners whose work resonates. Turns "I want to help mothers" into "Here's the organization doing exactly what you care about."

🤝
Area 2 — Convert

Personalized Pathways to Give

The gap: Donors on a partner page have intent but not enough conviction — last-mile hesitation is where donations die.Establishes a clear, simple pathway from motivation to meaningful contribution. AI-driven personalization across channels — seasonal campaigns, donor re-engagement, targeted outreach — makes the act of giving easier for everyday donors across different countries and communities.

📊
Area 3 — Infrastructure

Public Health Data Pipeline

The gap: Philanthropic data is siloed — AI cannot translate a gift into meaningful impact without a structured public health data layer.The IHME/LiST public health data pipeline makes philanthropic data visible to AI — connecting country-level mortality data to partner program outputs to produce credible, evidence-based impact estimates for donors.

From intent to impact — the causal pathway

Donor arrives with intent

Motivated to support maternal health — but unsure where to give among 50+ organizations

AI reduces discovery friction

Conversational matching recommends 1-3 partners based on donor values, geography, and preferences

AI profiles donor preferences

Builds understanding of the donor — country, type of support, and urgency level — to create personalized matches and segment into actionable personas

Evidence contextualizes the need

WHO/UNICEF data grounds each partner's work in real maternal health gaps — not abstract appeals

Donor gives with confidence

Higher conviction, better match, to the organization whose work resonates most

Feedback loop sustains giving

Donor sees their $10 become $30 through the Pregnancy Boost Fund, funding 10 months of antenatal care. Personalized impact stories — not receipts — drive repeat giving.

More funding reaches frontline organizations

Where it's most needed and most aligned with donor values

Smaller partners get discovered

AI prioritizes under-funded organizations when donors are undecided — leveling the playing field within the coalition and directing funds where they're needed most

Sector-wide learnings published

Open data standards, matching patterns, and findings shared for replication

Two AI tools for two donor problems

When donors arrive at Every Pregnancy, they're motivated — but two distinct barriers stand between intent and donation. Each requires a different AI intervention.

Problem 1

Decision paralysis

"I want to give, but there are 50+ partners — which one should I fund?" The donor is motivated but overwhelmed. They don't know enough to choose, and the cost of choosing wrong feels high. Many leave without giving. Even those who pick one partner wonder if they're missing a better fit elsewhere.

What's needed: A way to narrow 50+ options to 1–3 matches — or skip choosing entirely by allocating across multiple partners at once, weighted by evidence.
Problem 2

Last-mile hesitation

"I'm on a partner page, but I'm not sure this is the right one." The donor has already found a partner — via a social media link, a friend's recommendation, or browsing. They have intent, but not enough conviction to click donate.

What's needed: Contextual evidence and impact clarity right on the partner page — reducing hesitation at the moment of highest intent, not adding more to read.

These are different donors at different stages. Problem 1 is an entry point problem — "help me decide." Problem 2 is a last mile problem — "help me commit." Each gets its own AI tool below.

Three sources that power MamaMatch AI

Every Pregnancy Platform
Proven demand at scale
192,000+ donors across 136 countries, 15+ program countries, 4+ years of Ramadan giving baselines and conversion data.
The platform, donors, and giving history already exist. The grant funds the AI layer on top.
54+ Partner Organizations
Coalition knowledge base
Partner profiles, countries served, program types, and interview data Every Pregnancy has already collected — the coalition's own knowledge, ingested once and structured for AI retrieval.
Open Public Health Data
The evidence layer
IHME Global Burden of Disease — Impact Engine
Country-level cause-specific mortality rates that power all impact calculations and partner MMR rankings. The same source used by GiveWell, MSI Reproductive Choices, UNFPA, and the Gates Foundation. Free, annually updated, no license required.
LiST Methodology — Effectiveness Coefficients
The Lives Saved Tool (Johns Hopkins) is the field standard for estimating intervention impact — used by UNICEF, WHO, and governments. We apply its RCT-validated coefficients to what partners already report. No new data collection from partners required.
UNICEF + SESRIC — Donor-Facing Context
Human-readable health statistics for partner pages and chat. SESRIC (Statistical Centre for Islamic Countries) anchors credibility for the Muslim donor and Zakat audience at scale.
Also: UNFPA · World Bank · Lancet MNH series · Cochrane reviews · DCP3. All open access, no license required.

Two moments. One outcome: donors give more and give sooner.

Pre-Donation · Give sooner

MamaMatch AI

Identifies the right partner from 54+ organizations using cause, region, and urgency. What used to take hours of research takes a 3-minute conversation — reducing friction between donor intent and action.

Post-Donation · Give more

Impact & Re-engagement

After every gift, a personalized impact estimate and follow-up email show what the donation did. Return visits recognize the donor and make giving again effortless — turning one-time givers into lasting supporters.

Both moments are shown below.

Solving decision paralysis — "Help me choose"

MamaMatch AI guides donors from intent to giving in three steps: select a cause, choose a region and urgency, then commit to the right partner — or a portfolio. Donors who already know their partner go straight there. Every path is designed so partner relationships are protected and donor intent is honored.

EP

MamaMatch AI

AI-powered matching
Select a demo scenario
Guided Match
I want to help mothers in Gaza what partner works in Somalia zakat for pregnant women Bangladesh
Match Portfolio
$100 monthly for the most urgent need
Direct Match
my friend donated to SRD
The problem we're solving

Helping donors decide — without diverting funds from the partner they came for

Every Pregnancy's partners each work hard to acquire their own donor base. A donor who arrives knowing they want to support Partner X — because of a Ramadan campaign, a friend's share, or a trusted recommendation — represents committed intent that belongs to Partner X. The published evidence is unambiguous about what happens if MamaMatch AI pushes alternatives or splits at this moment: both total donations and the number of successfully-funded partners go down.

Intent override is documented to backfire
When donors with a chosen recipient are offered competing alternatives, total contributions and the number of successfully-funded projects both decrease. Corazzini et al., Journal of Public Economics — peer-reviewed (Tier B).
Paradox of choice has a measurable threshold
More than 3 options at any single decision point reduces both participation rate and average gift size. The 3-option ceiling is the published optimum for charitable-giving decisions. Iyengar/Lepper lineage — Tier B.
AI persuasion needs concrete framing
AI bots delivering data-driven framing (numbers, sources, specific outcomes) outperform the same bot delivering abstract appeals — p < 0.001 across 4 experiments, ~1,000 participants. Wang et al. 2025, MDPI — peer-reviewed (Tier A).

Our solution maps directly onto the evidence:

  • Ternary entry gate (Partner / Cause / Help me find) — exactly 3 options, the published optimum.
  • G1–G6 partner-intent guardrails — once a donor signals partner intent, no portfolio or split UI ever appears, eliminating Corazzini-style intent override.
  • Concrete data framing throughout — every partner card shows IHME MMR with country source citation, every Zakat claim is verified, every dollar amount maps to a specific outcome.
  • Up-sell only after the gift is locked in, framed as additive (a separate new gift), never as a re-allocation.

Full source list and verification tiers in docs/chatbot-flow/research-conversion-evidence.md.

MamaMatch AI branches at the very first question — two options, one per intent level. Each branch has guardrails baked in so we never cross partner interests. Persistent ↶ Back to step and ↻ Start over controls live in the chat header so a donor can revisit any prior decision without losing the conversation.

