Human-in-the-loop: how we build AI agents you can actually trust

Every business owner we talk to has the same two reactions to AI agents. First: "that would save us so much time." Second: "but I'm not letting it email my clients unsupervised." Both are right. The trick is designing systems that respect the second without throwing away the first.

The pattern: agents do the work, humans make the calls

We split every workflow into two kinds of step. Work steps are research, checking records, drafting, scheduling and chasing. The agent does these, fast and tirelessly. Judgement steps are anything that commits the business to something: sending, paying, promising. A person approves those, ideally in batches so it takes minutes, not hours.

Four rules we build into every agent

1. It can't act externally without approval until you explicitly decide it can, for a specific, low-risk action.
2. It checks history before acting. Has this person already been contacted? Is there an open conversation? An agent that doesn't check will embarrass you.
3. It never invents. If it can't verify a name, an email or a fact, it discards or escalates. A missing record is fine; a made-up one isn't.
4. It escalates ambiguity. Bounced email? Messy history? It flags it for a human instead of guessing.

What this looks like in practice

We run our own outreach on exactly this pattern. In its first eight days live, the agent researched 81 firms, verified 26 real decision-makers, and launched four campaigns, with every single email approved by a person before it went out. Zero duplicate sends. Zero missed follow-ups. Read the full case study.

Speed from the agent. Judgement from the human. Trust from both.

When to loosen the reins

Approval is a dial, not a switch. After a few weeks of reviewing outputs you'll see which categories are always right. Those can move to "send, then notify". The rest stay supervised. You stay in control of where that line sits.