An AI assistant that sounds confident isn’t necessarily right. In the first six weeks (2 Sep to 17 Oct) since launch of AI Evaluations (v2), 20 NGOs ran AI Evaluations; 19 of those ran it more than once, 14 came back on a different day, and together they ran 220 evaluations. Of these I have captured and shared the journey of five NGOs who found out qualitative ways to improve the bot. 

One caught its bot inventing registration links and had it scoring 5 out of 5 eight minutes later. Another discovered its own test set was the problem, rebuilt it, and went on to test 29 datasets across three languages. This blog shared what each of them found, what they changed, and the habit they share.

But before I go into the details from org’s eval runs, a bit about the process of running evaluations itself.

How AI Evaluations works

You upload a golden dataset: the questions your users actually ask, with the answer you’d want your assistant to give. Glific runs every question through your assistant, several times over, and an AI judge scores each answer out of 5 on three things: does it match the expected answer, does it follow your prompt’s rules, and is it grounded in your knowledge base.

You get an overall score, a verdict, and an AI summary. The summary is the part that matters most: it names the specific questions that went wrong and says where the fix probably lives, in the prompt, the knowledge base, the model or the test set itself.

FMCH: finding what the knowledge base didn’t know

Foundation for Mother and Child Health (FMCH)’s aim is creating a system that builds resilience in mothers by providing consistent, high-quality nutrition and care information, visit here to know more about their work

The assistant: PoshanSakhi, which answers mothers’ questions on pregnancy, breastfeeding and child nutrition.

What the eval found. FMCH ran 15 evaluations between 8 Sep and 7 Oct, and the summaries surfaced different kinds of problems.

• A language bug. In September, the bot replied in Hindi to almost every question asked in English, breaking its own same-language rule.
• Gaps in the knowledge base. On convulsions, an unresponsive baby, the Village Health and Nutrition Day and spacing between pregnancies, the knowledge base had nothing. The bot answered confidently anyway, with first-aid steps and clinical claims no one had approved.
• Prompt adherence issues: The bot sometimes failed to follow configured instructions, particularly around referrals, emergency responses and fallback guidance.
• Understanding context: The bot sometimes misunderstood the user’s question, such as confusing a mother’s diarrhea with a child’s diarrhea.
• Response quality: lengthy responses, even for simple questions, sometimes missed important recommendations.

What changed: By October, the language issue had been fixed, emergency responses followed the prescribed format more consistently, and answers became shorter and more suitable for interactions. Scores on the new evaluation datasets ranged from 3.8 to 4.2, with one run scoring 4.74. However, some knowledge-base gaps, particularly around VHND, persisted. The bot also continued to make unnecessary referrals to AWWs or health centres, even when the question did not require one. This highlighted the need for continuous evaluation and refinement of AI responses.

Still in progress: Knowledge-base scores remain low. Since the content is currently in PPT format, org plans to investigate whether the format is affecting the AI’s ability to retrieve relevant information. Org also found that the bot does not consistently share relevant videos, even when video links are available in the knowledge base. Both areas need further investigation and improvement.

Takeaway: AI evaluation is not a one-time exercise but an ongoing process. Every low score is an opportunity to learn, not just a sign of failure. Sometimes the fix lies in the knowledge base, sometimes in the prompt, and sometimes in how we evaluate the AI.

MVS: the test set was the problem

Manav Vikas Sansthan (MVS) works towards empowering rural communities with sustainable development, read more about their work here

The assistant: a farm advisory bot answering crop questions in English, Hindi and Punjabi.

What the eval found. MVS’s first five runs on 19 Sep scored between 2.52 and 4.00. The summary was blunt: the bot wasn’t the problem, the test was. Many golden answers were placeholders like “use locally recommended practice”, while the bot was giving specific, knowledge-base-backed advice. Other test questions were about wheat, which the paddy bot correctly declined to answer.

What changed. On 7 Oct, MVS came back with rebuilt datasets for paddy and wheat in all three languages. Scores jumped to 4.17–4.68. With the noise cleared, the summaries could point at a real bug: the bot kept asking farmers when they sowed, even when the question already said so.

What happened next. Over the following ten days MVS expanded to cotton, pear and poplar, added datasets just for language handling and off-topic questions, and started testing photo-based questions. By 17 Oct it had run 71 evaluations across 29 datasets, averaging 4.5–4.7.

Takeaway: a low score is a question, not a verdict. Sometimes the answer is to fix your test set first.

MOWO: a made-up registration process, fixed in eight minutes

Mowo Social Initiatives Foundation (MOWO), works towards making roads equal spaces for women, read more about their work here

The assistant: a bot that helps women in Hyderabad join a driving and livelihood programme.