STEP 0 · ENTRY — "Which issue matters most to you today?" Prevent mothers dying Premature & newborn Train midwives Emergency obstetric Zakat-eligible Browse by region Partner in mind DIRECT MATCH GUIDED MATCH PATH A · SINGLE PARTNER 1Searchable pickerAll partners 2Walkthrough onlyIHME data · Zakat 3Checkout 100%No split, no diversion 🛡 G1 GUARDRAILNo portfolio button on this path PATH B · GUIDED MATCHING 13 preference questionsCountry · Type of support · Urgency 2Top 3 + See allInline portfolio CTA 3Pick OR portfolio+ Increase / Add / Both 🛡 G2 GUARDRAILPortfolio only on matches screen COMMIT GATE · STRIPE Donation processed — intent locked in Resume flow if donor closes modal mid-payment + POST-DONATION · IMPACT ESTIMATE + ADDITIVE UP-SELL "Your $50 contributed to ~20 safe deliveries in Pakistan" Quantified impact card + suggest 2–3 similar partners as separate new gifts 🛡 G4 NEVER A SPLIT
Direct Match · highest intent
Guided Match
Guided Match · region-first
Commit gate
Guardrail
0
Step 0 — Entry: "Which issue matters most to you today?"
Six cause options: Prevent mothers dying in childbirth / Premature & newborn survival / Train midwives & birth attendants / Emergency obstetric care / Zakat-eligible giving / I have a partner in mind. Cause-first entry is the key design choice — opening on outcomes, not navigation. A seventh option ("Browse by region") gives donors who want to filter without committing to a cause access to the region-first Guided Match.
Direct Match · highest intent
"I have a partner in mind"
Focused single-partner path
1Searchable picker — full partner list, vertical 2-column grid
2Walkthrough — that partner only · IHME health context · Zakat status
3Checkout — 100% to this partner · single Stripe transaction
🛡 G1 · No portfolio button anywhere on this path
Guided Match · cause-first
Cause selected at Step 0 → 2 questions
Country & urgency — cause already captured
1Country / region · Urgency — 2 questions max
2Top 3 of N matches · inline "Try Impact Portfolio" CTA · "See all" expansion
3Pick partner OR open portfolio → checkout
🛡 G2 · Portfolio only on the matches screen, never after a partner is picked
Guided Match · region-first
"Browse by region"
Single-tap filter, no questions
14 region pills — pick one, see all partners in that region instantly
2Filtered match list → pick partner → checkout
3Escape hatch: "I want both — help me find" → switches to Guided Match
Commit gate — donation processed via Stripe
Whatever the donor chose — single partner or portfolio — the payment lands first. The donor's original intent is now locked in. Closing the modal mid-payment shows a Resume/Change amount/Start over recovery in the chat.
+
Post-donation — impact estimate + optional additive up-sell (single-partner gifts only)
"Your $50 contributed to safer pregnancies in Pakistan. ≈ 20 mothers supported through safe delivery." Quantified impact card replaces the old celebrate bubble. Below it, suggests up to 3 similar partners as separate additional gifts. Skipped entirely after portfolio donations.
🛡 G4 · Always additive · never a re-allocation of the locked-in gift
Header controls — always visible in every state
↶ Back to step opens a dropdown of every prior decision the donor reached (Step 0 entry · Guided Match questions · Top matches · Walkthrough for each partner viewed). Selecting one re-enters that step additively so the prior conversation stays visible above. ↻ Start over wipes everything and restarts from Step 0.
Core principle
MamaMatch AI never moves a donor's intended gift away from the partner they came to support. Every branch, every button, every up-sell, and every back-step option is designed so that partner trust is protected and donor intent is honored.

AI reads what donors type. Rules do the matching.

The matching engine is deterministic — rule-based filters, mortality ranking, Zakat eligibility flags. The AI has one job: when a donor types instead of clicking. That first free-text message is where intent is fragile, phrasing is unpredictable, and keyword routers break. The AI extracts what it can, asks only for what's missing, and hands off to the rule-based engine. These examples show the full range of what it handles.

