by Mitali Badkul and Aishwarya Shekar
The Glific AI Chatbot Accelerator was Glific’s largest accelerator so far. It took 34 Indian NGOs from curiosity about AI chatbots to working pilots in six months. This report tracks the same participants from the April baseline to the September endline, and lets their own words show what changed.
The programme opened with a two-day in-person kickoff on 20–21 April, where each team built its first AI-enabled WhatsApp flow and planned its first pilot. From May to September, every NGO had a dedicated Glific mentor, weekly office hours and check-ins, and ran a second pilot in the field, building on learnings from the first.
Education was the largest sector in the cohort, with 9 of the 34 NGOs, followed by health (5), gender equality and women’s empowerment (5), and child welfare and protection (3). The other 12 spanned nine more sectors, from agriculture and governance to disability and criminal justice (full list).
Share of matched participants who rate their team at “good working knowledge” or better 19 people from 15 organisations who answered both the baseline and the endline survey.
What we found
| Cohort | 34 NGOs in 13 sectors; 28 continuing after six months (82%); 28 submitted a pilot report |
| Matched participants | 19 people (15 organisations) answered both baseline and endline |
| Average self-rating, seven AI skills | 1.76 at baseline to 3.42 at endline (scale of 1–4); 18 of 19 improved |
| Overall experience | 4.36 out of 5; all 28 respondents rated 4 or 5 |
| Impact of the chatbot on the ground | 3.79 out of 5; none below 3 |
| Support from the Glific team | 4.61 out of 5; 68% gave the top rating |
Almost everyone moved, in every skill. Among matched participants, the average self-rating rose from 1.76 to 3.42, and 18 of 19 reported a higher average. The biggest gains were in running evals (from 1.3 to 3.2) and voice-enabled flows (from 1.4 to 3.2).
Evals went from unknown to routine. None of the 18 matched participants rated their team “good” at running evals in April. By September, 15 did. For an AI programme, learning to test an assistant before trusting it is the outcome that matters most.
Confidence followed capability. Average confidence in building independently rose from 2.6 to 4.1 out of 5. Of the 10 participants who started at “1” or “2”, 8 ended at “confident” or “very confident”.
Mentoring was the common thread. Mentors or support are named in about 16 of 28 “what worked” answers, ahead of any platform feature.
The challenge shifted from skills to sustainability. In April, participants worried about technical know-how and whether AI could be trusted with their communities. In September they described staff time, user adoption and funding. Those were also the reasons six NGOs dropped out at the end of the programme.
Capability: what changed for the same people
Matching the baseline and endline surveys by person shows how each participant compared with themselves.
| Skill area | Answered both | Baseline avg | Endline avg | Gain | Improved | Unchanged | Lower |
|---|---|---|---|---|---|---|---|
| Flows with file search | 18 | 2.06 | 3.61 | +1.56 | 17 | 0 | 1 |
| Conversational AI flows | 18 | 1.78 | 3.56 | +1.78 | 17 | 0 | 1 |
| Voice-enabled flows | 18 | 1.39 | 3.22 | +1.83 | 17 | 0 | 1 |
| Writing and iterating prompts | 18 | 2.00 | 3.33 | +1.33 | 13 | 4 | 1 |
| Running evals | 18 | 1.28 | 3.22 | +1.94 | 16 | 2 | 0 |
| Guardrails and AI safety | 19 | 1.68 | 3.47 | +1.79 | 18 | 1 | 0 |
| Piloting on the ground | 17 | 1.71 | 3.29 | +1.59 | 14 | 2 | 1 |
| Average of seven areas | 19 | 1.76 | 3.42 | +1.66 | 18 | 0 | 1 |
Across the whole cohort, the average of all baseline responses was 1.73 and of all endline responses 3.42.
Confidence to build and maintain an AI chatbot independently Matched participants (n=19). The question was worded slightly differently in the two surveys.
April concerns, September reflections
The baseline asked for the single biggest challenge in using AI. The endline asked what worked and what did not. Below, each organisation’s answers from the kickoff and the close-out sit side by side, showing how their thinking matured.
| Organisation | April: biggest challenge | September: what worked |
|---|---|---|
| Learning Links Foundation | “our GPT-integrated webhook occasionally generates ‘hallucinations’” | “We were able to understand, find and list all the areas in which the chatbot can play an important role in our project implementation.” |
| Collective Good Foundation | “Ensuring the bot is used by our beneficiaries in a meaningful manner” | “building a chat bot from scratch with community support” |
| Protsahan India Foundation | “the uptake of the chatbot by our target audience” | “The mentor office hours were very helpful for all the support I needed.” |
| Women’s Organisation for Rural Development | “We have not used any AI/Chatbots for our organisation.” | “Ease of use and navigation while preparation of flows for the AI chatbot.” |
| Catalysts for Social Action | “Apprehension about how an AI interaction will work with our target population” | “actually first time building a chatbot” |
| Adhyayan Foundation | “Building the capacity of the internal team to use AI and chatbot tools regularly” | “mentor support and fast query resolution” |
| MukkaMaar | “handling real, unstructured user input” | “such a beautiful bot (handling multiple categorisations)” |
Two patterns stood out. Concerns that were abstract in April (hallucination, uptake, unstructured input) became things teams had tested and, in several cases, solved. And fears about people did not disappear: one participant who worried in April about how an AI interaction would suit a population needing individual attention still wrote in September about “the apprehension about how our users will accept the chatbot”.
