“Human in the loop” was the right phrase for the first wave of AI adoption. It isn’t strong enough for the wave we’re in now. As AI agents move from pilots into critical technologies , the question boards are asking isn’t “did a person review this?” — it’s “who is accountable for what this produces?” We call that standard expert in command (Eric Dingfelder, President & CEO of Data Innovations, from Fireside Chat) and it’s becoming the difference between AI that’s engineered and AI that’s just enabled.
By 2028, Gartner projects the typical Fortune 500 enterprise will be running more than 150,000 AI agents — yet only 13% of organizations believe they have governance frameworks adequate to manage them [1]. Separately, SAP’s 2026 Agentic AI survey found that 98% of companies have already deployed AI agents or plan to, while less than half have real visibility into an inventory of what those agents are doing [2]. That gap — massive deployment, thin governance — is exactly the problem “expert in command” is designed to close.
1. What does “expert in command” mean, and how is it different from “human in the loop”?
“Human in the loop” implies a person is present. “Expert in command” implies a person is accountable. It means a named individual has reviewed the AI-assisted output, understands it well enough to explain it, and is willing to put their name behind it as if they’d produced it themselves.
Eric Dingfelder, President and CEO of Data Innovations, put it this way in a recent fireside chat with Gorilla Logic CEO Drew Naukam: his teams don’t get to “acquiesce responsibility to some AI thing that could actually have better results.” Whatever comes off the line still needs a human signature — what he calls “their John Hancock at the bottom” — because ultimately, real people are affected by what the software produces [3].
2. Why isn’t “human in the loop” strong enough for high-stakes AI use cases?
Because presence isn’t the same as ownership. A reviewer who’s technically “in the loop” but hasn’t been given the authority, the time, or the expectation to actually stop and challenge the output isn’t providing governance — they’re providing a rubber stamp. As AI-generated work scales, that gap between watching and owning is where quality and accountability quietly erode.
3. What does “signing off” on AI-assisted work actually require from a team?
Three things: the expertise to evaluate the output, not just skim it; the authority to pull the “stop cord” and say it isn’t good enough, even under deadline pressure; and a culture where doing that is rewarded, not penalized. None of that is a policy you publish once — it’s a standard you have to engineer into how teams actually work.
4. How does engineering discipline fit into an AI governance model?
Security, governance, testing, scalability, maintainability, and proven engineering process aren’t separate from your AI strategy — they are your AI governance model, applied to a new kind of output. Anyone can spin up an agent in an afternoon. Very few of those agents are built to be secure, auditable, and maintainable a year later. That’s the gap between something that demos well and something enterprise-grade — and it’s why engineering discipline is what makes AI adoption sustainable rather than a liability waiting to surface.
5. Doesn’t more governance mean slower AI adoption?
Not in practice. The organizations struggling most right now aren’t the ones with too much rigor — they’re the ones that scaled agents faster than they built the visibility to manage them, which is exactly what Gartner’s numbers above describe. Governance done well doesn’t sit on top of your AI work as a checkpoint at the end. It’s built into the workflow, so speed and accountability move together instead of trading off against each other.
6. What role does an outside engineering partner play in keeping AI accountable at scale?
A partner who has seen this problem across dozens of organizations brings pattern-matching your internal team hasn’t had the reps to build yet — where agent sprawl typically starts, which controls actually hold up under audit, and which shortcuts look fine today and become expensive in eighteen months. That’s a different value than a consultant handing you a framework: it’s the accumulated judgment of an engineering organization, not just the person sitting in your standup.
7. How does an organization start building an “expert in command” model?
Start by naming it: for every AI-assisted workflow that touches customers, revenue, or compliance, identify who signs off and make that explicit rather than assumed. Then build the review step into the process itself — not as a bolt-on audit, but as engineered infrastructure with the same rigor you’d apply to security or testing. The goal isn’t to slow AI down. It’s to make sure that when it moves fast, someone can still stand behind what it produced.
Enterprise AI doesn’t fail because the models aren’t good enough. It fails because governance gets bolted on after the fact instead of engineered in from the start. If you’re trying to figure out what “expert in command” looks like inside your own organization, talk to our team — it’s the same conversation we have with clients building AI into regulated, high-stakes environments every day.
Sources
- Gartner, cited in “AI Agent Sprawl: Why AI Governance Is Now a Board-Level Issue,” SAP News Center, August 2026.
- SAP LeanIX, 2026 Agentic AI Survey, cited in the same SAP News Center article.
- Eric Dingfelder, President & CEO of Data Innovations, in a recorded fireside chat with Gorilla Logic CEO Drew Naukam.