Guided Match region typed → 1 clarification → single partner
I want to help mothers in Gaza
EP
Got it — partners in Gaza. Any particular cause you’d like to focus on?
Prevent maternal deaths Emergency obstetric care Open to anything
Emergency obstetric care
EP
Healing for Gaza is your best match — they deliver emergency obstetric care and safe delivery in active conflict zones, fully Zakat-eligible, MMR 1,047. Would you like to proceed to donate?
Yes. I want to donate for $100.
EP
Opening secure checkout for $100 to Healing for Gaza…
Healing for Gaza · $100
Gaza · Zakat ✓ · Emergency obstetric care
Checkout →
Zakat path · Direct match full context typed → instant match, no questions
zakat for pregnant women Bangladesh
EP
Mercy Without Limits is your best match — they provide Zakat-eligible prenatal care and safe delivery support in Bangladesh, top-ranked for your region and giving type. Would you like to proceed to donate?
Yes. I want to donate $75.
EP
Opening secure checkout for $75 to Mercy Without Limits…
Mercy Without Limits · $75
Bangladesh · Zakat ✓ · Prenatal care
Checkout →
Match Portfolio undecided → 1 clarification → split gift
$100 monthly for the most urgent need
EP
$100/month · highest urgency — got it. Which region should this reach?
Middle East East Africa Anywhere critical
Anywhere critical
EP
Here are the 3 highest-need partners — your $100/month split by mortality urgency. Proceed with this portfolio?
Yes. Split it.
EP
Opening secure checkout for $100/month across SRD, MedGlobal & Amoud…
SRD · $38
MedGlobal · $35
Amoud · $27
Impact Portfolio · weighted by IHME data
Checkout →
Direct Match · partner in mind partner named → straight through, G1 guardrail
my friend donated to SRD, I want to do the same
EP
SRD is your match — they deliver emergency obstetric care and skilled birth attendant training in Somalia, one of the highest-need countries in East Africa. Would you like to proceed to donate?
Yes. $50.
EP
Opening secure checkout for $50 to SRD…
SRD · $50
Somalia · MMR 621 · Emergency obstetric care
Checkout →
Guided Match · region browser region typed → filtered partners → select → checkout
what partner works in Somalia
EP
3 partners in East Africa (Somalia) — ranked by maternal mortality urgency:
SRD · MMR 621 Amoud Foundation · MMR 621 Zamzam Foundation · MMR 621
SRD
EP
SRD is your best match — emergency obstetric care and skilled birth support in Somalia, highest maternal mortality in the region. Would you like to proceed to donate?
Yes. $75.
EP
Opening secure checkout for $75 to SRD…
SRD · $75
Somalia · MMR 621 · East Africa
Checkout →
Guided Match · vague intent emotional phrasing → 1 cause question → match
I want to do something meaningful for mothers
EP
That’s a great place to start. What would you like this gift to do?
Prevent maternal deaths Train midwives Open to anything
Train midwives
EP
Midwives for Mothers is your best match — they train skilled birth attendants in rural Sierra Leone. Would you like to proceed to donate?
Midwives for Mothers · West Africa
Sierra Leone · Midwife training
Checkout →
Knowledge Q&A · question mid-flow donor asks about a partner → grounded answer → checkout
Is SRD financially transparent? I want to know before I give.
EP
Yes — SRD is registered with the UK Charity Commission and publishes audited annual accounts. Administrative costs are 12% of expenditure. Would you like to proceed to donate?
Yes. $100 to SRD.
EP
Opening secure checkout for $100 to SRD…
SRD · $100
Somalia · Answer grounded in partner knowledge base
Checkout →

When a donor types instead of clicking, the AI is what prevents the "I don’t understand" dead end. It reads intent, fills in gaps, answers questions — then hands the rest to the rule-based engine. The structured flow handles the journey. The AI handles the door.

After the gift — impact and re-engagement

The donor journey doesn't end at checkout. On a donor's return visit (7+ days later), MamaMatch AI recognizes them via client-side localStorage and opens with a personalized follow-up — limited to the same device and browser without cleared history. Two post-donation features close the loop from barrier "Understanding impact" and "Knowing what to do next."

EP

MamaMatch AI

AI-powered matching · return visit
EP
Welcome back. Thank you for your last donation…
$50 to Mercy Without Limits in Bangladesh
— That gift helped support safe deliveries for ~20 mothers
MOCKUP · based on data captured at checkout · no partner submission required

Stored under mamamatch.donations — partner slug, amount, timestamp, impact string. No email, name, or device fingerprint.

Right after donation — personalized follow-up
Email preview · sent immediately after donation
From Every Pregnancy <impact@everypregnancy.org>
To donor@example.com
Subject Your impact with Mercy Without Limits — and what we found for you
EP
Every Pregnancy
Connecting donors to maternal health impact

Thank you for your generosity yesterday.

Your donation
$50 → Mercy Without Limits, Bangladesh
April 18, 2026 · Confirmed · Receipt #EP-29481
Estimated impact
~20 mothers supported through safe deliveries
Based on MWL's cost-per-service data + IHME GBD Bangladesh baselines. Modeled estimate, not a measured outcome.

During your visit, you asked about a few things. Here is what we found:

From your conversation

You asked about safe delivery programs in rural Bangladesh and Zakat eligibility. Mercy Without Limits works in Sylhet — where just 47% of women access antenatal care. Your gift directly funds the gap they are filling.

This summary reflects only what you asked — not assumptions about you.

Your interest in maternal health in South Asia means there are similar partners you may want to know about when you are ready to give again.

