How Non-Profits are using AI-based Innovations to Scale Impact

Abstract

The panel examined a four‑month AI Cohort Programme (Sept.–Dec. 2022) that brought together seven NGOs to design, prototype and pilot AI‑enabled solutions for social impact. Project Tech4Dev conceived and hosted the cohort; the Agency Fund supplied a pooled technical‐resource model; the NGOs shared the problems they tackled (teacher‑assist chatbots, personalised learning pathways, AI‑generated mentorship scripts, etc.), the challenges they faced (data quality, hallucinations, integration with existing workflows), and the lessons learned about responsible AI, mentorship, evaluation frameworks, and collaborative design. The discussion concluded with practical recommendations for NGOs looking to embark on AI projects and a brief audience Q&A.

Detailed Summary

• Moderator (Manohar) began by noting the unconventional use of an AI “assistant” (named Elizabeth) to introduce the participants, underscoring the theme of AI‑mediated interaction.

• Elizabeth read out short bios for each panelist:

  1. Tamina Madon – Co‑founder, The Agency Fund; works at the intersection of technology, philanthropy, and economic mobility.
  2. Erica Arya – CEO, Project Tech4Dev; leads open‑source platforms and advisory services for over 200 NGOs in the Global South.
  3. Mainak (Mynak) Roy – Co‑founder & CEO, Simple Education Foundation; partners with Indian state governments to scale teaching‑and‑learning, exploring responsible AI for educators.
  4. Steven Suting – Director of Technology & Product, QUEST Alliance; heads AI‑driven digital learning platforms for Indian youth.
  5. Pritam Sukumar – Technology & Research Lead, Avanti Fellows; builds data‑driven mentorship tools for public‑school students.
  6. Manohar Sreekanth – Partner & CTO, Sattva Consulting; moderator, works at the tech‑data‑AI intersection for social‑sector scaling.

• The moderator joked that by next year a “digital assistant” might sit on every panel, highlighting the community’s growing comfort with AI tools.

2. Framing the AI Cohort Programme

2.1 Genesis & Objectives (Erica Arya)

• Origins – Project Tech4Dev had already run a Data‑Catalyst cohort (with Dasra) that demonstrated the value of small, deep‑dive programs for NGOs.

• Motivation – Two observations drove the AI Cohort:

  1. Hands‑on mentorship need – Many NGOs possessed promising use‑cases but lacked engineering capacity, mentorship and funding to move from idea to prototype.
  2. Learning loop – A cohort enables iterative feedback from a tight group of NGOs, feeding insights back into Tech4Dev’s open‑source platforms.

• Structure –

  • Open application with screening calls to ensure a good fit.
  • Free participation for NGOs (pilot funding from Project Tech4Dev).
  • Resource commitment – NGOs had to allocate a dedicated staff member and secure leadership buy‑in.
  • Mentor model – Each NGO paired with one‑or‑two mentors (including members of Project Tech4Dev and external experts).

• Scale – Seven NGOs were selected for the four‑month pilot; three of them (Simple Education, QUEST Alliance, Avanti Fellows) presented during the panel.

2.2 The Agency Fund’s Cohort Philosophy (Tamina Madon)

• Cohort analogy – Tamina likened the model to Y Combinator and the South Park Commons (SPC) accelerator, where early‑stage founders benefit from peer support, shared resources, and collective learning.

• Resource pool – The Fund created a shared pool of 10 technical staff (engineers, product managers) that NGOs could draw upon as needed, rather than requiring each NGO to hire full‑time AI engineers.

• Why a cohort? –

  • Loneliness mitigation – AI development is often solitary; a cohort offers community and emotional support.
  • Economies of scale – Centralised resources reduce duplication of effort across NGOs.
  • Accelerator experience – The Fund had already run an AI for Global Development year‑long accelerator, informing the design of the shorter four‑month program.

• Product‑management gap – Tamina highlighted that many NGOs lack product managers; the pooled staff filled that gap, increasing the likelihood of successful delivery.

3. Problem Statements & AI Solutions from the NGOs

3.1 Simple Education Foundation – “Teacher Buddy” (Mainak Roy)

• Core problem – Teachers need a rapid, evidence‑based pedagogical recommendation for each class; manual research is time‑consuming, especially for teachers handling eight lessons a day.

