TL;DR:
- A 1979 IBM memo warned: computers can’t be held responsible, so people must stay in charge. Truer than ever.
- There’s a real gap between watching AI work and being in charge of it.
- People tend to rubber-stamp AI’s output instead of checking it (automation bias).
- AI has shifted from waiting for your instructions to running tasks on its own so there are fewer natural points where a human reviews the work.
- 92% of nonprofits use AI, but almost none have a real plan. That’s the danger.
- Virtuous Momentum drafts in your voice but never sends, so a person stays the author.
One month ago, I stood on a stage in Dallas to close out day two of our Respond conference, the one week each year when our customers come together to connect, to explore, to learn, and to dream about what comes next.
My keynote was called “Human First + AI Forward.” It is the phrase we use at Virtuous to describe a single conviction: that people come before the technology, and that we can move boldly into AI without moving away from our humanity.
Near the end, I put a single slide on the screen. It was not a chart or a product roadmap. It was a photograph of an old IBM training slide, written, as the story goes, for an internal meeting in 1979.
It read: “A computer can never be held accountable. Therefore, a computer must never make a management decision.”
The room got quiet…
A sentence drafted before most of the people in that audience owned a calculator was describing our present moment more precisely than almost anything I had read all year.
It was true when it was written. It is truer today. The only thing that has changed is how easy we have made it to ignore.
We are living through the most aggressive period of delegation in the history of work. We are handing tasks to systems that can draft, decide, summarize, score, and increasingly act on our behalf. And we are doing it faster than we are building the structures to stay responsible for what those systems do. The memo did not age. We did.
From Human in the Loop to Human at the Helm
For the last few years, the phrase we have leaned on to describe responsible AI is “human in the loop.”
It is a good phrase. It captures a basic and worthy commitment: that a person should sit somewhere in the process, reviewing what the machine produces before it reaches the world. Almost all of our governance language, our policies, and our compliance checklists are built around it.
Not too long ago, I was in a Fundraising.AI meeting with Ben Miller, who is one of the people I respect most when it comes to keeping the human at the center of philanthropy. He is a friend and someone whose instincts on this question I trust deeply. In the middle of the conversation, almost in passing, he used a phrase that stopped me.
He said the goal is not to keep a human in the loop. The goal is to keep a human at the helm.
I had heard the phrase before. It has been circulating in the AI conversation for a couple of years now. But coming from Ben, in a year when AI is learning to act on its own, it landed with a weight it had never carried for me before.
The Difference Between a Position and a Responsibility
The difference is small in language and enormous in practice.
Being in the loop is a position. Being at the helm is a responsibility.
One describes where a person is standing. The other describes who is steering. You can be in the loop and still be a formality, a required signature on a decision that was effectively made before it ever reached you.
You cannot be at the helm and be a formality. The helm is the place where accountability actually lives.
Why Approval Is Often Just a Reflex
Here is the uncomfortable truth about being in the loop. Most of the time, when a human approves an AI output, the approval is not oversight. It is a reflex.
Sol Rashidi, who has led more enterprise AI deployments than almost anyone I know of, calls this automation bias, and it is one of the most documented and least discussed risks in AI today.
When a person is asked to review a machine’s work under time pressure, the easiest thing in the world is to agree. The output looks fluent. The clock is running, and the next task is already waiting. So we click approve, and we tell ourselves we reviewed it.
Picture a gift officer at a busy nonprofit. An AI system has drafted forty personalized notes to major donors and scored each relationship for its likelihood to give. The officer has twenty minutes before the next meeting. The drafts are good. The scores are plausible. What is the incentive to slow down and interrogate the model’s reasoning on donor number twenty-three? The system has made saying yes effortless and saying no expensive. And more so than a character flaw in the officer, it is a design flaw in the process.
The Three Conditions for Real Oversight
For a human review to mean anything, three conditions have to hold:
- The person has to understand what they are looking at well enough to catch an error.
- They have to hold the context the model does not have: the history with that donor, the reason the easy answer is the wrong one.
- And they have to be genuinely free to say no, without the quiet institutional pressure to simply keep things moving.
Most organizations have none of those three. They have a signature step, and a signature step is not oversight. It is automation bias with a human name attached to it.
You Can Outsource Thinking, but Not Understanding
Andrej Karpathy has been repeating a line lately that captures the whole problem. You can outsource your thinking, he says, but you cannot outsource your understanding.
It is worth knowing who is saying this. Karpathy was a founding member of OpenAI and later the director of artificial intelligence at Tesla, where he led the team building the vision system behind the company’s self-driving cars. He has spent his career at the frontier of autonomous machines. And the warning he keeps returning to, made squarely about this new era of AI agents, is that as the systems take over more of the doing, the human’s irreplaceable job becomes understanding: knowing what is worth building, which result looks suspicious, which tradeoff is acceptable.
A signature outsources the thinking.
The helm requires the understanding.
When the person at the helm no longer understands what they are approving, the helm is empty, no matter who is sitting in the chair.
2026: The Year AI Stops Being a Tool and Becomes a System
This is not a hypothetical worry, and it is about to get sharper. At the turn of this year, I published a set of predictions for the sector, and the central one was this: 2026 is the year agentic AI stops behaving like a tool you prompt and starts behaving like a system you delegate to.
The distinction matters more than it first appears. A tool you prompt waits for you. You ask, it answers, and you decide what happens next. You are always in the loop because nothing moves without you. A system you delegate to does not wait. It pursues a goal, plans, acts, observes the outcome, and adjusts across many steps, often without checking back in. Much of the private sector crossed this line in 2025. The nonprofit sector is crossing it now.
When You Delegate, the Loop Disappears
When you delegate, the loop disappears. What remains is the helm.