Every Pregnancy · 54+ partners across 15+ countries
Unsubscribe · View in browser
How the donor profile is built — no PII collected

Email content is generated at checkout from three signals: (1) donation data (amount, partner, timestamp from Stripe); (2) conversation summary — the AI compresses the chat session into topic tags at session end; (3) country context from the partner selected, not IP geolocation. No name, device fingerprint, or browsing history is used. The personalized email is composed once and queued via Stripe webhook — the donor's email address comes from the checkout form they complete themselves.

Longitudinal partner outcome reporting — where partners submit updated field data months after a gift — is out of scope and requires complementary funding. Year 1 delivers three things: the immediate post-donation message in chat, the Ramadan return recognition via device memory, and the 3-month follow-up email — all generated from data captured at checkout, with no partner data submission required.

A second email — three months later

Three months after the gift, donors receive a follow-up grounded in donation-time data: the partner they supported, the country context, and the impact estimate generated at checkout. No new partner data is required.

Email preview · 3-month partner update
From Every Pregnancy <impact@everypregnancy.org>
To donor@example.com
Subject 3 months ago, you gave to Mercy Without Limits — here's what happened
EP
Every Pregnancy
A 3-month update from Mercy Without Limits

Three months ago, your $50 went to Mercy Without Limits in Bangladesh. Here is where that gift stands.

Your gift · Ramadan 2026
$50 to Mercy Without Limits, Bangladesh
April 18, 2026 · Confirmed · Receipt #EP-29481
Estimated impact
Your gift to Mercy Without Limits last Ramadan supported maternal health in Bangladesh. Based on Mercy Without Limits's program data, an estimated 20 lives may have been protected.
Based on IHME GBD Bangladesh baselines + MWL cost-per-service data captured at checkout. Modeled estimate.

Ready to give again? Ramadan 2027 is when your next gift will have the most impact.

Every Pregnancy · 54+ partners across 15+ countries
Unsubscribe · View in browser
How this email is triggered

The email is scheduled at the moment of donation and sent automatically 90 days later. All content — the impact estimate, country context, and partner summary — is drawn from data captured at checkout. No partner data submission is required for this email to send.

Partner outputs → estimated lives saved

This engine powers two things
The chatbot's real-time impact estimate — shown immediately after checkout: "Your $50 → ~20 safe deliveries."
Both post-donation emails — the immediate email and the 3-month follow-up both draw on these coefficients and country data.
Without this engine, we say "thank you for giving." With it, we say "here's what your money did."

When a partner's reported output maps to a validated effectiveness coefficient, MamaMatch AI generates an evidence-based impact estimate using country-specific baselines from IHME Global Burden of Disease. When it doesn't, the partner's activity is shown as reported. Every modelled estimate passes a three-condition validity gate and is clearly labeled as a projection, not a measured outcome.

① PPH Bundle (E-MOTIVE)

Postpartum Hemorrhage Prevention

Uncontrolled bleeding after childbirth causes ~27% of all maternal deaths. The E-MOTIVE bundle (oxytocin + tranexamic acid + guided response) cuts severe PPH by 60%. Partners report: bundles delivered + country. Everything else is public data.

Partner example: MedGlobal delivered 563 PPH bundles in Nigeria (program report, Q1 2025).
Evidence: E-MOTIVE Trial, Nature Medicine 2024 · RCT · N=10,000+ across Kenya, Nigeria, South Africa, Tanzania
Formula
deaths_averted
  = bundles × 0.60
  × PPH_share(country)
  × CFR_severe(country)
For each bundle delivered: 60% chance of preventing a severe bleed, multiplied by how often severe bleeds are fatal in that country.
Example calculation — MedGlobal · Nigeria
563 × 0.60 × 0.24 × 0.20
= 16–30 maternal deaths averted
Nigeria PPH share: 24% · CFR severe: ~20% · IHME GBD 2022
Country variables: IHME Global Burden of Disease (free, public) · Effectiveness: E-MOTIVE trial
Why IHME + LiST methodology?

IHME Global Burden of Disease is Gates Foundation-funded. LiST (Lives Saved Tool) from Johns Hopkins is the standard used by UNICEF and WHO. When a Gates reviewer reads "we apply LiST methodology with IHME country data," they recognize it — it is the same standard their program officers use internally. This makes estimates auditable, not proprietary.

