A framework by the Altruist League

ReadyForAI

A step-by-step method for the fast, low-risk deployment of AI in philanthropy.

Three stages · Twenty-one actions

Scroll

The Method

Adopt AI carefully, and cheaply

AI can bring real benefits to a philanthropic foundation, but it arrives with risks (privacy, hallucinations, bias) and obstacles (cost, data availability, internal buy-in). ReadyForAI is a step-by-step process that manages both while keeping costs low.

It moves in three stages: from zero-cost, zero-risk experiments, to putting your own data to work, to boosting impact across the organizations you fund. Each stage is a set of concrete actions. Start at the top and go as far as your foundation is ready to.

Stage 1

Quick wins, zero cost, zero risk

Build confidence and habits with free tools, with a human always in the loop.

Actions 1–7
Stage 2

Use your data, engage with regulation

Turn your own data into an asset, and get ahead of the rules.

Actions 8–14
Stage 3

Boost impact, empower partners

Extend AI beyond your own walls, to grantees, coalitions, and the field.

Actions 15–21
Stage 1

Quick wins, zero cost, zero risk

Quick wins · Zero cost · Zero risk
01 Talk to your team about AI use Address job-security fears openly, and find your early enthusiasts. +
People worried about their jobs often raise valid concerns (privacy, bias, hallucinations) as a proxy. Be clear about what you plan to do and why, so as to reassure them. Then identify the people who are curious, willing to learn, and ready to explore intelligently.
02 Experiment without spending money Use the free tiers of the leading assistants before committing to anything. +
Resist both the hype and the fear of missing out. Do not jump into a big project, not even a "pilot." Sign up for the free versions of the main assistants (ChatGPT, Claude, Gemini, and others) and low-cost image and video generators. The economics have only improved: model prices have fallen sharply, so capable tools are now close to free. Follow a newsletter such as One Useful Thing to keep track of what the models can do without getting overwhelmed.
03 Put a human in the loop Your first AI policy can be one sentence long. +
Generative AI can produce output that is incomplete, biased, or wrong. A machine on its own is fast but unreliable. A skilled human, aided by the machine and aware of its limits, is powerful. Your first policy can read: "AI output is reviewed by humans before it is used."
04 Use AI to summarize non-critical content Stop doing by hand what the machine now does in seconds. +
Modern LLMs can digest book-length documents and produce summaries. The same goes for generating reports from raw material, for example turning a collection of cryptic meeting minutes into a road map with actions and timelines. Your team should stop doing these jobs manually.
05 Let the team learn basic prompting Give the model a role, context, and examples. Then practice. +
Everyone working with these tools should know how to give the model a role, a context, and examples, and how to upload documents to it. Skip the search for magic bullets and "perfect prompt" cheat sheets. The skill comes from spending time with the models and learning to improve their output through clear instructions and iteration.
06 Use LLMs to draft and update key documents Speed up the feedback loop on your core documents. +
Examples include your mission statement, strategy, list of key performance indicators, data governance rules, and AI use policy. AI can speed up the feedback process: feed people's comments (text or audio) into the machine so that the documents improve quickly.
07 Secure management support Lead with results, not theory. +
Be ready to address the primary concerns: privacy and data protection, bias, hallucinations. Focus on what AI does, not what it is; treat it as a tool. You are more likely to win support when you already have success to show, and when you can point to what peers and partners are doing.
Stage 2

