Stanford’s Social and Language Technologies Lab just revealed something that many of us know inherently: employees resist AI because they don’t feel they can trust it. 

According to recent research from SALT, the top most common fears that workers have about AI automation in their work are “lack of trust (45%), fear of job replacement (23%), and the absence of human touch (16.3%).”

At StackAI, we understand that complex, process-heavy workflows in regulated industries cannot simply be automated away. But we also believe that the opportunities for major operational and time savings lie in agentic automation. That’s where our focus on human-in-the-loop workflows enters. 

The most valuable AI workflows we see in production are clear about where human judgment belongs, and they design everything around that. AI models might extract, draft, retrieve, search, and propose, but humans always approve, edit, or reject. Along the way, every step gets logged. What is produced is work that an auditor, a customer, and a CFO can all respect.

Human-in-the-loop (HITL) is the design pattern that enterprise AI actually needs, and it captures what we think AI should be. Not a black box that spits out answers no one can explain, but a glass box that a real human stands in front of. 

What Human-in-the-Loop actually is

A Human-in-the-Loop workflow runs the agent until it hits a step it shouldn't take on its own authority. Then it pauses and sends what it has to a person: a drafted email, a proposed CRM update, a risk score it wants confirmed. That handoff goes through whatever channel the person already works in: Slack, Teams, email, etc. While the expert reviews and responds, the workflow remains paused the whole time. When the response lands, it resumes running, integrating the human decision into the next step.

HITL creates a deal between agentic AI and the enterprise: AI can do everything up to the point where human expertise and judgment is required, and at that point it has to stop and ask. In return, the organization gives it access to real tools/functions, data, and access to the steps before and after the gate.

Why this is the version of AI we believe in

When we talk about AI that runs right, HITL is what we mean.

Running right means the workflow is legible. Open the log of any run and see the path it took, the tools it called, the data it pulled, the prompts it used, and how long every single step took. And it means every consequential action has an approver by design. 

The agents in these workflows read hundreds of pages of policy, draft answers to questions that used to eat hours of someone's day, call external APIs, and reason across multiple sources. That's what models are good at. But people are good at judgment in the rocky spots, edge cases, and nuanced situations that take years of human experience to navigate. 

After a workflow runs with a human gate for a few months, the team has a clean record of where people approve with zero edits and where they do real rework. Wherever humans are stamping their approval, AI has earned more trust and can take that step on its own. Autonomy gets earned, one step at a time.

Use cases we see

The shape of a HITL workflow stays consistent across industries. The agent prepares, the human approves, the system executes and logs. These are examples I’ve seen from real production deployments. 

Financial services and compliance is where HITL carries the most weight. In a typical anti-money-laundering review, the agent pulls transaction history for a flagged account, screens it against external watchlists, retrieves prior case notes, and drafts a structured risk memo. The agent can do all of that on its own. The disposition of the case is where a person has to weigh in: clear it, escalate it, or file a suspicious activity report. So the agent posts its memo into a Slack channel the compliance team watches, with the disposition options as buttons. A licensed officer makes the call, the workflow resumes, writes the decision back to the case management system, and closes out. 

The same shape recurs in arrears analysis, where the agent reviews delinquent accounts and proposes outreach but a senior analyst signs off on the wording before anything goes out. In alternative-investment fund-flow review, the AI classifies paid and received movements and routes each one for approval before the entry hits the ledger. 

Sales teams see some of the most obvious ROI here, in RFPs and RFIs. Responding to an RFP used to take weeks, with subject-matter experts answering the same questions they'd answered fifty times. In the HITL version, the agent retrieves prior answers from a knowledge base, drafts a response to each question, and pings the deal owner in Slack, one question at a time or in batches, with an option to approve, edit, or reject. The owner only spends attention on the questions where the AI is unsure or wrong. 

Lastly, IT, operations, and support workflows often touch sensitive customer data, which makes the case for HITL plain. Invoice approval might be the most archetypal case of all: the agent extracts line items from incoming invoices and posts them to a finance channel, and payment only queues after a human clicks.

Looking ahead

Over months, the human gate tends to move. From the data that a team’s own approvals generate, they learn which steps the AI handles well and which it doesn't. The reliable processes get promoted to full automation, and the gate slides downstream to the next consequential decision. The less certain steps either get better prompts, better retrieval, and better tools, or they stay manual intentionally.

This is the version of enterprise AI we believe in. It isn't full autonomy that hopes no one notices the failures, and it isn't caution so paralyzing that the AI never touches anything meaningful. It's AI that runs with a person in the loop wherever the loop needs one. Over time you end up with a portfolio of workflows where the person does less of the rote work and more of the hard—and human—work.