Despite the rush to pilot AI, progress has stalled for many organizations, in large part because they don’t have the right underlying infrastructure in place to scale agentic workflows. Airtable research from 1,001 companies found that 56% are still in the earliest stages, where AI use is primarily at the individual employee level, and that only 21% of organizations are allowing AI to operate with real autonomy. So how do you graduate to the next level?

The most advanced companies are building agentic workflows across a shared operational surface, where agents and humans access the same single source of truth and have real-time visibility into each other’s actions. This allows these orgs to make the shift from employees directing personal AI assistants to supervising AI agents that can perform workflows that benefit the whole team. But there’s more to it; even as your operations become more autonomous, success relies on keeping humans in the loop (or “on the loop,” as Deloitte puts it.)

Our findings align with industry research and predictions. According to Gartner, “By 2028, over half of all enterprises will stop paying for assistive intelligence (such as copilots and smart advisors) and instead will favor platforms that commit to workflow results.” This means the time is now to build the foundation needed for reliable, observable agent workflows that fundamentally change the way humans and agents collaborate.

What is human-in-the-loop (HITL) in AI?

Human-in-the-loop (HITL) AI is a design approach where humans are embedded into an AI system's decision-making process at intentional moments throughout a workflow. These moments identify where a person should review, approve, correct, or redirect what the AI has done before work moves forward.

It might seem counterintuitive to build human checkpoints into an autonomous process, but it helps to safeguard outcomes andThe importance of human-in-the-loop AI for agents

AI agents can now plan, reason, and execute complex multi-step tasks — often without any human involvement. That capability is powerful for your business, but it must be thoughtfully implemented. You have to ask: when an agent makes a mistake, causes a compliance issue, or takes an action that has consequences down the line, who is accountable? Without human-in-the-loop design, that question becomes hard to answer.

Agents are capable, but not always consistent. Even well-designed agents can drift, hallucinate, or misinterpret context — especially when they encounter edge cases or ambiguous inputs. HITL creates checkpoints that catch errors before the work continues.

The level of human involvement may depend on the workflow. Some agentic processes run easily in the background, analyzing data and surfacing insights to internal teams. Yet others can have real-world consequences, especially if they are high-profile or customer-facing use cases. (Search results on ‘AI fails’ return any number of examples, often costly.) So it’s important to consider the scope that agents have. Even a small use case, to help you respond to customer return requests faster, for example, can have a large impact when your agent is updating a customer record or triggering a payment. These are tasks that need to be precise, and by adding a human approval step, you create accountability and a clear audit trail.

Beyond consistency and accountability, HITL workflows are also beneficial for trust and compliance. Many industries that handle sensitive data are building compliance requirements around AI-generated outputs, which may require you to build-in human oversight as regulations are developed. But human oversight also helps you adjust and expand your automation with confidence over time. Evaluate your AI readiness, and whether you have the right foundation in place to build and scale for human-agent collaboration.

How to design agentic workflows with a human in the loop

Good HITL design means being deliberate about where humans add the most value. Here are five tips to design agentic workflows that keep a human in the loop.

1. Separate recommendations from execution

One of the most practical HITL patterns is staging agent actions: the agent proposes what should happen, and a human reviews before it's applied. This means designing fields or workflow states like "Agent recommendation," "Approved," and "Override" — so nothing is executed until someone signs off. This is especially useful early in deployment, when you're still calibrating how much to trust a given workflow.

2. Define the human checkpoints before you build

The best time to add humans into the fold is when you’re designing an agentic workflow. Map which decisions require human review and which don't. For example, a content generation agent might run autonomously on drafts, but require human approval before anything is published or pushed to a customer. A compliance review agent might flag issues automatically, but require a qualified reviewer to clear them.

3. Build visibility into the agent’s reasoning

Effective HITL design relies on agent observability. You should be able to see both the agent outputs and the reasoning behind them to help course correct mid-flight or troubleshoot after the fact. This means surfacing agent recommendations in solutions that are accessible to non-technical teams, and that provide a shared operational surface where your teams and agents access the same data.

4. Hard-code your non-negotiables

While agents can reason through many things and make judgement calls, don’t rely on the agent to remember every constraint, every time. Instead, build defined rules directly into your workflow. (You might think about it like a form that you can’t submit unless all required fields are completed.) This way, your system has guardrails in place via status fields that can't advance until conditions are met, or required approvals that can't be skipped, to ensure rules are automatically enforced.

5. Design for gradual automation

Begin with more human oversight than you think you need, then dial it back as agents prove themselves reliable. Track where agents are accurate, where they need correction, and what patterns those corrections reveal. HITL workflows aren’t a constraint on automation, so much as safeguard that your implementation delivers the value it promises.

Examples of human-in-the-loop AI agent workflows

Fanatics Betting and Gaming operates in a compliance-intensive environment: offering real-time sports betting in accordance with state-by-state regulatory requirements, and 24/7 content demands. The brand’s team uses Airtable agents to monitor live player injury data and automatically flag when content needs to be updated — but humans stay in the loop for every compliance review and approval. Every asset is logged with its copy, disclaimers, and approval record, giving the team an audit trail that reduces compliance review from days to minutes. As Creative Operations Lead Zach Granowitz put it: "We're managing over 100 briefs at a time. With Airtable, everything is in one place, so creatives and operations hand off seamlessly."

Or take MGA Entertainment , the toymaker behind legendary L.O.L. Surprise!, Miniverse, and Little Tikes. The brand manages thousands of marketing assets across dozens of brands and hundreds of retailers — with new product launches every 90 days. Their marketing team uses Airtable AI agents to parse unstructured creative briefs, scan for SKU numbers, and automatically pull relevant assets from their digital library. Human project managers stay in the loop to review and schedule the resulting tasks. This cut creative brief processing time down by roughly 60%, with asset searches reduced from hours to seconds. As VP of Franchise and Integrated Marketing Ashlee Meese described it, "Airtable didn't just help us get organized. It keeps us organized, so we can operate with greater clarity, speed, and confidence."

Best tools for human in the loop AI workflows

Successful HITL workflows rely on a clearly defined infrastructure. They’re not just automations that humans check in on sometimes. Here’s what matters most as you build yours:

Airtable is built for human-agent collaboration

The organizations seeing the most success with agentic workflows are those who've built systems where humans and agents work together, with the right visibility, the right checkpoints, and the right accountability structures in place.

This is what Airtable is designed for. Both humans and agents operate on the same surface, with shared data, shared state, and shared context. Humans can observe what agents are doing, inspect their outputs, and step in before results become final. In this way, the workflow definition becomes a governance layer. Airtable also provides a structured operational context that enables the people who know the workflows best to oversee it, allowing them to step out of long development queues and into production.