The six-month accelerator experience
These questions were asked only at the endline, so they describe the whole cohort of respondents (28 people from 19 organisations) rather than a before-and-after change.
Endline ratings, average out of 5 28 respondents
100% rated the overall experience 4 or 5 and none went below 4. On-the-ground impact is rated lower than the experience (3.79 against 4.36), which fits what participants wrote about adoption: building the chatbot was within their control, getting every user to adopt it was not. Logistics and pre-workshop communication (4.11) also scored well.
Confidence about moving forward on their own was high: all 28 respondents who answered rated 3 or above and 75% rated 4 or 5.
In participants’ words
What worked
“The mentor office hours were very helpful for all the support I needed.”
Protsahan India Foundation
“Selfless support from the glific team, especially the mentor and the flexibility of the platform to integrate with our existing digital platforms.”
Youth Dreamers Foundation
“Regular follow ups and support received from the team mentors was very useful.”
Learning Links Foundation
“Learning about AI product evaluation and building a chat bot from scratch with community support ensured our rich learning in the accelerator”
Collective Good Foundation
“Having the in-person accelerator in April which helped deep dive and troubleshoot.”
Point of View
“Induction meeting and guidance from our mentor.”
India Literacy Project
| Theme in “what worked” (28 answers) | Mentions | Share |
|---|---|---|
| Mentor, office hours and fast support | 16 | 57% |
| Learning by building (flows, evals, first chatbot, pilots) | 14 | 50% |
| Platform features (image evaluator, speech-to-text, translation, functionality) | 8 | 29% |
| In-person opening and need-based webinars | 4 | 14% |
Having two chances to pilot also mattered. All six award winners used feedback from their first pilot to refine their bots, improve the experience and strengthen user engagement in the second.
What did not
“Bandwidth issues, no dedicated person working on Chatbot.”
Endline participant
“On ground hiccups on technology adoption and initial low response rate to HSMs was a challenge”
Endline participant
“The apprehension about how our users will accept the chatbot”
Endline participant
“first three months were slow due to schools being shut for summer vacations”
Endline participant
| Theme in “what didn’t work” (26 answers) | Mentions | Share |
|---|---|---|
| No dedicated person or bandwidth | 8 | 31% |
| On-ground adoption, connectivity, response rates | 7 | 27% |
| Platform asks (handover, knowledge base, interface, URLs) | 6 | 23% |
| Nothing to add | 4 | 15% |
Suggestions were concrete: another hands-on session mid-way, monthly milestones to track progress, and more co-working time with other organisations.
What the surveys point to on the ground
The feedback is backed by programme outcomes. 28 of 34 NGOs submitted a pilot report and 28 are continuing, with paid subscriptions beginning on 1 October. At the close-out, four winners shared ₹2,20,000 in prizes, and two honourable mentions each received $100 in WhatsApp messaging credits:
- Protsahan India Foundation (1st): trauma-informed guidance for caregivers. It launched its pilot on 22 May with a government official present and stayed on track through a change of point of contact. Its second pilot onboarded 164 people in Bihar and 168 in Manipur and Assam, taking CareBuddy to four states (see its story).
- Collective Good Foundation (2nd): climate-health alerts and guidance for frontline workers, with live weather data and a custom dashboard. In the second phase, sanitation workers and community leaders send photos of unhygienic public spots so the municipality can act quickly.
- Foundation for Responsive Governance (joint 3rd): with government approval, block-level officers in Meghalaya used the chatbot to see maternal-health data for their blocks during field visits. The second pilot added dashboard access, feedback flows and content managed through Google Sheets.
- Youth Dreamers Foundation (joint 3rd): an AI assistant for scholarship queries, linked to Moodle so students can explore scholarships and check status and renewals. The second pilot added phone (IVR) access through Exotel.
- MukkaMaar (honourable mention): an AI classifier that reads adolescent girls’ free-text answers and gives personal feedback, now running inside its main self-defence flows.
- ATREE (honourable mention): biodiversity-friendly farming guidance for Sikkim and Darjeeling, with a Nepali-language flow built at users’ request and live weather data. Its AI evaluations scored well.
