Six months building an AI back office for a non-profit
We spent the last six months replacing spreadsheets, paper forms and inbox archaeology at a working non-profit. Here is what we built, why we built it ourselves, and what it changed.
Most writing about AI for non-profits is about what could happen. This blog is about what did.
For the last six months Murmur AI has worked with one non-profit in India. It has a few dozen staff, a large volunteer base, grant-funded programmes and field teams in several states. The brief was to take the back office, the part of the organisation that donors never see and staff never have enough time for, and rebuild it around AI.
We aim to publish one post a week on how that went. This first one is the overview.
What we built
We don't name the organisation or the tools here. What each tool does matters more than what it is called.
An HR assistant. Staff ask it policy questions in plain language and get answers drawn from the organisation's own policy documents, with the source document and page shown beside each answer. It also runs leave applications from start to finish: checking the leave balance, applying for leave, manager approval, cancellation, emergency leave after the fact. There is no paper form behind it. The conversation is the form.
A volunteer management system. It tracks who volunteers, what they did and how many hours they gave. It recognises their work with points, tiers and digitally signed certificates they can put on LinkedIn. When a coordinator creates a task, it suggests the volunteers best suited to it and explains why. A person always makes the final choice.
A project and grant management system. It holds every programme in one place: activities, dependencies, budgets, spending, deliverables and evidence from the field. When one activity slips, it shows what else slips with it. It produces the donor report as a document, with an approval chain, from data the team has already entered.
An analytics pipeline and an analytics agent. Field surveys flow into a data warehouse, are cleaned and checked, and appear on dashboards. When one programme's incoming reports nearly tripled in two months, from 823 to 2,337 across 14 districts, the pipeline absorbed it without anyone adding staff. Managers can ask the agent questions about the data in plain language.
Managed infrastructure. All these solutions are cloud hosted. Somebody has to patch the servers, renew the domains, test the backups and watch for people spoofing the organisation's email. That is now us, and most of it is automated.
Why bespoke apps and not off-the-shelf software
Small and medium non-profits have specific needs. Commercial software is built for many kinds of customer, so it comes with features a small team never uses, and those features still have to be paid for and learned. Administering it is also cumbersome and takes time that a small team does not have. We found that a simple tool built around how the organisation already works does better than a large one the organisation has to bend to fit.
What it changed
Speed is the first thing. The project management system went from kickoff to live, with staff signing in through their work accounts, in three days. We could move that fast because we did not invent a new process. We reverse-engineered the spreadsheets the team already ran its projects on. The system shipped with about 350 automated tests. A year ago that would have been a three-month project, or more likely a licence for commercial software that did not fit how a grant-funded organisation works.
The second is that information stopped living in people's heads. The organisation's old project workbook had a hand-typed "achievement" figure that no longer matched the session records beneath it. Nobody had done anything wrong. Spreadsheets drift when the same number is typed in two places. In the new system that figure is calculated from the records and cannot be typed at all.
The third is privacy. The HR assistant uses a large language model, but the model never sees a real name or email address. Identity stays inside the application, and the model works on anonymised text. We designed it that way from the first day because India's data protection law demands care, and because staff should not have to wonder where their leave history goes.
What's coming
Over the next few weeks we will share what we learned, covering both what we built and the efficiencies it is bringing to the non-profit.
If you lead a non-profit and want to know whether any of this would work for you, get in touch. The first conversation is free.
Would this work for your organisation?
Tell us what you're trying to achieve. The first conversation is free.