Mercy Corps on what AI makes possible in humanitarian work

Mercy Corps, which will become Prosper Global later in 2026 as part of its organizational evolution, is a global humanitarian organization running more than 200 programs across 35+ countries. Nayid Orozco is its AI Solutions and Delivery Manager, responsible for the organization's Claude deployment, from administering the Enterprise workspace to setting the guidelines that govern what data can be used. We spoke with Orozco about Mercy Corps' 13-team pilot, what changes when field teams hold the tools themselves, and where AI fits in the future of humanitarian work. 

Anthropic: With 13 teams in the mix, what has AI actually changed for an organization like yours?

Nayid Orozco, Mercy Corps: In a lot of sectors, the debate is whether AI competes with people, or whether it just helps teams do what they already did a bit faster. In ours, it is different: AI lets us do things we simply could not do before or were never going to be resourced to do—even though we knew they mattered. Synthesizing and drawing insight from large bodies of evidence and learning used to be out of reach for us in most cases. Claude puts it within reach, and it does it at the speed our work moves at, which on the humanitarian side is fast. That is what lets an organization like ours take a real step toward being a learning organization rather than just aspiring to one.

Anthropic: You're the one running this deployment day to day. What does that involve, and who uses Claude most?

Orozco: I work in our Technology for Development group and serve as a link between the teams using Claude and our legal and data protection colleagues. The teams that use Claude most heavily are our Community Accountability team and our monitoring, evaluation, and learning (MEL) team.

Anthropic: What does the MEL team's work with Claude look like?

Orozco: MEL is the cornerstone for measuring the effectiveness and outcomes of our programs; it is core to helping us understand if we are improving the lives of people. A core challenge has been establishing consistent analytical frameworks not in just one program, but across our portfolio of more than 200 programs. That includes building and refining the logic models that are key to program and strategy design, indicator frameworks that establish the metrics by which we measure outcomes, processing datasets and running statistical analysis in real time, and turning messy source material into program evidence summaries, often across English, Spanish, and French, ready for donor reporting and for data-driven decision making. One advisor described Claude working through a full indicator framework, across an impact statement, four outcomes, and fourteen outputs, as a genuine thought partner rather than just a formatter.

Anthropic: Can you give us a concrete example or two of work that wasn't realistic before?

Orozco: A market dataset analysis for our Sudan research that would have taken a very different and less rigorous form, if it was possible at all, was compressed from about a week into a day and a half. The processing script that came out of it is now in production, allowing for repeated analysis that will build a larger understanding of market disruption and shifts in one of the most vulnerable places in the world. 

In Pakistan, Claude helped our team diagnose a complex failure in the DHIS2 health information system we support for the national TB program, across 7 provinces and more than 15,000 facilities, in a single conversation that would normally take a senior developer several days, if not weeks. And on the research side, a design we estimated at roughly 280 hours was completed in about 20. It is not only the hours saved. The depth of work has expanded significantly, allowing teams to conduct more in-depth analysis and modeling than was previously possible.

In Kenya, our JobTech Alliance team offers another example: its members, who do not come from a software engineering background, built their own MCP server to query a million-row SQL database. That’s the kind of technical capability that used to require dedicated engineering support.

Anthropic: Some of this work now sits with country teams rather than at headquarters. How do you think about that shift?

Orozco: That shift is immense. Country teams hold the deepest knowledge of the contexts they work in, and giving them Claude lets them build on that contextual knowledge and do their own analysis instead of waiting on the center.  We've learned critical lessons about bringing field teams on board in new contexts. They need persistent context, so they are not rebuilding background every session. They need version-aware accuracy on technical systems, a low-bandwidth option for remote districts, and short practical training rather than long sessions. The result is more direct program action, more data-driven decision making, and a more organic approach to adaptive management, one that moves away from traditional top-down models toward locally led action. More than any single efficiency gain, that is what we think matters most about expanding Claude to the field.

Anthropic: How do you think about data protection and human oversight?

Orozco: Two things protect the communities behind the reports. First, on data: we screen and redact direct identifiers and the most sensitive content before anything goes into Claude. Inputs and outputs are retained for a default period for abuse detection and are then deleted. Under our Enterprise commercial terms, customer data is never used to train the models. We documented all of this in a formal privacy impact assessment. Second, on oversight: we kept a human in every loop that mattered.

Anthropic: Was there anything in how staff responded that you didn't expect going in?

Orozco: What surprised me most was how often people stopped describing Claude as a productivity tool and started calling it a thought partner. One person said it asks them the right questions to get to a better answer, which is not what I expected to hear about software.

Anthropic: What would you tell another humanitarian organization that wants to use AI responsibly but isn't sure where to start?

Orozco: Start small and bounded, and put data protection first rather than last. Do a privacy impact assessment for each real use case before you scale it, because the risk lives in the use case, not in the tool. Pick tasks that are high volume or have a high cognitive load, because that is where the value shows up fastest. Keep a human in the loop for anything that affects a person's safety or dignity. And invest in training and shared prompts early, because the difference between people who love the tool and people who bounce off it is almost always whether someone showed them how to use it well.

Anthropic: Where does this go next for Mercy Corps?

Orozco: We work with a wide range of mostly local partners, and we see real opportunity in helping them access and use AI too, not only using it ourselves. Much of the sector is focused on its own use of AI. We think the bigger prize is improving the performance of local and hyper-local actors, because that is where impact is actually made.

Some of this is already taking shape across the Impact Alliance, a collaborative effort across three leading NGOs to work in a more connected, efficient, and locally led way. The alliance includes Mercy Corps, Save the Children, and CARE. In that group we have a CommunityInsights Engine workstream, where community feedback will be analyzed using Claude. That points to where this can go beyond Mercy Corps alone. 

Our Innovation Catalyst platform is another effort taking shape, weaving AI for social impact through our impact investing, innovation, and market-facing capabilities.

Anthropic: Where does AI take the humanitarian sector from here?

Orozco: AI can move us from a sector that consistently underutilizes the learning from its own work to one that uses it consistently and efficiently. That change feeds better design, better implementation, and greater impact for people in the places where we work. Claude in particular is well suited to this, because it is strong at operational research, learning work, and analyzing qualitative datasets. That’s exactly the kind of data learning our work depends on. The way we most want to influence where AI goes in humanitarian work is pushing the benefits outward to local actors, rather than keeping them inside large organizations.