3 conditions before showing any estimate
  • The intervention has a published RCT or Cochrane coefficient
  • The country has IHME baseline mortality data
  • The partner has reported a verifiable output count

If any condition fails, show "Impact data not yet available for this partner" — never a fabricated number.

Intervention Library — Published Effectiveness Data
Intervention Evidence Effectiveness Primary Source
PPH Bundle (E-MOTIVE) RCT 60% reduction in severe PPH Lancet 2023, E-MOTIVE trial (WHO-led, 10 hospitals)
Antenatal Corticosteroids Cochrane 30% reduction in neonatal deaths Cochrane 2020, Oladapo et al. — 30 RCTs, 7,774 women
Kangaroo Mother Care Meta-analysis 20–41% reduction in neonatal deaths Cochrane 2023, Conde-Agudelo & Díaz-Rossello
Modern Contraception DCP3 Model 38/100K maternal deaths averted per 1% coverage gain DCP3 2016, Bhutta et al.; IHME GBD cause fractions
Skilled Birth Attendant Multi-country 21% reduction in neonatal deaths Lancet Global Health 2019, Semrau et al.
Obstetric Fistula Repair Clinical series 65% quality-of-life improvement post-repair WHO 2018 clinical guideline; Arrowsmith et al. fistula outcomes

All effectiveness coefficients are from peer-reviewed published sources. Applied via LiST methodology with IHME GBD country-specific baselines. Estimates are labeled as modeled projections, not clinical claims.

What each source provides for Every Pregnancy
IHME Global Burden of Disease
Formula engine
  • Country-specific MMR (deaths/100K live births)
  • Cause-of-death fractions by intervention type
  • Free API — gates-funded, no license required
healthdata.org/gbd
UNICEF Data Warehouse
Donor-facing context
  • ANC coverage rates by country
  • Skilled birth attendant coverage
  • Contextual stats for post-donation impact cards
data.unicef.org
SESRIC (OIC Countries)
Muslim donor credibility
  • OIC member-state health indicators
  • Maternal health data for Zakat-relevant countries
  • Adds authority for Muslim-majority audience framing
sesric.org/stats

Understanding givers to serve them better

Before AI can match donors to partners, it needs to understand who's giving and why. We're building a donor intelligence layer from partner interviews, behavioral data, and persona-driven A/B testing.

🎤
Research

Partner Interviews

Interview partners to understand their existing donor base — age, gender, geography, giving patterns. These insights seed the initial persona models and reveal what motivates real donors.

👤
Segmentation

5-6 Donor Personas

Build actionable donor personas that guide segmentation, communication, and AI matching. Each persona represents a distinct motivation and giving pattern — from first-time Zakat givers to recurring monthly donors.

🧪
Optimization

A/B Testing & Lead Scoring

Test persona-driven experiences against controls. Iterate based on what segmentation works. Build toward automated lead scoring that improves matching accuracy over time.

How we build MamaMatch AI

Six workstreams that turn a working prototype into a production AI system — from infrastructure to measurement. Each phase builds on the last; the knowledge base and partner data pipeline feed every downstream component.

  1. 1 Infrastructure
  2. 2 Knowledge Base
  3. 3 Core AI Development
  4. 4 Configuration & Calibration
  5. 5 Integration & QA
  6. 6 Measurement & Reporting
Key challenges to overcome

Building a production LLM chatbot for donor matching involves challenges beyond standard web development. These are the hard problems we need to solve:

Knowledge base completeness

The AI must not miss information when responding. Incomplete or outdated partner data leads to wrong matches or missing context. Requires continuous validation that the knowledge base covers all partners, programs, and health data without gaps.

Conversation guardrails

Questions and answers must stay focused on Every Pregnancy's mission — maternal health, partner programs, and donation guidance. The model must not drift into medical advice, political topics, or off-brand responses. Requires robust topic boundaries and fallback handling.

Cost optimization at scale

Every conversation incurs token costs. At public scale (thousands of donors during Ramadan), costs can escalate quickly. Requires prompt engineering for efficiency, caching frequent patterns, and potentially distilling to smaller models for common flows.

Factual accuracy & grounding

Training the model to answer with the right data — not hallucinate partner details, fabricate impact numbers, or cite outdated statistics. Requires retrieval-augmented generation (RAG) with strict grounding to verified sources and partner-approved content.

