A Practical AI Roadmap for National NGOs

National NGOs should start with AI in five stages: pick one repetitive, text-heavy, low-risk problem; put data protection in order before any pilot; run one 90-day pilot with a measured baseline and a named reviewer; write a one-page AI policy; and build capability through AI champions and a shared prompt library. Never enter beneficiary-identifying data into public AI tools, and keep AI away from decisions about people.

Published 2026-08-14 · By Shahzad Asghar

Over the past twenty years I have built data and technology systems for humanitarian and development organizations in Pakistan, South Sudan, Kenya, Switzerland, Jordan, and Lebanon. In the last two years, one question has followed me into almost every meeting with NGO leaders: where do we start with AI?

The honest answer is that most advice on this subject was not written for you. It was written for corporations with large budgets, full IT departments, and data that does not describe vulnerable people. National NGOs work under different conditions. Funding is tied to projects. The IT function is often one person, sometimes half a person. The data you hold is sensitive by definition. Your credibility with communities and donors is the asset you can least afford to damage.

This article is my attempt at a one stop starting point for that reality. It draws on systems I have built, mistakes I have watched organizations make, and a set of working documents that I now share with NGOs on request. Details on those are at the end.

Where this comes from

I write from practice, not theory. At UNHCR Jordan I led the Data Analysis Group for seven years. My team built an automated phone appointment system that served more than 700,000 refugees. We trained an Arabic language model that classified community feedback with 83 percent accuracy. We built a job matching platform with the Ministry of Labour that reduced placement time by 45 percent. We used open source tools to find duplicate entries across more than 680,000 registration records. None of this required a Silicon Valley budget. It required a clear problem, usable data, and patient testing.

Later, at the UN Economic and Social Commission for Western Asia, I governed a portfolio of more than 90 digital platforms and contributed to AI governance work across the Arab region. That experience taught me the other half of the story. Whether an organization succeeds with technology has very little to do with the technology itself. It has to do with how the organization prepares, decides, and follows through. The roadmap below is built on that lesson.

The roadmap in five stages

Stage one: start with a problem, not a tool

Do not begin by asking which AI tool to buy. Begin by listing the tasks that consume staff time without requiring much judgment. Every NGO has them. Field visit reports that take a day to summarize. Donor reports that start from a blank page every quarter. Documents that wait weeks for translation. Meeting minutes that nobody writes.

Good first candidates share three features. They are repetitive. They are text heavy. And they are low risk if the first draft is wrong, because a person checks the output before it goes anywhere. Bad first candidates are the opposite: anything that makes a decision about a person, anything that faces a beneficiary directly, and anything that touches money. You will get to some of those later, with proper controls. You should not start there.

Write down three candidate problems. Pick one. That single choice will do more to shape your success than any tool comparison.

Stage two: put your data in order before anything else

Before any pilot, answer three questions in writing. What data will staff be allowed to put into the tool? Where does that data go once it leaves your computer? Who could be harmed if it leaked? If you cannot answer these, you are not ready to pilot. You are ready to prepare.

Preparation is not glamorous. It means agreeing where files live, deleting data you no longer need, and drawing one bright line for everyone: no names, case numbers, phone numbers, addresses, or any detail that identifies a person you serve is ever entered into a public AI tool. Free tools in particular may use what you type to improve their models. Treat every prompt as a document you might one day have to disclose.

This stage matters more for NGOs than for almost any other kind of organization. A company that leaks customer data faces a fine. An NGO that leaks beneficiary data can put people in danger and destroy community trust that took a decade to build. Data protection is not the compliance step in this roadmap. It is the foundation.

Stage three: run one small pilot with clear success measures

One process, one team, one tool, ninety days. That is the whole formula.

Spend the first two weeks measuring the current state. How long does the task take today, and what does acceptable quality look like? Without this baseline you will end the pilot with opinions instead of evidence. Spend the next two weeks choosing a tool, writing one page of rules, and training the three to five staff members who will use it. For the following six weeks, run the process with AI assistance, with a named person reviewing every output before it is used. In the final two weeks, compare the results against your baseline and decide: scale it, fix it, or stop it.

All three outcomes are respectable. A documented decision to stop is worth far more than an undocumented habit that quietly spreads through the organization.

