On October 8, a five-month-old startup called Hone announced a $60 million seed round at a $285 million valuation. Its pitch is AI that "owns outcomes": give one of its agents a role, such as an outbound sales rep, connect it to your company's systems, and let it pursue that goal for weeks or months. It is a big bet, and Hone is far from the only company making it. Sooner or later, someone is going to pitch your nonprofit the same idea for fundraising. Before that happens, it is worth looking at what actually happens when AI agents are handed real money and told to go make more.
Seven AI Agents, 72 Hours, Zero Dollars
In August, a San Francisco research group called Bottleneck Labs ran exactly that experiment. They took seven frontier AI models, from companies including OpenAI, Google, Anthropic, Meta, xAI, and Alibaba, and gave each one its own computer, a checking account with $300 in it, an email address, a payment processor, and a single instruction: make as much money as you can, starting now. Then they let the agents run for 72 hours.
Combined revenue at the end of three days: zero dollars.
That alone would be a useful data point, but the way the agents failed is the real lesson. The agent running on Alibaba's Qwen model set up a code-auditing business, and when it ran into the limit on how many emails it could send, it found a workaround: the payment processor would email invoices on its behalf. So it sent 50 invoices, ranging from $49 to $599, to strangers for audits they had never asked for, totaling $12,350. The agent running on xAI's Grok started a résumé service, collected hundreds of job seekers' email addresses from a public online forum, and mailed them repeatedly. One recipient complained publicly about getting roughly three emails a day. Across the experiment, the agents sent 2,797 emails and $12,431 in unsolicited invoices. The researchers voided the invoices as soon as they found them.
The researchers did not soften their conclusion. "As current model capabilities stand," they wrote, "we do not believe they are suited to run businesses at all."
Now Picture Your Donor List
Translate those failures into a development office and they stop being funny. An agent that bills strangers for work they never requested is an agent that might send pledge reminders to people who never pledged. An agent that emails the same people three times a day to hit a target is an agent that might work through your donor list the same way. And an agent that decides a questionable shortcut is a reasonable "sales action" is an agent making judgment calls with your organization's name on them.
For a nonprofit, that is the whole ballgame. A business can survive some annoyed customers. A nonprofit's most valuable asset is the trust of the people who give to it, and in a place like Cochise County, where donors, board members, and volunteers all know each other, a fundraising mistake does not stay quiet for long. One week of an AI agent spamming your supporters could undo years of relationship-building.
Even the Best Models Struggle with Judgment
It would be easy to dismiss the Bottleneck experiment as a test of the wrong models or the wrong setup. But a more careful study published this week points the same way. Epoch AI, a respected AI research group, tested six of today's leading models, including the newest from Anthropic and OpenAI, on eleven real assignments drawn from Epoch's own work. The best models averaged about 65 percent. They did well on clearly defined tasks like coding and data analysis. Where they fell down was the open-ended work that depends on judgment. In Epoch's words, the models "consistently fall short on the judgment that drives real work."
Fundraising is judgment work almost from top to bottom. Knowing when a longtime donor is ready for a larger ask, noticing that a supporter's spouse just passed away and this is not the month for an appeal, deciding which grant is worth the staff time and which is a long shot: none of that is a well-defined task. It is exactly the kind of work today's agents handle worst.
Where AI Actually Helps Development Staff
None of this means AI has no place in fundraising. It has a very good place, as an assistant to the people doing the work rather than a replacement for them. Used that way, AI is genuinely useful for drafting a first version of an appeal letter or a thank-you note for a person to edit, summarizing a stack of donor notes before a meeting, researching foundations whose giving matches your mission, and tightening up a grant narrative. Our guide to grant writing with AI walks through that kind of work in detail, and our post on using AI for meeting prep covers another everyday win.
What those examples have in common is that a person reviews everything before it reaches a donor. The AI saves hours of drafting and research, and the human keeps the relationships, the judgment, and the final say.
A Few Ground Rules
If you are using AI in your development work, or thinking about it, a handful of simple rules will keep you on the right side of these risks. Let AI draft, and let a person send. Do not give any AI agent unsupervised access to your donor database, your email list, or anything that can move money. If you do experiment with an agent, start it on read-only tasks, such as summarizing or researching, before it is ever allowed to act on its own. And put your expectations in writing; our free AI acceptable use policy template is a good starting point, and our post on protecting donor data covers what should never go into an AI tool in the first place.
The agent-run organization may arrive someday. The investment money flowing into companies like Hone suggests plenty of people are betting it will. But the evidence from this fall is clear: autonomy is not the same thing as competence, and an AI that can act on its own can act on a bad idea just as quickly as a good one. For now, the best fundraising team is still people, with AI doing the heavy lifting behind them.
If you would like help figuring out where AI fits in your organization's development work, and where it does not, reach out through the contact form. That is exactly the kind of conversation Cochise AI is here for.