• Solution architecture –

  • Platform – WhatsApp‑based chatbot (chosen because WhatsApp is ubiquitous among teachers, even in remote Indian districts).
  • Workflow – Teacher sends class‑type & objectives → AI (presumably LLM‑powered) returns a proven pedagogical strategy.

• Challenges –

  • Conversation initiation – The bot mis‑interpreted the first teacher message (often a QR‑code scan) as a “high” instead of a prompt, breaking the data capture flow.
  • Question scope – Teachers sometimes strayed from the intended query set, requiring guardrails and clarification prompts.

3.2 QUEST Alliance – Personalised Learning & Behavioural Nudges (Steven Suting)

• Core problem – Heterogeneous learner profiles (different confidence levels, gender, access) clash with a one‑size‑fits‑all “chalk‑and‑talk” classroom model; teacher shortage (≈1:30) reduces individual attention.

• Solution sketch –

  • AI‑driven digital learning platform that captures learner signals (tone, speed, confidence) to generate personalised nudges and career‑development advice.
  • Behaviour‑science focus – The system treats the AI‑learner interaction as a behavioural‑science problem rather than a pure software issue.

• Current status – The project remains in the “pie‑in‑the‑sky” conceptual stage, with a team of engineers and behavioural scientists building the scaffolding.

3.3 Avanti Fellows – AI‑Generated Mentorship Scripts (Pritam Sukumar)

• Core problem – Large‑scale online learners (≈200 k) receive only score‑based feedback; lack of actionable, human‑like mentorship hampers progress.

• Iterative use‑case journey –

  1. Initial attempt: Improve student reports → insufficient impact.
  2. Pivot: Replace teacher‑student mentor conversation with AI‑generated script summarising performance, attendance, chapter strengths/weaknesses.

• Deployment – Pilot with 15 teachers; each teacher conducted ~57–75 AI‑guided mentorship sessions.

• Issues observed –

  • Hallucinations: LLM sometimes inverted trends (e.g., “decreased” instead of “increased”). Prompt‑engineering was required to correct.
  • Trust gap: Teachers remained wary until hallucinations were mitigated.

3.4 Cross‑NGO Themes

• Resource constraints – Common lack of in‑house AI engineers; reliance on pooled technical staff or mentors.
• Data quality & model reliability – Hallucinations and inaccurate outputs were a universal pain point, leading to extensive prompt‑tuning.
• Adoption barriers – Integration with existing workflows (e.g., WhatsApp vs. bespoke UI) required careful user‑experience design.

4. Mentorship Model & Operational Learnings

• Mentor involvement – Each NGO was paired with 1‑2 mentors (often from Project Tech4Dev). Mentors provided technical guidance, helped define problem‑to‑solution pathways, and facilitated peer learning.

• Practical challenges –

  • Bot initiation glitches (Simple Education) → solution: “First‑message‑as‑high” workaround.
  • Guardrails – Need to restrict bot’s answer space to avoid inappropriate or irrelevant responses (e.g., “What should I do?” vs. policy‑restricted queries).

• Collaboration & Knowledge Sharing – In‑person workshops allowed NGOs to showcase progress, receive feedback, and avoid duplicate effort.

5. Responsible AI & AI‑Safety Framework (Erica & Tamina)

• Knowledge partners –

  • Digital Future Labs – Provided expertise on integrating Responsible AI principles into design (fairness, transparency).
  • Tattl – Focused on AI safety, offering guard‑rail plugins (e.g., profanity‑filter “slur list”).

• Four‑Level Evaluation Framework (outlined by Tamina):

  1. AI System / Model Evaluation – Ensure reliability, safety, reduce hallucinations, apply content filters.
  2. Product Evaluation – Measure activation, engagement, retention among pilot users (e.g., teachers).
  3. User Evaluation – Survey‑based assessment of changes in beliefs, confidence, behaviour (e.g., teachers’ confidence in using AI).
  4. Impact Evaluation – Large‑scale outcome measurement (health, education, livelihood) – rarely reached in private‑sector projects but central to social‑sector impact.