This is the part the excitement tends to skip. Agentic AI is genuinely thrilling. It compresses timelines and lets a small team operate with the leverage of a large one. I am not here to talk anyone out of it; I have spent years arguing that the sector cannot afford to wait.
But the more we delegate, the fewer natural moments there are when a human pauses to review, and the more accountability has to be designed in on purpose rather than assumed. In this way, autonomy actually raises the need for human responsibility.
The Org Chart Is Inverting
Consider what this does to the shape of an organization. For roughly a century, the logic of an org chart was simple: you identified the jobs that needed doing and hired people to do them. The human was the worker. That logic is now inverting. Increasingly, AI can audit what work needs doing, perform a growing share of it, and leave people to direct and oversee the machine. The human is becoming the manager of the AI rather than the doer of the task.
When that inversion is complete, the human’s job is no longer to do the work. It is to remain accountable for it.
That is the helm, and it is harder than it sounds, because there is a failure mode that looks almost identical to it. Picture the worker who once built a product by hand and now stands beside a production line, watching a thousand units race past, pulling the occasional defect. That worker is in the loop. They are present, even useful. But they are not in charge of anything. They cannot change where the line is headed, or whether it should be running at all. Oversight reduced to that is not oversight. It is the appearance of a human, kept in frame for reassurance.
The helm is the opposite. It is the place where someone can still say stop, still change course, and still be answerable for the destination.
The One Asset We Cannot Replace
For the nonprofit sector, none of this is an abstract management puzzle. It is a question about the one asset we cannot replace.
As models grow more capable and more woven into everything we do, the line between what a human created and what a machine generated is going to blur, and then it is going to vanish. A thank-you note, a story of impact, a personal appeal: before long, you will not be able to tell, from the artifact alone, whether a person or a system produced it.
That is not, by itself, a catastrophe. But it raises a question the sector has to answer honestly. When the gesture of connection is generated rather than given, is it still generosity? Or is it a counterfeit of generosity, convincing on the surface and hollow underneath?
How Virtuous Momentum Keeps the Human as Author
This is the question that shaped one of our products at Virtuous called Virtuous Momentum. It uses AI to draft donor outreach in a gift officer’s own voice. What it deliberately does not do is send that draft to the donor. It hands the draft back to the officer, to read, to edit, to make true, and to put their name behind it before anything goes out. The AI does the drafting. The person keeps the deciding, and the accountability that comes with it. The connection stays genuine because a human still authors it.
Try out Momentum in a demo now.
Trust Can Only Be Extended to What Can Be Held Accountable
Trust is the foundation of the nonprofit sector.
It is the reason a stranger gives to a cause whose results they will never personally see. And trust has a peculiar property that sits at the center of everything I am describing: it can only be extended to something that can be held accountable. A person can be. A model cannot. You can trust a person to keep a promise because, if they break it, there is someone to answer for the breaking. There is no one to hold accountable inside a model. That is the entire argument of a fifty-year-old IBM memo, and it is the entire argument for the helm.
The Adoption–Strategy Gap
There is a gap in our sector that makes all of this both harder and more urgent, and it is also, I believe, the opportunity. AI adoption among nonprofits is now remarkably high, around 92 percent. But the number of organizations with an actual AI strategy, one that embeds these tools into how the organization makes its real decisions and names who remains accountable for them, is still in the single digits.
That gap is where the danger lives, because adoption without strategy is exactly how you end up with automation bias wearing a human signature. But it is also where the work is, and the work does not start with a tool. It starts with working backward from a few honest questions:
What are we actually trying to achieve?
How do we make it safe for our teams to experiment and to learn, to say “I do not understand this output” without being treated as behind?
And where, deliberately, will we keep a human at the helm, not because the machine could not perform the task, but because we have decided this is a place where a person must remain answerable?
Human First + AI Forward Is an Order of Operations
“Human First + AI Forward” is not a tagline we invented for a stage in Dallas. It is an order of operations. Human first. Then AI, moving forward with intent. It is also a form of restraint, a willingness to slow down precisely where slowing down protects something worth protecting. The sector’s caution about new technology has often been read as a weakness, a sign that we are behind. I have come to believe it is closer to wisdom. In a moment when much of the world is racing to hand its judgment to machines, the instinct to keep a human answerable may turn out to be the most valuable instinct we have.
The machines will keep getting faster. That was never the question. The question is whether we will stay at the helm while they do.
A computer can never be held accountable. Half a century later, that is still true. It is also still our job.
FAQs
What does “human at the helm” mean?
It means a person remains genuinely accountable for and able to redirect what an AI system does, not just positioned somewhere in the process as a required signature. The helm is where someone can still say stop, change course, and answer for the outcome.
How is that different from “human in the loop”?
Being in the loop describes where a person stands; being at the helm describes who is steering. You can be in the loop and still be a formality, but you cannot be at the helm and be a formality.
What is automation bias?
Automation bias is the tendency to approve a machine’s output without real scrutiny, especially under time pressure when the output looks fluent and the next task is waiting. It turns human review into a reflex rather than genuine oversight.
Why does agentic AI raise the stakes for nonprofits in 2026?
A tool you prompt waits for you, so you are always in the loop. A system you delegate to plans and acts across many steps without checking back, which removes the natural review moments and forces accountability to be designed in deliberately.
What is the nonprofit AI adoption-strategy gap?
Roughly 92% of nonprofits use AI, but only single digits have a real strategy that embeds the tools into decisions and names who stays accountable. That gap is how organizations end up with automation bias wearing a human signature.
How does Virtuous Momentum keep a human as the author?
Momentum uses AI to draft donor outreach in a gift officer’s own voice, but it never sends the draft. It hands the draft back to the officer to edit, make true, and put their name behind before anything goes out.