Six NGOs are not continuing, and for five of them the reason was staff or money, in line with the “no dedicated person” theme above. IBD India and Fadhila Health had promising pilots and are looking for funding to return. The board of Women’s Organisation for Rural Development decided to stop because low literacy and digital remoteness among the farmers it serves made it hard to reach enough users to justify the cost. Manav Vikas Sansthan and Youth4Jobs Foundation had no one who could own the chatbot. SELCO Foundation found its existing WhatsApp integration already met its needs; it paused its subscription and may revisit later.
Two further obstacles came up outside the surveys. Some pilots waited weeks on internal or government sign-offs, or on something as simple as a SIM card for the bot’s number. And Meta’s increase in WhatsApp pricing worried several NGOs about running costs.
Who the pilots reached
The NGOs’ own pilot reports give reach figures for 24 of the 34 organisations. Together they show how far the chatbots went in six months, who used them and in which languages.
| People reached in pilots | About 2,300* across 24 NGOs, from 26 government officers in Meghalaya to 507 teachers and school leaders in Tripura, Arunachal Pradesh and Goa |
| Messages exchanged | Over 38,000* at the 4 NGOs that reported a total: Collective Good Foundation (15,186), Adhyayan (13,300), The Kind Citizen (6,070) and Protsahan’s first pilot (3,488) |
| Languages | 13 languages plus Hinglish, with voice input built into at least 10 bots |
| Geography | At least 19 states and union territories, plus pan-India groups reached online |
Languages the bots worked in
Assamese · Bengali · English · Hindi · Hinglish · Kannada · Malayalam · Marathi · Nepali · Odia · Punjabi · Tamil · Telugu · Urdu
Voice mattered most where reading and typing were hard. At ATREE, voice was the main way farmers used the bot, mostly in Nepali; Collective Good Foundation’s health workers called voice notes the most appreciated feature; and Ayang Trust took questions in Assamese by voice from farmers in Majuli.
Where the pilots ran
Andhra Pradesh · Arunachal Pradesh · Assam · Bihar · Delhi · Goa · Jammu and Kashmir · Maharashtra · Manipur · Meghalaya · Odisha · Puducherry · Punjab · Sikkim · Tamil Nadu · Telangana · Tripura · Uttar Pradesh · West Bengal
Who used them
- Frontline health and care workers: ASHAs, ANMs and community health officers (Collective Good Foundation, NariCare), creche workers (Mobile Creches) and women health entrepreneurs (Collective Good’s Sadhikas).
- Teachers and school leaders: Government school teachers and leaders (Adhyayan, Learning Links, Nilay, Saturday Art Class) and career-guidance facilitators (India Literacy Project).
- Students and young people: Scholarship applicants aged 16–22 (Youth Dreamers), Grade 9 students (Life Lab), care leavers aged 18–26 (Catalysts for Social Action) and young people aged 15–25 (Dream a Dream).
- Adolescent girls and women: Girls aged 11–18 (Milaan, MukkaMaar), women training as drivers (MOWO), and women and gender-diverse people online (Point of View).
- Farmers: Smallholders in Majuli, Assam (Ayang Trust), Sikkim and Darjeeling (ATREE), Punjab (Manav Vikas Sansthan) and Koraput, Odisha (WORD), and paddy producer groups (SELCO).
- Child protection and government: Social workers, counsellors, child-protection officers and police (Protsahan); block and state officers in Meghalaya (Foundation for Responsive Governance).
- Communities with specific health needs: People living with inflammatory bowel disease (IBD India); people living with HIV, sex workers and LGBTQ communities in Delhi (Deepshikha Samiti).
- NGOs, volunteers and vendors: NGOs, CSR teams and individual volunteers (The Kind Citizen) and clean-energy technology vendors (SELCO).
*Staff-only tests and NGOs without a head count are left out. Definitions of “reached” and “engaged” vary between NGOs, so these totals are indicative.
Stories from the field
Excerpts from three NGOs’ accelerator journeys show what the programme looked like when it worked at scale, when it was blocked, and when trust was the first thing to build.
CareBuddy · Protsahan India Foundation
Child protection · Delhi, Bihar, Manipur and Assam · over 400 people reached
Frontline child-protection workers are often the first or only contact when a child is in trouble. CareBuddy gives them three things on WhatsApp: AI guidance for a live situation, grounded in POCSO, the Juvenile Justice Act and the NIMHANS–SAMVAD guide; short videos with an AI quiz; and art- and play-based practices for their own wellbeing.
