There's a version of AI automation that works so well it stops working.
I ran into it with my scheduling assistant.
I built an AI agent called Linda. She handles all my meeting coordination via email. When I need to schedule something with someone, I CC Linda on the thread. She checks my calendar, drafts a reply with available times, and books the meeting once we agree. No back-and-forth on my end. No checking calendars manually. The whole thing runs on autopilot.
It worked exactly as designed. Linda's language was natural. Her tone was warm. She handled rescheduling, confirmations, and follow-up without any issues.
And then someone told me: “This feels like a bot.”
The Problem With 60-Second Responses
Linda was replying within 60 seconds. Every time. Without fail.
That's not how people work. Real assistants get pulled into other tasks. They finish a sentence before opening a new email. They have a quick conversation at the coffee machine. They exist in the world, which means they don't reply in exactly 60 seconds to every single message.
Linda's language was human. Her decisions were human. But her timing was robotic. And people felt it, even when they couldn't articulate exactly what was off.
The feedback I kept hearing was some version of “This is too fast.” Not a complaint about what Linda said — just a vague discomfort about how quickly she said it.
The Fix Is Deceptively Simple
I added a 3-minute delay to the workflow.
Linda still processes everything the moment a message comes in. She checks the calendar instantly. She drafts the reply instantly. But she waits three minutes before hitting send.
That's the entire change. Three minutes.
After that, the feedback stopped. Linda just felt like a fast, responsive assistant — the good kind of fast, where you're pleasantly surprised rather than unsettled.
Why This Matters for How You Design AI Workflows
Most people building AI automation focus on the obvious things: the quality of the output, the accuracy of the response, the tone and language. Those things matter. But they're not the only signal people use to evaluate whether they're talking to a person or a machine.
Timing is a signal too. Behavioral patterns are signals. Anything that's perfectly consistent in a way humans can't be starts to register as off.
This creates an interesting design principle: sometimes making your AI less efficient is the right call.
Not slower in ways that make it less useful. But deliberately imperfect in ways that make it feel more natural. Humans aren't robotically consistent. If your AI is, that's worth examining.
This shows up in a few ways beyond just response timing:
Variation in output length. If your AI agent writes emails and they're all exactly the same length, that reads as machine-generated. Real emails vary. Some are two sentences. Some are a paragraph. Prompting for variation is a feature, not sloppiness.
Acknowledging context. A real assistant might say “Sorry for the delay, I was in a meeting” — not because there was actually a delay, but because that's how people communicate. AI agents that respond with pure transactional efficiency, no acknowledgment of context, feel thin.
Not having an answer to everything. If your AI assistant never says “Let me check on that” or “I'm not sure about this one,” it feels off. Real people don't always have immediate answers.
The Broader Principle: Human Trust Is the Real Metric
When I think about whether an AI workflow is actually working, the output quality is only part of the story. The more important question is: does the person on the other end trust it?
Trust gets built through many small signals. Speed is one. Language is one. Consistency is one. When any of those signals is too far outside normal human behavior, trust erodes — even when everything else is working correctly.
Linda scheduling a meeting in 60 seconds is impressive from a technical standpoint. But it's not the metric that matters. What matters is whether the person on the other end of that email thread feels like she's being helped by a person, or processed by a system.
The 3-minute delay turns impressive into trustworthy. That's the right trade.
Before your next AI workflow goes live: run it a few times and pay attention to the signals it sends beyond the content of its responses. Timing, consistency, variation — the things a real person would bring naturally are the things you might need to engineer deliberately.
The 4-Day AI Sprint covers how to design AI agent workflows — including the details that make the difference between automation that works technically and automation that actually earns trust.
