AI is already in your foundation – Is it serving your mission?

Nearly two-thirds of foundations and non-profits report using AI, across tasks ranging from internal administration and communications to research and programme delivery. Yet many organisations remain unsure what good use looks like in practice. Widespread use is not the same as better decisions or stronger mission delivery.

In a recent survey, only about one-third of professionals working in social impact sectors believed their organisation was using AI very effectively. Another survey found that among 1,321 non-profit respondents, 76% report their organisation did not have a formal AI strategy. The pattern is consistent: many are adopting AI, but far fewer organisations have decided where it should create value or built the structures needed to capture it.

Impactful AI use cases are nevertheless emerging. A 2026 review mapped over a hundred non-profits already using AI in programme delivery, across personalised support, service coordination, knowledge discovery, screening, career development and client matching. AI is being used to anticipate crop vulnerability before humanitarian crises intensify, connect families to food assistance, flag potential safeguarding concerns in calls with older people, diagnose diseases with high accuracy and conduct complex biomedical research. These applications differ substantially in their maturity, but they share a pattern: AI can change how social impact organisations convert information to action.

The Bertelsmann Stiftung Data Science Lab has been tackling the question of how the use of data and AI can help our foundation achieve its mission since its inception six years ago. This is what we have learned in the process.

1. Start with a problem

The question I hear most often is: What is a good use of AI? It is the right question, but there is unfortunately no universal answer. A promising AI use case usually combines three features: a recurring bottleneck, relevant data the foundation can access and an outcome that matters. The specific opportunity will depend on the foundation’s work, data availability, risk profile and resources.

At the Bertelsmann Stiftung, one of our first use cases was knowledge management. We have more than 3,000 studies and reports comprising about 150,000 pages. Every year, we publish an additional 15,000 to 20,000 pages – the equivalent of roughly three months of full-time reading. The bottleneck was not producing knowledge. It was making sure that organisational knowledge could be used to understand a topic, test ideas or develop new projects. We therefore developed the PublikationsBot, which enables colleagues to ask questions across the publication corpus, retrieve relevant knowledge and follow the underlying sources.

Before approving a pilot, leaders should be able to state which problem it addresses, which information it may and should use, how success will be measured, and where human review is mandatory. Start with bounded, low-risk applications and learn from actual use.

2. Evaluate continuously

Pure adoption numbers do not prove that AI is delivering value. Generative AI in particular can produce convincing outputs which can contain mistakes and biases. An organisation’s use of AI should be evaluated against the specific goals it is meant to achieve.

For our PublikationsBot, that means testing whether it retrieves the right evidence, grounds its answer in that evidence, and helps a subject expert to improve the existing work. We have been developing a repeatable evaluation process, centred on human expert judgments, that can be flexibly reused to improve the tool. In an assessment of an earlier version, senior experts rated 86% of responses as helpful. The weaker responses showed us where the system needed to improve – whether finding the right sources, interpreting them or presenting the answer.

When using AI, leaders should require representative test sets fit for the use case, human expert reviews, repeatable measurements that allow tracking the system as models and functions change and explicit criteria for restricting or stopping a use case. Evaluation is not an audit at the end. It is how the system – and the organisation – improves.

3. Build internal expertise

AI solutions are becoming easier to buy. But using them effectively is another matter. Vendors can provide technology and implementation support; they cannot define the mission problem, decide which data are appropriate, set acceptable error thresholds, or remain accountable for the consequences. Those responsibilities must stay inside the foundation.

At the Bertelsmann Stiftung, our Data Science Lab brings together data science, data stewardship, AI and governance expertise. We work closely with Corporate IT and the AI and Digital Organisational Development team to establish an operating model for AI use, in close collaboration with senior management and all staff.

Leaders should ensure that the organisation has sufficient capacity to set requirements, assess AI performance, challenge vendors and retain what it learns. Smaller organisations can combine roles, pool expertise with peers or rely on shared sector infrastructure. The model can vary; the ownership cannot. Peer learning can accelerate this capacity building. Philea’s Data Science and AI Group gives foundations a place to exchange experiences and develop stronger practices together.

An organisation’s success in using AI is not measured only by how many of its people use it. Success should instead be judged by which decisions, services or workflows have improved; what evidence demonstrates improvement; who helped define success; who bears the risk when a system fails; and what would justify stopping its use. AI is already here; the task now is to shape its use deliberately in service of our missions.

In the companion piece to this article, Catherine Miller examines AI from an outward-facing perspective, focusing on its societal risks and on how foundations can respond through their grantmaking. This article takes a different approach; illustrating how AI also offers significant potential for mission-driven organisations. The question is how foundations can realise that potential while managing their limitations and risks responsibly.

The author used ChatGPT for drafting and editorial support. The author reviewed and verified the final text and is responsible for its content.