What the eval found. In the programme, you register simply by messaging the bot; there is no form and no link. But when MOWO tested four registration questions on 12 Oct, the bot invented a whole process: application links, registration forms, camps and orientation sessions. It scored 1.25, then 1.51. The AI assistant was giving incorrect and incomplete information about training details (vehicles available, location, fees). Depicting high fall-back response rate. 

What changed. Within minutes of running the evaluation, they fixed most of those issues, reduced fall-back response and inaccurate information. 

Still in progress. Earlier runs (26 Sep) caught the bot inventing employer names and training durations. MOWO ran seven evals in an hour that day, lifting the score from 3.46 to 3.86, but the employer names were still being flagged on 13 Oct. Some fixes take more than one sitting.

Takeaway: a bot that invents a link can send real people to the wrong place. An eval catches it before they do.

“The current evaluation feature flags all incorrect responses made by the AI and automatically makes corrections in prompt structure without any manual intervention. This has helped save lot of time, energy and made the fixes fast and seamless.”

– Kaumudhi, Team member, MOWO

Deepshikha Samiti: getting the small facts right

Deepshikha Samiti (DS) works towards socio-economic development of communities with a main focus on women, MSM/Transgender and youth. Read more about their work here

The assistant: a helpline bot in Hindi, Hinglish and English for people looking for HIV testing and treatment services in Delhi.

What the eval found. DS’s first run on 2 Sep scored 4.29, a “Good” verdict. The summary still found errors that matter. In Hindi and Hinglish, the bot gave the wrong opening hours for testing centres (9 AM–4 PM instead of 10 AM–3 PM), though it got them right in English. It also invented a room number at one hospital, and on one question it looped until its reply broke.

What changed. DS ran six more evaluations that evening on the same 75 test pairs. Within 46 minutes the testing-hours error was no longer flagged and the score reached 4.65, then 4.72.

What happened next. With the facts fixed, the summaries turned to crisis handling: the bot sometimes asked for age and PIN code during a message about suicidal thoughts, and placed helpline numbers mid-reply instead of first. On 30 Sep, DS evaluated a second assistant for maternal health, which scored 4.45–4.47 on its first runs.

Takeaway: “Good” isn’t the same as right. For a health service, the wrong opening hours mean a wasted trip, and the eval checks those details in every language your bot speaks.

ATREE: checking a good bot before it grows

Ashoka Trust for Research in Ecology and the Environment (ATREE), is focused on environmental conservation and sustainable, socially just development. Read more about their work here

The assistant: a bot for spice farmers, answering questions on large cardamom and other crops in English and Nepali.

What the eval found. ATREE’s bot started strong, at 4.35–4.55 on 4 Sep. The summaries still found things worth fixing. In the Nepali test set, one question had been paired with the wrong reference answer, so it scored 0 every time through no fault of the bot. And on drying cardamom, the bot always described the traditional 35–40 hour method, never the improved dryers that take 17–24 hours.

What changed. When ATREE added red cherry pepper and dalle chilli in mid-October, it evaluated them before relying on them. They scored 4.62–4.67. The summaries then caught a subtler problem: the FAQ and the resource sheet gave different numbers for ploughing and seed rates, so the bot’s answer depended on which one it happened to read.

Takeaway: a good score doesn’t mean nothing to fix. Evals can find conflicts inside your own source material that no one would spot by reading it.

What the five have in common

None of these NGOs ran one eval and stopped. Between them they ran 117 evaluations, and the improvements came from going round the same loop: run, read the summary, fix one thing, run again.

Three habits stand out.

• They read the summary, not just the score. The score says how much is wrong; the summary says which questions and where to look. MVS’s low score came from its test set, and FMCH’s colostrum error came from its source material. Neither would have shown up in the score alone.
• They re-ran right away. MOWO fixed its registration answers in eight minutes. MVS, FMCH and DS each ran several evals in a single sitting.
• They kept their test sets. Re-running the same questions is what let DS and MOWO confirm their fixes, and it’s the quickest way to catch a change that breaks something. A golden dataset is worth keeping and growing.

Try it with your assistant

1. Write 20–50 real questions your users ask, with the answer you’d want. Include a few your bot should decline. Read more on how to put together the Golden QnAs here
2. Run an evaluation in Glific. Read more steps to run the AI Evaluations here
3. Read the summary’s top three issues. Check whether each points at your prompt, your knowledge base or your test set.
4. Fix one thing and run again. Keep the same questions so you can see what moved.
5. Re-run after every change to your prompt, knowledge base or model, so a fix in one place doesn’t break something elsewhere.

If you’d like help setting up your first evaluation or reading a summary, reach out to the Glific team at info@glific.org to learn more and get targeted help.