• Implementation – NGOs incorporated these evaluation steps into their development cycles, embedding responsible‑AI checks from day 1 rather than as an after‑thought.

6. Ecosystem‑Level Collaboration

• Beyond the cohort –

  • Two health NGOs (not on the panel) collaborated on a high‑risk pregnancy prediction model after discovering overlapping use‑cases.
  • A third NGO built an assessment‑rubric tool for answer‑sheet grading; Project Tech4Dev facilitated a joint discussion with two other NGOs tackling similar problems, fostering shared codebases.

• Contrast with for‑profit world – Tamina noted that early‑stage cohorts are rare in corporate settings; social‑sector collaborations enable rapid learning, reduce duplication, and accelerate impact.

7. Audience Interaction & Key Learnings

• Audience composition – A sizable portion of the live audience were NGO representatives, prompting a focus on practical takeaways.

• Learnings shared by panelists

  • Problem‑first mindset – Start from concrete pain points, then evaluate whether AI is a suitable tool (instead of “AI‑first” solutions).
  • Avoid the “shiny‑object” trap – Resist jumping on AI hype merely to please funders; focus on sustainable, need‑driven solutions.
  • Scale‑ready design – Build with scalability and responsibility in mind (e.g., “golden data set” for consistent output).
  • Leverage open‑source – When possible, adopt existing open‑source platforms (e.g., Superset for dashboards, OpenAI models, Tattl safety plugins) rather than reinventing entire stacks.

• Q&A Highlights

  • Question: “Do you evaluate existing large‑model offerings (e.g., Gemini) versus building custom solutions?”

  • Answer (Erica): The team generally integrates open‑source components (OpenAI models, Superset dashboards). Even when building a new platform, they stitch together existing tools and add custom layers only where gaps are identified.

  • Follow‑up: Audience urged panelists to share documentation; Erica confirmed that Project Tech4Dev publishes detailed blogs and whitepapers on the cohort.

• Closing Remarks – The moderator thanked the panel and audience, reaffirmed the importance of sharing resources, and encouraged NGOs to reach out for mentorship or documentation.

Key Takeaways

• Cohort‑based mentorship accelerates AI adoption for NGOs by providing pooled technical talent, peer learning, and shared resources, reducing the need for each organization to hire full‑time AI engineers.
• Start with a concrete pain point; evaluate AI as a means‑to‑solve, not as a default solution. This avoids costly “shiny‑object” projects that may not align with organisational needs.
• Responsible AI must be baked in from day‑one – leveraging partners (Digital Future Labs, Tattl) to embed safety, fairness, and guardrails early prevents downstream trust issues.
• Open‑source integration is the norm – most NGOs and the coordinating organisations stitch together existing tools (OpenAI models, Superset dashboards, Tattl plugins) rather than building everything from scratch, saving time and money.
• Evaluation needs a multi‑layered framework: model reliability → product usage → user behavioural change → societal impact. NGOs should adopt all four levels to demonstrate effectiveness and attract further funding.
• Collaborative ecosystems unlock synergies – NGOs working on similar problems (e.g., predictive health models, assessment rubrics) can co‑develop tools, share data, and avoid duplicated effort.
• User‑experience matters – simple, familiar channels (e.g., WhatsApp) lower adoption barriers, but designers must anticipate conversation‑flow quirks and build guardrails to maintain data quality.
• Hallucinations remain a practical challenge for LLM‑based tools; iterative prompt engineering and continuous testing are essential before scaling.
• Leadership buy‑in and dedicated staff time are critical – without organizational commitment, AI pilots stall despite technical support.
• Documentation and open dissemination (blogs, whitepapers) amplify the impact of cohort learnings and help other NGOs replicate successes.

---

Prepared by the AI Conference Summarisation Team.

See Also:

• trusted-ai-practical-pathways-for-the-public-sector
• ai-innovators-exchange-accelerating-innovation-through-startup-and-industry-synergy
• empowering-the-human-edge-building-a-future-ready-workforce-in-the-age-of-ai
• flipping-the-script-how-the-global-majority-can-recode-the-ai-economy