Use your data, engage with regulation

Use your data · Engage with regulation
08 Have a sensible data policy Classify data as public or not, and mind the grey area. +
As a start, classify your data as either public or private. Grant data is a good example: some foundations publish it, some do not. Mind the grey area, which is the information you provide when you prompt a model. In theory providers do not train on it, or you can opt out, but you may not want to take the risk. Truly private data should not be fed into an internet-connected model or published on your website.
09 Start a data-gathering culture The more (well-governed) data you hold, the more AI can do. +
Capture minutes from meetings and calls, and record conversations with partners, with their consent (for example with a tool like Fireflies). Gather data on grants and their performance, needs assessments, scenario analyses, sustainability metrics, potential partners, and all the feedback you receive. This should always follow your data policy.
10 Clean up your data Most of the work in AI is preparing the data. +
A common saying in AI is that 90% of the job is just cleaning and structuring the data. The most complex cases are for specialists, but the code-execution tools built into the assistants can help with simpler jobs, for example standardizing how your different spreadsheets or databases look. They also excel at quick and effective data visualization.
11 Explore open-weight models Self-hosted models now rival the best proprietary ones. +
The most secure way to use these models is to run an open-weight one (such as Llama, DeepSeek, Qwen, or Mistral) inside a trusted environment, not connected to the internet. Open-weight performance now rivals the best proprietary models, though you may need a few days of professional help to deploy them. An alternative is a turn-key solution from a local supplier, so your data does not travel to other jurisdictions.
12 Have a model learn your data Query your own knowledge base in plain language. +
You will not spend hundreds of millions of dollars training your own model, but you can take an existing one and add your data to its knowledge base. This is called Retrieval-Augmented Generation (RAG), and it is fairly standard by now, often not requiring much technical skill. It lets you query your own data and learn from it, though recall is imperfect and the model can still hallucinate.
13 Score applications with AI The model learns your methodology as the human corrects it. +
If the model has digested your methodology and some applications you have already scored, with explanations or annotations, it can score and sort new ones for you. It will not work like magic at first, but it improves as the human in the loop corrects it. This is of real value to people reviewing grant applications. Keep in mind, though, that machines can be biased against certain groups of people or causes.
14 Understand and shape regulation The rules are now in force. Know which parts apply to you. +
The EU AI Act entered into force in 2024. Its bans on certain practices and its rules for general-purpose models already apply; the high-risk obligations were pushed back under the 2026 simplification package (the "Digital Omnibus") and now phase in through 2027 and 2028. In the United States, NIST published its AI Risk Management Framework in 2023 and a companion Generative AI Profile in 2024. Tools such as the EU AI Act compliance checker help you see where you stand, and philanthropy has a role in shaping the debate.
Stage 3

Boost impact, empower partners

Boost impact · Empower partners

AI agents can now carry out multi-step tasks (research, drafting, data work) with limited supervision. That widens what a small, capable team can accomplish at this stage.

15 Build relationships with suppliers Vendors want a credible non-profit story. Use that. +
Competition among AI suppliers is fierce, and they would love to profile themselves as ethical and supportive of the non-profit sector. The major providers run programs aimed at non-profits. Partnering with them can unlock expertise and resources.
16 Help the organizations you fund use AI Pass the capability down to your grantees. +
Explore jointly how AI can improve their operations. Use your environment-scanning capability to help them find new and better donors. Show them how to write stronger grant applications, personalize outreach with custom video, audio, and other content, build the kind of online presence that lets donors find them, and use tools that predict churn and future donations.
17 Share data to build coalitions Coordination is the hard part. Standard formats make it possible. +
The most difficult aspect of philanthropy is often coordination; it is easy, but far less impactful, to focus on our own operations in isolation. To be effective, we have to understand the actions of other groups (government, business, civil society, academia) and work with them. A start is standardizing the data streams we share (CSV, XML, JSON) and our publishing formats (IATI, 360Giving).
18 Expand your dataset of potential partners Build a unique pipeline, not a bought list. +
You need a systematic process for finding the top investments. You can buy data to start, but you want a dataset that is your own: analysts embedded in local communities who find unique opportunities, recommendations from existing partners, and information gathered from the web, including with AI tools. The mix of techniques will be slightly different for each organization.
19 Track KPIs directly Monitor results yourself instead of trusting glossy reports. +
AI-written annual reports will all look the same and report on output metrics, so telling organizations apart is hard. It should be the donor's job to monitor results directly. The KPIs you measure should be concrete: growth of support, mentions in Tier 1 media, legal cases influenced, sentiment change in the local community, and so on. What used to be difficult and expensive to track can now be monitored by one AI-savvy analyst across hundreds of organizations at very low cost.
20 Build competence in-house Train domain experts in AI, not the other way round. +
It is tempting to hire PhDs to do all the AI work, but without deep knowledge of your field they can struggle: they form hypotheses that are too obvious and build models overfitted to historical data. Your instinct should be to train motivated internal people who have the domain knowledge and build their AI competence. More broadly, a baseline of AI skill should be expected of everyone, just like writing an email or using a word processor.
21 Structure your AI team as an assembly line A repeatable process beats one clever strategy. +
For predictive modelling, you want a team of specialists, following best practice from for-profit asset management: some people clean the data, others generate hypotheses, others test them through simulation. The team's job is not to find the single strategy that maximizes your KPIs, because that strategy decays quickly as conditions change. You want a process that generates and tests many strategies fast.