AI That Stays In-House at a Humanitarian NGO

With the shared reporting platform on Azure Databricks that we built for this NGO, the groundwork for AI was already in place: connected data, clear rules on who may access what, and an environment in which AI workloads can run under the same governance.

On top of that foundation we took on the role of AI coordinator. For this organisation, one requirement came before all others: much of its information is sensitive, so AI is only useful here if that information stays protected. Every project below started from that requirement.

After each project, project managers record what they learned. Over the years that has grown to thousands of lessons learned, and a colleague facing the same problem elsewhere could benefit from every one of them, if only they could find the right one. A search on exact words misses a lesson that describes the same problem in different words.

So we created a function app that connects to the organisation's monday.com, where project managers enter their lessons, and enabled vector search on top. Vector search turns each lesson into a series of numbers that captures its meaning, so a project manager can ask a question in plain language and get the lessons closest in meaning, whichever words they use. Ask a question below, and compare the two ways of searching.

The organisation also conducts interviews with people in areas hit by disaster. These conversations are personal and very sensitive, so that information cannot be exposed to third parties. Sending them to an outside service to be processed was not an option.

The same vector search works for these interviews, with one important difference: the vectorisation runs on a local model. The step that turns an interview into numbers happens entirely within the organisation's own environment, so the text never leaves. Switch between the two below and follow where an interview goes.

Before, these interviews were hard to find. Each one had to be tagged by hand, and a search only returned what carried the right tag. That is suboptimal by nature, because a tag can only capture a fraction of a conversation, and it took a team of people working full time on retrieving and categorising these tags.

With vector search, thousands upon thousands of interviews are indexed automatically and can be searched quickly by what they are about. The full-time work of retrieving and categorising tags is no longer needed. Pick a theme below. Each dot stands for an interview; we show no text, because that is the whole point.

People want to use AI in their work, and public AI tools are one click away. For an organisation that handles this much sensitive information, that is a risk: whatever someone pastes in leaves the organisation, and banning the tools does not take the need away.

So we implemented an internal AI assistant. Colleagues across the organisation can use it and talk to it the way they would with a public tool, but without the data going to the AI provider. That gives everyone a safe environment to work in. Pick a task below and see where the text travels.

The organisation works in many places, and colleagues everywhere need marketing material, often without a designer nearby. Left to their own devices, each location would produce something different, and the organisation would stop looking like one organisation.

So we also built an image generator with preset designs. People can easily create marketing material wherever they are, and it all stays in the same style. Choose a format and a message below, and switch the presets on.

Project managers now search thousands of lessons learned in their own words and find what colleagues learned before them. Interviews that could hardly be found before can now be searched quickly, without tags and without the full-time work behind them, and without the text ever reaching a third party.

Colleagues have an AI assistant they can use safely and an image generator that keeps every piece of marketing material in one style. And because it all builds on the foundation that the shared reporting platform laid, the NGO got practical AI without giving up control of its most sensitive information.

• The role of AI coordinator, building on the data foundation of the shared reporting platform.
• A function app connected to monday.com, with vector search, so project managers can search thousands of lessons learned in natural language.
• Vector search across thousands of sensitive interviews, using a local vectorisation model so the text is never exposed to third parties.
• An end to manual tagging, and to the full-time work of retrieving and categorising tags.
• An internal AI assistant, giving colleagues a safe environment without their data going to the AI provider.
• An image generator with preset designs, so marketing material can be created anywhere and stays in the same style.