It ran in three rounds: about 75 people in Delhi in May and June, 164 in Bihar over six weeks, and 168 in Manipur and Assam. Users were social workers, counsellors, ASHA and Anganwadi workers, police and child-welfare officials. They completed 1,093 flows, 89% would recommend CareBuddy, and some Bihar users came back 15 to 17 times. Five more states (Tripura, Mizoram, Meghalaya, Punjab and Jharkhand) asked to join without being approached.
Pilot 1 users found its mandatory-reporting tone off-putting, so the team rebuilt the knowledge base around the seven-step SAMVAD guide while keeping a legal safety floor underneath. When a user disclosed a live situation inside the wellbeing section, the bot paused and pointed them to urgent help and 112.
“Who supports the supporter? Frontline caregivers hold everyone else’s crisis, often with no one holding theirs.”
Avantika Jain, Associate Director, Programs, who built CareBuddy without a dedicated tech team
Sakhi · Peepul
Education · Grade 6–8 maths teachers · Hindi-first
Sakhi is built for maths teachers in government schools, who lose workshop support once they are back in class. Teachers pick a grade and chapter from Hindi menus and get a lesson plan, a hook, a quiz or an assessment grounded in NCERT textbooks and Peepul’s training material, or ask a question by voice or text.
The live teacher pilot could not run because permission to put Sakhi in front of teachers did not come through in time. Instead, about 25 Peepul programme staff in Bhopal tested it and raised 20 issues in six themes, including search returning the wrong chapter, PDFs with tables and images breaking, message delays, long replies and having to type “Hi” first. The team narrowed the scope to maths for Grades 6–8 and is converting its sources to Markdown ahead of a structured teacher pilot. It is the clearest case for the recommendation to clear approvals before the kickoff.
“Once we saw a teacher-ready script appear in seconds, we knew this was worth getting right.”
Peepul team
BalSneh · Mobile Creches
Early childhood care · creche workers and caregivers · 82 people across two pilots
BalSneh answers questions on child nutrition, health, immunisation, sleep, feeding and early development, based on Mobile Creches’ childcare guidelines. The first pilot, on 10 June, brought together 29 people, mostly creche workers with their supervisors and programme staff. The second, on 2–3 September, onboarded 53 caregivers through project staff, and the team reports more accurate answers and stronger trust than in the first round.
The first question was whether caregivers would trust an AI at all, and what is at stake if it gets something wrong:
“If the chatbot suggests a wrong diet and a child becomes ill, parents may blame the caregiver.”
Mobile Creches pilot report
Next, the team plans to validate its content with child-health experts and widen its reach.
What this means
- Teach evaluation early and in local languages. AI evals produced the biggest gain, and participants still asked for more ways to evaluate responses, especially in Nepali, Hindi and Hinglish.
- Treat mentoring as the core of the model. It is the most cited reason the programme worked, so the cost of mentor time is the cost of the outcome.
- Clear approvals before the kickoff. Get internal and government sign-offs, the bot’s phone number and its SIM in place before teams arrive, so pilots do not wait weeks to start.
- Plan for the second half. Add a mid-way hands-on session and monthly milestones, as participants suggested, and require a named owner and a backup.
- Add usage data to the next update. Flow usage, users reached and OpenAI credit use per NGO would turn self-reported capability into measured adoption.
The cohort by sector
The 28 organisations continuing after six months.
| Sector | NGOs | Organisations |
|---|---|---|
| Education | 9 | Adhyayan Foundation, Diksha Foundation, India Literacy Project, Learning Links Foundation, Life Lab (WOSCA), Nilay Foundation, Peepul, Saturday Art Class, Youth Dreamers Foundation |
| Gender equality and women’s empowerment | 4 | Deepshikha Samiti, Durga India, MOWO Social Initiatives, Point of View |
| Health | 3 | Samyak (Collective Good Foundation), NariCare, Saukhyam Foundation |
| Child welfare and protection | 3 | Catalysts for Social Action, Mobile Creches, Protsahan India Foundation |
| Girl safety and empowerment | 2 | Milaan Be The Change, MukkaMaar |
| Agriculture and climate | 1 | Ayang Trust |
| Disability and inclusion | 1 | Nayi Disha |
| Environment and conservation | 1 | ATREE |
| Youth empowerment | 1 | Dream a Dream |
| Governance | 1 | Foundation for Responsive Governance |
| Criminal justice | 1 | Prison Aid + Action Research |
| Volunteering and CSR | 1 | The Kind Citizen |
Further reading
Blog posts by the Glific team on their experience of the accelerator:
- Glific AI Chatbot Accelerator: A Glimpse: Radhika Bhagwat, Head of Glific, on her observations, learnings and objectives from the April kickoff.
- Building for Impact: Inside My First AI Accelerator Experience: mentor Fawas Chemba on his first Glific accelerator, after the April kickoff.
- The Weather Changed. So Did the Journey.: Fawas Chemba’s closing learnings and reflections from the September close-out.