Room to try AI with real beneficiaries
Over the past year, OpenAI credits have given NGOs on Glific (an open-source chatbot platform) room to try AI with real beneficiaries before committing their own budgets. Small and mid-sized NGOs rarely have a budget line for AI experiments, and finance teams are slow to approve API spend that varies month to month. The credits removed that barrier. Kaapi, our open-source AI layer, connects Glific’s WhatsApp chatbots to OpenAI models, so the credits went directly into the conversational flows, knowledge-base search and voice features that NGOs built and tested in the field.
34 NGOs, from first chatbot to field pilots
The clearest results come from the Glific AI Chatbot Accelerator. Between April and September 2026, 34 Indian NGOs took part, and most of them had never built an AI chatbot before. Each NGO had a dedicated mentor and ran two pilots in the field. Together, the pilots reached about ~2,500 people across 24 NGOs, in 13 languages plus Hinglish and in at least 19 states. Users included frontline health workers, teachers, farmers and child-protection staff. Protsahan India Foundation’s CareBuddy guided child-protection workers in four states, and 89% of users said they would recommend it. ATREE’s farmers mostly asked their questions by voice, in Nepali. On average, participants rated their own AI skills at 1.75 out of 4 at the start and 3.42 at the end. 28 of the 34 NGOs (82%) are continuing expand their pilots into their existing programs. AI Cohort 3.0, which focuses on climate organizations, started this month with LLM credits built into the program.
Evals: testing a bot before beneficiaries rely on it
The credits also let us build what NGOs need before they can trust AI with their communities. Evals 2.0, now live in Glific, uses an AI judge to score each bot answer from 0 to 5. It checks whether the answer is correct, whether the NGO’s documents back it up, and whether it follows the prompt. A 500-question test run now takes under four minutes instead of about 45. Between September 2 and October 17, 20 NGOs ran 220 evaluations. Deepshikha Samiti, which runs an HIV testing helpline in Delhi, found its bot giving the wrong testing-center hours in Hindi and Hinglish and making up a hospital room number. The team fixed both within 46 minutes, and its score rose from 4.29 to 4.72. In the Accelerator, 19 of 22 participants rated their team good at running evals by end of cohort, while none knew what Evals was when they started.
Guardrails: safety checks on every message
Kaapi (our open-source AI platform) has also shipped four open-source Guardrails that NGOs can switch on for each assistant in Glific:
- Privacy Protection hides Aadhaar, PAN, phone numbers and other personal details before a message reaches the model.
- Content Safety filters abuse in Hindi, English and Hinglish.
- Scope Control blocks topics the NGO chooses, such as medical claims or legal advice.
- Gender Neutrality rewrites gendered assumptions in the bot’s replies.
The checks run both before a message reaches the assistant and before the reply goes back. To keep guardrails fast and affordable on every message, we benchmarked OpenAI’s Decisions API against three LLMs. On our 21-query test set, it was 100% accurate, gave the same answer every time, responded in 180 ms and cost ₹8.12 per 1,000 queries. Everything is released under AGPL v3, so any NGO or developer can inspect it, self-host it or build on it.
Along with the Glific Accelerator, we shared credits with AI cohort NGOs as well
Open AI credits removed the fear of failing and encourage them to experiment, try new ways of doing the same thing, failing and ietrating. As we wrote when kicking off AI Cohort 2.0, grants from OpenAI and Anthropic meant we could give participating NGOs enough credits to experiment, iterate and run a pilot. This was key and it mattered more than it sounds, because API spend is a variable, unbudgeted cost sitting inside restricted grant funding. Without credits, NGOs get one shot. With credits, they can afford to fail a few times first. That showed up in the work done during the AI cohort program. Samanvay Foundation built on the OpenAI API to help nonprofits create their own apps. By Cohort 2.0 we were running the same use cases across Claude, OpenAI and Gemini to see which model suited which task, which is the kind of comparison nobody can afford to make on a programme budget.
Thus, the mindset of experimentation helped NGOs get started on their journey to integrating AI into their workflows.
What more credits would make possible
More credits would help the next group of NGOs turn their pilots into lasting services. In the Accelerator, running costs and staff time were the main reasons a few NGOs dropped out.
Further Reading:
