June 9, 2026
I keep sitting in meetings where someone says "it uses 40 agents" like that's the most impressive thing they could tell me.
And every time, my brain goes to one place: does anyone in this room know what any of them are actually doing?
Not theoretically. I mean: can you name them? Can you explain, in plain language, what decision each one is making, what data it's touching, and who is accountable when one of them does something unexpected?
Usually the answer is some version of no.
That's not a technology problem. It's a governance problem that got dressed up as an innovation story.
The research confirms what I keep seeing
A recent Omdia report found that 56% of organizations are deploying AI agents faster than they can establish proper governance frameworks, and that infrastructure, not model performance, is the real constraint for organizations trying to scale.
That tracks. In my experience, the organizations moving fast on agentic AI aren't doing so because they have everything figured out. They are moving fast because the speed feels like progress. And because nobody has slowed them down yet to ask the uncomfortable questions.
Capability is not the problem. Accountability is.
Agentic AI is genuinely interesting. The ability to chain AI models together to complete multi-step tasks, make decisions, and take actions without constant human input, there are real use cases here, including in the nonprofit sector.
But the conversation that needs to happen before any of that isn't about capability. It's about accountability. And those aren't the same conversation.
Before your organization deploys AI agents at any scale, someone needs to be able to answer three questions:
Who can explain what each agent does? Not at a conceptual level. In plain language. What decision is it making? What is it authorized to do? What happens when it gets it wrong?
Who owns it when something goes wrong? Not "the vendor." Not "the platform." A specific person, with a specific role, who has the authority to pause it, roll it back, or escalate. If you cannot name that person, you don't have ownership. You have hope.
What data is it touching? This is the question that gets skipped most often. AI agents do not just process information. They access it, retain it, pass it between systems, and sometimes send it places nobody anticipated. In a multi-agent workflow, tracing exactly what data moved where is not straightforward. If nobody has mapped it, that's not a missing detail. That's a risk.
Why this matters more for nonprofits
Your clients, donors, and community members trust your organization with sensitive information. That trust isn't abstract. It is the foundation of your ability to do the work.
When an AI agent makes a decision that affects service delivery, communications, or case management, and nobody can explain how that decision was made, that trust is at risk. "The AI did it" is not an explanation your board, your funders, or the people you serve are going to accept.
This is not an argument against using AI. It's an argument for knowing what you have agreed to before you deploy it.
What to do before you go anywhere near 40 agents
Start smaller and slower than the demos suggest. Before you evaluate any agentic AI solution, make sure you can answer those three questions for every agent in the workflow. If the vendor can't answer them for you, that's not a gap to fill in later. It's a red flag.
And if your organization is still getting clear on the basics, what AI tools you are already using, what data they touch, what rules your team is operating under, start there first. The Nonprofit AI Readiness Toolkit walks you through that process before you take on anything more complex.
Download the free toolkit here
Lisa Montague is Partner and CEO at Coat Rack, a nonprofit technology consulting firm based out of Cedar City, Utah.
