AI Uplift as a Service - Helping Nonprofits Have a Slightly Greater Impact

Impact Ops is experimenting with AI coaching and consultancy as a service. Since joining our team, Alejo Acelas has spent two months working at 80,000 Hours and coaching staff from Ambitious Impact and its incubated charities.

Based on what we learned, we're designing new services around two ideas: helping staff expand their AI toolkit through regular experimentation, and building tools in short, collaborative sprints to address big visible pain points.

If you'd like to be one of the next organisations we work with, please let us know by filling out this form.

The rest of this post is Alejo's account of what he worked on over the past two months, ways in which his work fell short of his initial expectations, and the approaches he's currently most excited by.

A brief motivation for AI uplift

For me, the central case for AI uplift is:

There's things you can do with AI that are worth doing and that you haven't given a proper try yet^[1].

If you already know what some of those things are, I don't need to convince you of much. If not, here's some things I've observed, for your consideration:

1. People in similar roles report getting very different amounts of value from AI.
2. A few people I’ve met started getting much more out of AI soon after a single interaction with someone they trusted who was using AI more ambitiously.
3. We have some early examples of AI use translating to tangible wins at organisations.

These don't imply that you or your team could quickly start using AI much better, or that doing so would result in better outcomes for your organisation. But it's still somewhat suggestive.

Here's a couple stories of how I've tried to help organisations get more value out of AI, to help you make up your mind.

Coaching

Most of my coaching consisted of one-off 50-minute calls. There I'd ask people about their role, suggest a task that seemed very amenable to AI automation, and very often just launch into tackling it during the call so I could give live input on their approach. I also had people fill out a pre-call form, which was pretty useful to get a sense of their starting AI fluency and their priorities for the call.

Thinking back over roughly 30 calls, coaching was most useful when I could open up some new space for AI exploration that made people want to keep trying things out on their own after the call. For example, I think people got most from calls with me when they either:

1. Had not tried AI much beyond single-task chats
2. Already used AI enthusiastically, but had not encountered many of the ‘AI power user’ patterns
   1. e.g., doing more planning upfront, delegating across subagents, or allowing agents to run independently for hours by removing one-off blockers

In both cases I had a very concrete value offer: “here's something immediately useful, which you had not tried using, that you can now reuse across a variety of projects”.

On many other calls I suspect the most valuable thing I did was suggest an easy-to-adopt tool for a problem people submitted to me. I spend more time than many of my clients trying out new AI tools, so I can often suggest something useful they had not considered.

I also have an intuition that maybe, through sheer force of personal excitement, I managed to prompt some people to play around and experiment more with AI after my coaching, but that's harder to tell.

Overall, I suspect most of my impact came from encouraging people to form a more frequent habit of experimenting with new tools and ways of using AI. And this is despite me doing little to follow up with them after the call. I already have some ideas for how to improve on this, but you’ll have to wait a bit to see them.

Bespoke projects

In addition to coaching, I also built a few tools and automations for staff at 80,000 Hours. In brief, I discovered that automating other people's work is much harder than automating your own and I didn't do a particularly good job of reorienting in light of that.

When I started working with 80k, my mindset was “you don't need to prioritise, as long as you're fast enough at building solutions.” Turns out I wasn't fast enough at building solutions.

Here's some ways in which I found building for others hard:

1. A few times I didn't spell out the desired outcome of a project upfront, and I wasted time flat out building the wrong thing. And even after doing things better, most tools needed repeated feedback from users, which introduced some coordination costs.
   1. Sometimes I could only automate part of a task, which means I also needed to work with users to design an interface that made it easy to complete the remaining steps.
2. Every moment someone spends helping you automate their task is a moment they can't spend doing it. And even if the automation would save them time eventually, there still needed to be a case that automating it should be a priority now.
3. I was no longer the owner of all the systems I wanted to rely on, which meant I had to learn to ask for access and permissions, and figure out reasonable scopes for them based on what the project required.

I'm still unsure which of these complications would be solved by better prioritisation and more practice, versus which are structural to AI uplift efforts. Still, I was expecting to have more tangible wins after two months at 80k, so I take this as some evidence that AI uplift is harder than I initially thought.

I think building solutions for your own problems is so much easier. So I'm very keen to test ways of building capacity within organisations so they can automate things for themselves. Hopefully I'll have more to say on this soon!

Work with us

If you’d like to explore how we could help your team get more out of AI, just let us know!

And, if you think you would make a good AI coach, feel free to leave us your details here. We can't promise anything, but if the experimental period goes well we may scale up this service.

1. ^{^}

    Much hides in the word ‘proper’ here. I would say more if only I had sorted out my thoughts on this.