Multilingual & cultural sensitivity

Donors span multiple languages and cultural contexts (Zakat/Sadaqah framing, regional naming conventions, faith-based giving norms). The AI must handle these naturally without stereotyping or misrepresenting faith traditions.

Latency & user experience

LLM responses must feel conversational, not sluggish. Donors expect sub-second interactions. Requires streaming responses, precomputed common paths, and fallback to deterministic flows when the model is slow or unavailable.

Budget allocation

Proposed allocation of the Grand Challenge grant across five workstreams over the 1-year project period.

Gates Foundation Grant
AI/ML development & integration
Product design & frontend
Data infrastructure
Testing, measurement & publication
Partner outreach & coordination

A practical plan to test, measure, and generate actionable learning

We use a stepped-wedge trial (Hemming et al., BMJ 2015) — MamaMatch AI rolls out sequentially to donor cohorts during Ramadan 2027. Each cohort serves as its own control before activation. Control: donors reaching Every Pregnancy partner pages directly (pre-launch Ramadan 2026 baseline from existing payment and analytics data). Treatment: Navigator-assisted donors during Ramadan 2027.

The "test" in the Gates requirement refers to measuring real donor behavior — specifically whether AI-assisted matching causes measurable change in giving rate and gift size compared to the pre-launch baseline. This is the stepped-wedge trial described below, not software QA.

1 — Primary
Donation volume lift
Total donation volume from chatbot-assisted donors during Ramadan 2027 vs. the Ramadan 2026 baseline. The clearest "give more" signal.
Collection: payment attribution tag (chatbot vs. direct) vs. prior year baseline
2
Completion rate
% of chatbot sessions that end in a completed donation. Measures whether the chatbot closes the gap between interest and action.
Collection: structured events log — session start → payment confirmed
3
Bounce rate reduction
% of chatbot visitors who leave without donating — compared to the Ramadan 2026 baseline of partner page visitors who left without giving. A lower bounce rate means the chatbot is converting more visitors into donors. The "give sooner" signal.
Collection: web analytics — visitor sessions vs. completed donations, Ramadan 2027 vs. 2026 baseline
4
Impact estimate qualification rate
% of completed donations where the 3-condition validity gate passed (estimate shown) vs. fallback message shown. Measures responsible AI execution in practice.
Collection: structured events log — estimate shown vs. fallback triggered per completed donation

Measurement infrastructure

Structured Events Log
Conversion funnel: step entered → cause selected → partner matched → checkout opened → donation completed
AI Quality Tests
Automated test suite run before each deploy — verifies intent parsing accuracy and that AI answers stay grounded in the knowledge base
Web Analytics
Public-facing funnel analytics for Every Pregnancy, Marleen, and Gates Foundation reporting

Year 1 measurement report — findings reported to Gates Foundation

1
Donor persona report — conversation patterns by global audience segment (UK, North America, Gulf/Muslim diaspora)
2
Intent → completion funnel breakdown by chatbot path (A/B/C) and drop-off step
3
Zakat/Ramadan giving pattern insights — seasonal behavior, Zakat-specific paths, language register effects
4
Impact translation accountability report — 3-condition validity gate results, impact estimate qualification rates, and responsible AI audit findings
Fallback clause: If MamaMatch AI cannot produce measurable lift in intent rate or overall donation impact by end of Year 1, the fallback is a smart static page with cause-tagging and urgency ranking — not a full chatbot rebuild. The impact translation and data infrastructure deliverables remain intact regardless.

Every Pregnancy operates in the gap

No existing AI tool solves the coalition navigation problem — routing donors across a curated network of organizations in a shared cause area.

PlatformApproachEvery Pregnancy's Differentiation
Charity NavigatorAI ratings across 1.5M+ charitiesEvery Pregnancy goes deep within one cause across a curated coalition
DaffyConversational AI for DAF holdersEvery Pregnancy serves mass donors, especially Ramadan/Zakat givers
Fundraise UpAI-optimized single-org donation formsEvery Pregnancy innovates on which org — the routing layer
DonorSearchAI tools for fundraising teams (B2B)Every Pregnancy's MamaMatch AI is donor-facing — donors are the users
Muslim Charity UKAI-personalized Ramadan emailsEvery Pregnancy applies AI across 50+ orgs with Zakat routing

Complete your donation

🔒 Secure checkout
$
edit amount