DEV Community

Cover image for Human-in-the-loop isn't collaboration. It's a queue.
Khoushik Reddy Ananth
Khoushik Reddy Ananth

Posted on Originally published at linkedin.com AI-assisted

Human-in-the-loop isn't collaboration. It's a queue.

We spent three years asking how much AI to let in. We never asked who decides, and who checks.

Every enterprise I talk to this year has the same two lists.

The first is the list of AI agents they've switched on. It's long and it grows every quarter. Some of the agents they built themselves. More of them came inside software they already pay for: the CRM, the service desk, the HR system.

The second list is the people who are supposed to be "in the loop" on those agents. It's short and it doesn't grow at all.

The gap between those two lists is the real story of AI at work in 2026. It isn't a gap in model quality. It's a gap in how people and AI actually work together.

How we got here

Looking back, we've tried three ways of putting AI to work. Each one broke in the same place.

First, AI as a tool. A person prompts and the AI answers. It's useful, but the person carries all the judgement and does all of the checking, and every answer starts from zero. Nothing the company learns builds up anywhere.

Second, AI as an autonomous agent. Give it a goal and let it run. This works brilliantly in demos and makes people nervous in production. When it's wrong, nobody can say who owned the decision or why it was made.

Third, the fix we bolted on: human-in-the-loop. Put a person between the agent and anything that matters, and make them click approve.

That third one is where most companies are now. The numbers show it isn't holding.

What the 2026 numbers say

  • Dynatrace surveyed 919 senior leaders. 69% of agentic AI decisions are still verified by humans, and only 13% of organisations use fully autonomous agents.
  • EY surveyed 202 senior AI executives. Among companies using agentic AI, 85% already have systems acting without real-time human involvement. 26% can't detect unauthorised agents running inside the company. 47% have bypassed their own AI governance to ship something urgent.
  • VentureBeat surveyed 157 enterprises. Half have shipped an agent that passed its internal evaluations and still failed a customer. Only 5% fully trust the automated evaluations behind their release decisions.

Read those together. We're asking people to verify most decisions, while most companies already have agents acting with no one watching in real time. We're checking everything and seeing very little.

That's what happens when the human's only job is to approve. The queue grows faster than the people. Approvals turn into rubber stamps. Then someone finds a way around the queue, because the business can't wait.

The gap, named

The problem isn't that AI is too weak, or that people are too slow.

We gave AI work without giving it a job.

When you hire a person, you never say "do things, and someone will approve you." You write down what the role is responsible for, what it can decide alone, what it hands off, when it escalates and to whom, what it must never touch, and who checks its work.

We skipped all of that for AI. Then we were surprised that the only control left was a person pressing approve.

There's one more reason approval alone can't work. The Harvard and BCG study of consultants using AI (Dell'Acqua et al., 2023) found what it called a "jagged frontier." On tasks inside AI's frontier, people with AI did much better. On a task just outside it, they did worse than people without AI, and they couldn't tell which side of the line they were on.

If people can't reliably see where AI's edge is, a person clicking approve at the end isn't a safety net. It's a hope.

What hybrid intelligence actually means

Hybrid here means AI and people. It doesn't mean office and home, and it doesn't mean hybrid cloud.

A Hybrid Intelligent Enterprise is a company where AI agents and people work as one team, in the same kinds of roles, under one set of rules. Neither one is the other's tool, and neither one is the other's approver. They're teammates, and each does its own share of the work.

In practice that comes down to a few plain rules:

  1. Every role gets a job description, AI or person. It covers the same six things: responsibilities, decision rights, handoffs, escalation, limits and checks.
  2. No one checks their own work. AI drafts and a person refines. A person does the work and AI checks it. One agent checks another. The maker and the checker are never the same teammate.
  3. Handoffs go both ways, with a reason. AI can hand a decision to a person, and a person can hand a task to AI, as long as they say why.
  4. Authority is earned, one task at a time. A new agent starts narrow, like a new hire. Its limits widen when its track record earns it and narrow when the record shows mistakes.
  5. Escalation goes to named people, decided by complexity and impact, never to "the system."
  6. Cost is known before the work runs, with a limit no one can exceed, AI or person.
  7. Every decision is recorded with its reasons, from the AI and from the person, on your side. Early on this feels like extra effort. Over time it becomes your enterprise memory, the one part of your AI stack that keeps compounding and survives a change of vendor.

What changes

Take an ordinary example: an invoice that doesn't match its purchase order.

Under human-in-the-loop, the agent flags it and a person approves or rejects it. Multiply that by a few thousand a month, and approving is now that person's whole job.

Under hybrid intelligence, the agent's job description says it resolves mismatches under a set amount on its own, with a second agent checking each one. Anything above that amount, or anything involving a new supplier, goes to a named person in finance, with the agent's reasoning attached. The person's decision and their reason go on the same record. When that record shows the agent getting the small ones right month after month, its limit goes up. When it shows mistakes, its limit comes down.

The person stops being a queue and goes back to being a colleague who handles the work that actually needs a person.

The question worth asking

For three years the question has been "how much AI should we let in?"

That's the wrong question. The useful one is: for this decision, who leads, who does the work, and who checks?

Ask it one process at a time, and the gap between your two lists starts to close.

Rent the capability. Own the judgement.

I'm curious where this breaks in your company. Is it the handoff, the checking, or the record? Tell me in the comments.


Originally published on LinkedIn. I'm the founder of LeafCraft, the Hybrid Intelligent Enterprise platform, and we build the LeafMesh ADK (pip install leafmesh).

Sources: Dynatrace Pulse of Agentic AI 2026 · EY AI governance survey, Sep 2026 · VentureBeat, Jul 2026 · Dell'Acqua et al., HBS working paper 24-013 (2023)

Top comments (0)