LSE Students Map How Humanitarian Orgs Actually Use AI in New ACCORD Report

While the AI industry obsesses over agent benchmarks and trillion-parameter rumours, a quieter question has been gnawing at the humanitarian sector: what should aid organisations actually do with this stuff? A new report from a team of MSc students at the London School of Economics and Political Science takes a stab at answering it — not with hype, but with interviews and desk research aimed at one specific slice of the problem: knowledge management.

The report, flagged on 9 June by ecoi.net, was produced for ACCORD — the Austrian Centre for Country of Origin and Asylum Research and Documentation, the outfit behind ecoi.net itself — as part of a consultancy project run through LSE's Department of International Development. That pairing matters. ACCORD's bread and butter is country of origin information, the carefully sourced research that feeds into asylum decisions across Europe. If any corner of the humanitarian world needs to get AI adoption right rather than fast, it's this one.

A deliberately narrow lens

The smartest thing about the report may be what it leaves out. The students drew a hard line around internal research functions — document summarisation, metadata tagging, translation support, and database management — and explicitly excluded operational AI uses like logistics or aid distribution. According to ecoi.net, that scoping was deliberate, and it shows a maturity that a lot of corporate AI strategy documents lack.

Why does the boundary matter? Because the risk profiles are wildly different. An AI system that mis-tags a document in an internal archive creates friction; an AI system that misallocates aid or botches a needs assessment can cost lives. By bracketing off the operational side, the report can talk about adoption in the lowest-stakes, highest-leverage zone — the drudgery layer of humanitarian work, where researchers drown in PDFs, policy papers, and multilingual source material every single day.

That drudgery layer also happens to be where current-generation language models are genuinely good. Summarising a 200-page situation report, tagging documents with consistent metadata, producing a working translation of a source in a language the team doesn't read — these are tasks where models perform well, errors are catchable, and a human stays firmly in the loop. The report's scope reads like a tacit endorsement of the 'boring AI first' school of adoption, and frankly, more sectors should subscribe.

Who they talked to, and how

Methodologically, this is a mixed-methods effort: a desk-based review of academic literature, AI policy documents, and strategy papers from humanitarian organisations, paired with semi-structured interviews conducted remotely between December 2025 and March 2026. The interview roster is worth noting — representatives from CARE International UK, UNESCO, the European Policy Centre, and UNDP, alongside specialists in AI governance and humanitarian action.

That's a revealing cross-section. You've got a major operational NGO (CARE), a UN normative agency that has been pushing AI ethics frameworks for years (UNESCO), a Brussels think tank watching the regulatory weather (European Policy Centre), and the UN's development arm (UNDP), which has been experimenting with AI tooling across country offices. In other words: the people writing the policies, the people implementing them, and the people who have to live with the consequences when the two don't match.

The interview window is also telling. December 2025 to March 2026 captures the sector mid-adjustment — after the initial generative AI panic-and-pilot cycle of 2023–2024, but while organisations are still converting ad-hoc ChatGPT usage into actual institutional policy. Snapshot research from this period is going to age fast, but it's exactly the moment when documentation is most useful: the habits being formed now will calcify into the sector's defaults.

Why the COI angle raises the stakes

Here's where the GLSRM-reader context kicks in. ACCORD isn't a generic NGO — country of origin information sits in a legal pipeline. COI research gets cited in asylum tribunals; its accuracy can determine whether a person is deported into danger. That makes AI-assisted knowledge management in this domain a fascinating stress test for everything the industry says about reliability.

Summarisation is the obvious flashpoint. Anyone who has shipped a retrieval-augmented summarisation product knows the failure modes: dropped caveats, flattened nuance, confident misstatements about what a source actually says. In consumer apps, that's an annoyance. In a research function that feeds asylum adjudication, a summary that strips out a critical qualifier — 'security has improved in the capital, but not in the provinces' — is the kind of error that lands in front of a judge. Translation support carries similar weight: machine translation of, say, a regional-language news report about persecution is enormously useful and enormously easy to over-trust.

Then there's the data protection question lurking under 'database management.' Humanitarian organisations hold some of the most sensitive personal data on Earth — information about refugees, persecuted minorities, and people in active conflict zones. Every decision about which AI tools touch which databases is implicitly a decision about whether that data transits a commercial API, where it's processed, and who could subpoena it. The report reportedly includes AI policy recommendations, and if those recommendations grapple seriously with vendor selection and data residency, they'll be worth reading well beyond the aid sector.

What builders and policy watchers should take away

For AI builders, the signal is straightforward: there is real, unglamorous demand in the humanitarian and public-interest space for tooling that does summarisation with verifiable citations, metadata tagging with audit trails, and translation with confidence flagging. These organisations don't need agents that book flights. They need systems that can be defended in front of an oversight board — provenance, logging, human review checkpoints. That's a product spec, and it's one most general-purpose AI tools still don't meet out of the box.

It's also worth appreciating the delivery mechanism here. LSE's International Development department has been running consultancy projects that pair MSc students with real client organisations for years, and ACCORD using one to map its AI options is a shrewd, low-cost move — the kind of capacity-building that resource-strapped documentation centres can actually afford. Expect more of this: universities as the de facto AI strategy consultants for the nonprofit world, filling a gap that McKinsey prices most NGOs out of.

The skeptical read? Student consultancy reports vary in depth, interview-based research captures what organisations say rather than what they do, and a four-month window in a field moving this fast guarantees some findings will be stale by autumn. None of that makes the exercise pointless — it makes it a baseline. The thing to watch next is whether ACCORD and the wider COI community translate the recommendations into published, binding AI usage policies, and whether the EU's regulatory machinery starts treating AI-assisted asylum research as the high-stakes application it plainly is. When the tools that summarise persecution reports meet the rules that govern high-risk AI, things will get interesting. This report is an early field note from that collision.