Stage four: write a one page AI policy

Long AI policies do not get read. Write one page. It should state what staff may use AI for, what they must never enter into it, who reviews outputs before they leave the organization, how you disclose AI use to donors and communities when it matters, and who to ask when something is unclear.

Assume that some of your staff already use these tools quietly on their phones. A policy that only says no pushes that use underground, where you cannot see it or guide it. A policy that offers a clear and safe channel brings it into the open. Most people follow rules that are short, reasonable, and explained.

Stage five: build capability inside the team

Appoint one or two AI champions. They do not need to come from IT. The best champion is usually the person who cannot stop fixing broken processes. Give them time, a modest budget, and direct access to leadership.

Train through real work, not generic workshops. A two hour session where staff apply a tool to their own reports teaches more than a full day of slides. And build a shared prompt library. When a prompt produces a good result, save it in a shared document with a note on what it is for. This is the cheapest knowledge management system your organization will ever run.

Budget honestly. The subscription is the smallest cost. Staff time to test, review, and train is the real one. Then revisit the whole effort every quarter and decide with evidence, not enthusiasm.

Practical tips from the field

1. Start internal. Choose first use cases where the reader of the output is a colleague, not a beneficiary or a donor. The cost of an error is a correction, not a crisis.
2. Keep AI away from decisions about people. Screening job applications, selecting beneficiaries, or evaluating staff with AI carries legal and ethical risks that a national NGO does not need. Draft with AI. Decide with humans.
3. Test in your working languages. A model that performs well in English may struggle with Arabic dialects, Urdu, or local languages. We learned at UNHCR that community feedback in dialect needed dedicated work before any model could read it reliably. Run your own test before you commit.
4. Name the reviewer before the pilot starts. Not after the first mistake. Nothing produced with AI should leave the organization without a person who has put their name on it.
5. Ask vendors for references. When a company promises an AI solution, ask for two references from organizations of your size and sector. If they cannot produce them, then you are the pilot, and you should be paying pilot prices.
6. Prepare for the donor question. Donors have started asking grantees how they use AI. A one page policy and one documented pilot puts you ahead of that conversation instead of behind it.
7. Ignore the news cycle. Tools change every month. Your problems change slowly. An organization anchored to its problems can switch tools easily. An organization anchored to a tool is stuck when the market moves.

Ideas worth borrowing, by function

For program and monitoring teams, the strongest early wins are summarizing field reports, producing first drafts of indicator narratives, and coding open ended survey responses for analysis. Translation of qualitative feedback for internal use also saves real time.

For grants and fundraising teams, AI is useful for first drafts of donor reports built from your own data, for adapting an existing proposal to a different donor format, and for background research on prospective funders. The judgment, the numbers, and the relationships remain human work.

For communications teams, start with translation drafts, with adapting one report into a press release, social posts, and talking points, and with subtitling videos. Every output still needs an editor who knows the organization's voice.

For operations, HR, and finance, the safe ground is drafting job descriptions, summarizing long policy documents, and producing meeting minutes. Keep AI out of final decisions on hiring, performance, and payments.

For leadership, use AI as a briefing assistant. Ask it to explain a new regulation, to summarize a long strategy document before a board meeting, or to argue against your own plan so you can find the weak points before someone else does.

Documents available on request

Over the past year I have turned this roadmap into a set of working documents that I share with NGOs at no cost. They currently include an AI readiness self assessment, a one page AI use policy template, a ninety day pilot plan template, a data protection quick screen for AI use cases, a vendor evaluation question set, and an outline for a half day staff introduction session.

If any of these would help your organization, request them through the contact page on this site and mention which one you need. My only ask in return is feedback on what worked and what did not, so the documents keep improving for the next organization.

For deeper reading on responsible data handling in our sector, two free resources are worth your time: the ICRC Handbook on Data Protection in Humanitarian Action and the OCHA Data Responsibility Guidelines. Both were written for humanitarian and development contexts rather than for corporations.

The point

AI will not rescue a weak process. It accelerates whatever you already do, good or bad. So fix the process first, protect the data always, and automate second. Start with one problem, run one honest pilot, write one page of rules, and build the skills to repeat that cycle.

National NGOs carry missions that matter and budgets that do not forgive waste. Used with discipline, AI can return hours to the people doing the actual work: the program officer in the field, the grants manager against a deadline, the coordinator holding it all together. That is the only measure of success worth tracking. The mission comes first. The technology is only useful when it serves it.

← All articles