.png)
AI in practice: Examples from CHEW members
Ella Nation-Tellery, Impact and Learning Manager at ClientEarth
Key takeaway: The most useful applications of AI start with a clear evaluation question and combine AI with human expertise and other research methods.
In our first blog, we shared practical recommendations for getting started with AI, from linking it to your strategy and values to setting your own commitments. In this second blog, we look at what that means in practice. Drawing on examples shared by members during our peer learning sessions, we explore how organisations are already using AI in their work, from qualitative data analysis to synthesising large volumes of data at scale.
Here we share a range of examples.
Synthesising large volumes of data
ITAD
What they did: ITAD used Co-pilot 365 to analyse and synthesise the large sets of documents spanning multiple years using the process outlined below:
They also developed a framework to assess the accuracy of the outputs and their confidence in them, rating the AI outputs against completeness, relevance, accuracy, and fabrication.
Why it’s useful: It helped funders make sense of their portfolios of grants over several years.
What to watch out for: More is not necessarily more! Ingesting everything, rather than using a sample risks outsourcing sense-making to machines. AI can miss nuances, use it as an aid in analytical processes to spot and validate patterns.
The Health Foundation
What they did: They developed a Knowledge Catalogue and AI Assistant to help make better use out of the data and insights gained from their research. They worked with their IT team and an external developer to collate all the outputs for the catalogue, clean and standardise the data for the catalogue and develop the AI assistant.
Why it’s useful: Staff can now ask questions on previous research to support scoping, business planning and responding to external requests.
What to watch out for: This demands a lot of capacity, accommodate accordingly and work closely with a developer to understand the different options and models available to balance cost, accuracy and coverage.
Supporting qualitative analysis
What he did: Steve used AI to conduct interviews, code answers (using human evaluator input), and cluster factors mentioned by sources together, helping people tell a story of what’s causing what in their world.
Why it’s useful: Harnessing AI in qualitative research can make it possible to work at greater scale, and may improve consistency and reproducibility, allowing for repeated monitoring and comparison across groups, places, timeframes and languages.
What to watch out for: Always use anonymised data and note causal mapping does not on its own warrant causal inference! AI can help identify and organise patterns in qualitative data, but this does not by itself establish causation. Human interpretation and appropriate research methods are still required.
Analysing sentiment and online conversations
Bradford Culture Company
What they did: They used Natural Language Processing (NLP) to analyse thousands of open-ended survey comments.
Why it’s useful: The NLP was able to code every open ended comment with a rating for sentiment, automatically sort comments into thematic categories and surface recurring themes across all events. This allowed the team to process comments at scale, surface patterns that might be missed manually, generate insight faster and free up evaluators time to focus on interpretation rather than data entry.
What to watch out for: Context can also be lost at scale and NLP categories always require human input and review.
Evaluating large events
Bradford Culture Company
What they did: they used AI to calculate crowd density and counterfactual footfall to evaluate Bradford City of Culture 2025.
Crowd density - the team collaborated with students at the University of Bradford extracting frames from drone footage, developing heat maps and overlaying them with audio recordings to estimate crowd density.
Footfall - a collaboration with Leeds Institute for Data Analytics (LIDA) tested and used multiple predictive AI models to predict what the city would have looked like without the event based on years of historical data, and compared the predictions vs the actual to assess the impact of the event.
Why it’s useful: Mass unticketed events leave no natural attendance record so counting footfall requires alternative methods. AI also allowed them to work at scale and more efficiently over hundreds of events with thousands of audience comments.
What to watch out for: Surveillance footage quality and camera angles affected the reliability and coverage gaps with models that calculate footfall mean that they can’t capture everyone. Aligning the different data streams (audio, visual, footfall) was also very complex.
What we've learned
Across all the sessions and areas, members shared many of the same recommendations:
Start with the question, not the tool: Let your research or evaluation questions shape how you use AI, not the other way round. Be clear about what you need to find out before choosing a tool.
Test more than one approach: Don't assume one model or machine learning approach will work best. Different tools perform differently depending on your data and purpose.
Combine AI with other methods: AI is most useful as one part of a wider mixed methods approach, not a standalone solution. Use it for different stages or components as relevant or appropriate.
Keep humans in charge of judgement and decisions: AI should support evaluators, not replace them. Use it to spot patterns or check your own findings, but leave interpretation, conclusions and decisions to people.
Plan for time and trial and error: Using AI well takes planning, iteration and resilience. It won't always work the first time, or at all, so build in time to experiment and have a back-up plan.
Build internal capacity: Invest in your team's skills and confidence so AI use doesn't depend on one person or an external provider.
How are you using AI in your work? We'd love to hear more examples and reflections from our members. Join the discussion in the ChEW LinkedIn group, or get in touch if you’d like to share your experiences at a ChEW event.
These blogs draw on the experiences shared through ChEW's AI peer learning sessions. We are grateful to Naomi Tyrell for leading a session on getting started with AI, to Rocket Science for co-facilitating the peer learning sessions, and to all the members who contributed examples, questions and lessons from their work. Thank you for helping make these conversations open, practical and useful to the wider network.
For anyone who missed the sessions, we encourage you to watch the recordings and access the accompanying resources in the Members section of the ChEW website.
Comments