Nonprofits are under growing pressure to adopt AI, but the path is full of pitfalls. In a recent exchange with Peekwire, Eshara Kohli, co-founder of ImpactFlow AI, shared a candid look at how mission-driven organizations can implement automation responsibly. The conversation covered everything from data privacy to the dangers of free AI tools, offering a practical roadmap for teams that want to scale impact without losing the human element.
Defining Ethical AI in Practice
Kohli is quick to cut through the buzzword. For her, ethical AI comes down to three pillars: guardrails, human supervision, and human-in-the-loop checkpoints. "A guardrail is a boundary the system cannot cross, agreed on before anything is built." In practice, that means asking four questions before any automation is deployed: what data can the AI access, what is it allowed to do on its own, which outputs require approval, and what happens if something goes wrong.
She emphasizes that AI can outperform tired humans in consistency, but it can't be held accountable. "Human supervision is still needed. That's because AI can’t be held accountable for its mistakes." Every system gets a designated human owner who reviews outputs and can pause the system. The distinction between supervision and human-in-the-loop is key: supervision is ongoing, while human-in-the-loop is a checkpoint built into the workflow. "The AI prepares, the human decides."
Data privacy is non-negotiable. Kohli warns against AI that trains on your data and advocates for zero data retention. "Using AI that trains on your data is an absolute no-go." ImpactFlow structures automations to comply with privacy rules even when clients can't afford enterprise subscriptions, and they avoid vendor lock-in by making provider switches a simple config change.
Common Mistakes Nonprofits Make
The biggest error? Treating AI as a productivity booster for staff rather than a system that works for them. "The difference between working with AI and having AI work for you is significant." The real value emerges when AI handles tasks in the background while humans focus on mission-critical work.
Another frequent misstep is relying on free AI chat tools. "Most managers and staff don't realise their data is used to train these models." Teams often paste sensitive information into free tools without reading the terms. "It's simply not safe to feed your organisation's data into free AI tools." Kohli notes this isn't about unwillingness—teams are overwhelmed and overworked. To address this, ImpactFlow includes a free AI literacy session with every build, which she says "massively increases the chance of a quick win."
Balancing Automation with Human Touch
Kohli's rule of thumb: automate the repetitive busy work, but never send communications without human review. She cites one exception—requesting quotes from vendors, where emails are simple asks with low stakes. Even then, the system drafts winner and decline emails, but a human reviews and sends them. "Whichever part of the system drafts the winner and decline e-mails doesn't have the ability to send them."
Training is essential to avoid what she calls "AI slop." "Staff need to know how to recognise when material feels AI-generated." The quality of outcomes depends on the people using the tool, so ImpactFlow insists that all team members join training sessions.
Best Tasks for AI—and What Stays Human
Kohli stresses that processes must be clear before automation begins. She recalls a client who asked for a simple automation but actually needed a full operating system. "The underlying issue was chaos." Instead of jumping in, they mapped processes, access levels, and ownership, building a system that now flags errors automatically.
Some tasks should remain human-led: capacity building, conflict resolution, and field missions. Donor emails to mid-level donors and above can be drafted by AI—pulling from CRM data on campaigns, events, and giving history—but must always be reviewed. "The cost of sending something that feels AI-generated is too high." AI can support human-led activities by making projects more sustainable, creating refreshers, or tracking progress.
Starting Small with Limited Budgets
For smaller nonprofits, Kohli recommends starting with one impactful automation that saves time or increases income. "The time or money saved can then be used to create impact elsewhere." She advises choosing an agency that understands ethical structuring and to "just start small. But start." Even simple invoice processing can free up hours weekly and reduce errors.
As AI becomes more accessible, the challenge for nonprofits is not whether to adopt it, but how to do so responsibly. Kohli's advice offers a clear starting point: define guardrails, keep humans in the loop, and prioritize data privacy. For more insights, explore ethical AI platforms for nonprofits or learn about AI ethics guidelines from UNESCO. Organizations can also consult data privacy best practices and AI literacy resources to build internal capacity.