When Nobody Can Explain the Campaign Email

Imagine you log into work on a Monday morning, all fresh and ready to take on whatever tasks you have for the week, and suddenly get a very urgent and confused call from the director of your organisation.

Turns out, the latest campaign response analysis email you sent to the donor on Friday mentioned that 82.6% of people were impacted. However, when their team analysed the surveys internally, it only showed 68.7%. None of their internal calculations backed up your 82.6% claim. That is when the donor sent an email to clarify, "Where did this number come from?"

The situation escalated further when the director, who was cc'd on the donor's email, scratched his head trying to understand where the claim came from and called you out by asking, "Yes, where did this number come from, and who approved this?" and for the first time you don't really have an answer. That is because last month, when your team was stretched thin, your organisation decided that the best and most obvious solution would be to let AI take more of the team's workload off the team's plate.

This led to creating AI agents and workflows to manage and automate daily tasks like analysing reports and campaign performance, and sending the outputs to the client without any human intervention. It worked wonders during testing, and everyone loved the results it produced, so human intervention was completely removed from the process to save time.

However, when that particular section of the report needed a number and the AI wasn't able to calculate the actual one due to a lack of access to the survey results, it fabricated a number, 82.6%, and sent it out to the donor. Because there was no governance in place and the number looked genuine, no one checked it, causing the escalation and a loss of trust from the donor.

By the way, this entire scenario isn't true. But in 2023, a sports journalism platform, Sports Illustrated, actually published articles with AI-generated headshots and bios, and an investigation revealed those articles were published under bylines of writers who didn't exist. This was done without proper oversight, and the fallout badly damaged one of the most trusted names in sports journalism and led to leadership departures.

This shows that while our scenario above wasn't true, it could very easily come true for any individual or team trying to replace their entire team with AI agents just to save costs and time.

The Two Issues

From the scenario above, we can see two problems come to light:

• Where did this number come from? This is not a real number, it was generated by AI to complete the campaign analysis.
• Who approved this? Due to a lack of governance, no human reviewed it.

While AI making a mistake in its analysis and providing the wrong information is certainly an issue, a much bigger issue is the fact that there were no safeguards in place to prevent the mistake, catch it, or make clear who was responsible for approving the output.

That is where AI adoption becomes more than just a question of what AI can do. It also becomes a question of how the work is governed, and the example above is exactly the kind of situation nonprofits can walk into when AI use isn't governed. Not just because AI can get something wrong, but because it can be wrong in a way nobody can properly explain!

The CURA Difference

Now imagine the same situation, but the AI agent you deployed is actually CURA. Here's how the story would go a little differently:

• AI would work from approved information: Instead of filling gaps in the results with information that just sounds right, the AI would work from the organisation's approved information. This would reduce the risk of unsupported claims or made-up data like the 82.6% impact figure.
• A human would review the output before it goes out: Since your agent can still make mistakes, the workflow would include a clear point for a person to check, question, and approve the output before it reaches a donor or other external audience. This step would have caught and fixed the issue before the email went to the donor.
• The AI would have the right context and access: Having information available doesn't always mean it should be used. Context such as permissions, sensitivities, or previous decisions would still stay connected to the work, so the AI would only have access to the information it needs for a specific task, increasing the safety of using it.
• The AI would listen to you: CURA agents can be told not to fabricate any information, and they will follow that instruction. You can also ask it to surface where information would be needed and add the required guardrails and instructions, keeping full control of the output yourself.

The main difference wouldn't be that CURA would somehow make your AI agent's output mistake-free, but that the human governance guardrail would have stopped the email from reaching the donor, protecting your organisation's trust and credibility.

That is ultimately what the future of AI in nonprofits will be defined by: not how much work we can automate, but how responsibly we can put AI to work. The real measure of progress is whether organisations can use AI to move faster and do more without losing the context, judgment, and accountability behind that work.

That means the goal isn't AI that works completely on its own. It's AI that can move work forward with human interventions in place to review, question, and intervene before an output becomes a decision or reaches the outside world.

For nonprofits, that distinction matters. Their work involves people, stories, funding, evidence, and decisions that carry real consequences. Speed matters, but not at the cost of human decisions and review.

The most useful AI, then, may not be the one that takes the most off a team's plate.

It may be the one that helps them do more, without taking responsibility off the table.