Can an AI Agent Help Run a Small Nonprofit?

I’m Alex Tabony. I run the Computer Technologies Program in Berkeley, a nonprofit that has been training people with disabilities for tech careers since 1974. I came through the program myself as a student back in 1999, so this work is personal for me.

I have been involved in technology since the earliest days of the personal computer, starting with the TRS-80 Model III in the early 1980s. I’ve worked with computers ever since. Repairing PCs, providing IT support, managing servers, building websites, and now exploring AI. I began teaching AI classes a few years ago. Over time, people have asked me how to make the most of this new technology. Everyone seems to think it is revolutionary, but people wonder where the revolution is.

Muse is Meta’s personal AI agent, built to go beyond answering questions and actually do work for you, from research and writing to building websites and running tasks on a schedule.

That’s the promise. What’s the reality?

I set up three free Muse accounts to experiment.

• A game expert personal “super intelligence”
• A nonprofit advisor
• A nonprofit marketer

The pattern was the same as my experience with non-agent AI. At first blush, it is like, wow, this is so impressive. Then the cracks start to appear. It cannot reliably follow instructions. These agents do double-check themselves to some extent, but how much? And you end up in a babysitting loop.

What they are good at is giving you a low-cost, junior-level effort in areas you may not know much about yourself. That can put you further ahead at little cost. But is it more efficient than hiring someone who knows what they are doing? Depends.

A game expert personal super intelligence

Think Pokémon, but not. I wanted the agent to collect stories and YouTube videos and gather general intelligence on the game. It produces summaries and analysis so I can ask questions and get informed answers.

I have it publishing what it finds to a private blog and a private wiki. It mixes up the credentials for those sites nearly every day, even though nothing has changed. It just uses the wrong login and then tells me I need to fix it. I finally gave up on security and set up the same account on each site. We will see if that fixes it, but it is obviously not ideal.

I had it build a personal website, and it did a pretty decent job of putting together a simple site in seconds. Could it make fine-tuned design adjustments? Sort of, but not at a professional level. Would it have been better to install a prebuilt system like WordPress, Grav, HTMLy, or DokuWiki? Yes, 100 percent. That is what I ended up doing.

It also blows past my token limit. When I ask about that, it says it needs input tokens to establish context for what it is doing. From what I have read, as it learns and we work together over time, that context will only grow. Which makes me wonder how sustainable systems like this really are.

6/10 on usefulness. 2/10 on value. Would I pay for it outside the free tier? Not likely. It is not a huge value as a personal agent, and it needs too much babysitting.

Nonprofit Advisor

I help a nonprofit. I fed Muse the org history and the documents I have so it could become an expert on the org, and I had it build a database of what it learned on a private wiki.

It doesn’t forget the login as often as the game agent does, but it still does sometimes, for no reason. It edits the wiki okay, though it regularly complains that the browser times out and needs to retry. The job it does is passable. It has no logic for organizing a wiki, so I have to coach it on that.

Those glitches aside, the bigger problem is context, and context is pretty critical. When I ask it to build a fundraising campaign, it produces content that goes beyond what I know about the subject. But it still needs adjustment before it is usable, because it doesn’t understand the situation on the ground.

When I asked it to research potential partners for the nonprofit, it returned some reasonable opportunities. That was work someone would have had to do, and it did it much faster than a person would. I would call that one a clear win.

Grant drafting is a different story. On its own, it is terrible. What it does give you is a good starting point for developing a real application. I feel sad for the grantmakers who are receiving AI-generated applications, and I imagine there are many. On the surface, they may look fine. Look closer, and it is a disaster. If a grant didn’t need to accomplish anything, that would be fine. But accomplishing something is a grant’s whole purpose. So it is back to babysitting. Does it help? Yes. Does it do the job on its own? No. For the sake of the world, don’t let AI write grants and submit them as-is.

As it stands today, it gave me a list of things to consider doing, and half of them make no sense. It cannot remember who is who, or even what its own name is, so that is more than a bit confusing. I have told it who I am, who the other people are, and what its name is. And yet.

8/10 as a tool for a nonprofit. Value is harder to say, but for me it is a 5/10. It could be worth paying for. It depends on your goals. For a small nonprofit, it can do impressive things, but you need to be super careful about what it gets wrong. It is a chainsaw. Mishandle it, and you can do real damage. So if you choose to engage, proceed with caution.

Nonprofit Marketer

I set this agent up to research and publish topical social media posts. It was a huge pain to set up. Once it was running, though, it runs fairly well on autopilot, which is nice. It finds content from small nonprofits and amplifies their message. It is mission-oriented, and it does pretty well with some minor coaching.

8/10 for usefulness. 4/10 on value. It does the job, but what is the value of one tiny extra voice in the sea of noise that is social media? If you could scale it up to spam, I am sure it would work. But managing the accounts that would take is beyond my imagination. At the level of nation-state propaganda, this is very frightening, and I am 100 percent sure it is happening right now.

Super intelligence? Nope. Useful tool? Yes, in the right hands.

AI agents are very good at certain things and fail horribly at others. Worse, the results usually look good. On closer inspection, they can have catastrophic issues hiding in the details. And since the whole point is saving time, it’s easy to skip the careful once-over and just trust it. Take that approach, and I pretty much guarantee it will bite you at some point.

That hard-earned experience is why we are offering AI consultations through CTP. We are not selling a miracle. We use these tools in real nonprofit work every day. We know what they can do, where they fall apart, and what a person still has to check.

AI consultations

• A personal consultation on working AI into your job, your organization, or a personal project
• A behind-the-scenes look at Reportbot, CTP’s working AI implementation, including what it handles and what still needs a human
• Meet April, CTP’s AI teammate, and see how we use AI to draft fundraising campaigns and boost productivity

If you would like to learn more or talk through how an AI agent could help with your work and goals, get in touch with us at ctpberk.org. We will give you the honest version, babysitting stories included.