There is a gap between what AI tools promise and what they actually deliver day-to-day, and understanding where that gap lives is worth the effort. The tools are genuinely impressive. They help. But the overhead around using them often stays constant, or quietly grows. That nagging feeling resolves once practitioners consider a distinction that sounds simple but changes a lot in practice: the difference between a prompt and a loop.
The invisible work around every prompt
Think about something routine — say, preparing for a school trip. Nothing on that list is technically hard. But the effort is not in any single task; it is in the coordination: checking the weather, confirming the pickup time, realising the calendar conflicts, remembering that last year’s raincoat no longer fits. All that routing between apps and systems lives, until you fix it, entirely in your head.
That wiring between surfaces is the loop — the recurring responsibility that wraps around every individual task. A prompt (a question or instruction you give an AI tool) helps with one step inside that loop. But if you still have to notice the trigger, gather the right sources, hold the context across sessions, judge the output, take action, and seed the next pass yourself, then the loop is still entirely your problem. The AI removed some typing. The recurring responsibility stayed put.
This is why the first wave of AI inside productivity tools has felt underwhelming to many practitioners. The individual surfaces got smarter. The hidden work between surfaces did not.
From prompts to loops
The reframe worth returning to is to stop asking “how can AI help with this task?” and start asking “what recurring responsibility no longer needs to be carried entirely by hand?”
A loop, in this framing, is a narrow and bounded thing. It has a trigger (what wakes it up), sources it trusts, memory of what has changed since last time, a set of actions it can take safely, an explicit stopping point where human judgment is required, and a record that makes the next pass smarter. That last element — the record — is easy to skip and costly to skip. Without it, every pass starts from scratch.
The stopping point matters as much as anything else. A useful loop does not act and surprise you; it drafts and stops. The question is not whether AI should handle judgment calls. It is where the boundary between safe action and required judgment actually sits. Locating that boundary precisely is most of the design work.
The five questions that map a loop
There is a clean set of questions that helps turn any recurring responsibility into this structure:
1. What is true right now? The current state of the trigger and its sources.
2. What changed since last time? So the loop is not re-doing work that is already done.
3. Which source should be trusted? Not all inputs are equal. Stale data, weak signals, and already-contacted people need flagging before action is taken.
4. What can happen safely, without me? The narrow set of actions the loop can take before it must stop and wait.
5. Which other recurring job needs to know? This is the question that turns a single loop into a loop of loops.
The fifth question is where the leverage shifts significantly. Loops that can notice each other — where the output of one becomes the trigger or context for another — start to replace the human as the integration layer, rather than just improving a single surface.
When loops start talking to each other
Rain changes packing; an earlier pickup changes the calendar; a promising candidate wakes reference checks, scheduling, and compensation thinking at the same time. We tend to think of tasks in isolation. Most of the relief in real working life comes when the relationships between tasks stop living entirely in your head.
Normal automation struggles with this because it assumes the world holds still. Real recurring responsibilities have stale sources, judgment calls, and history that matters. A set of narrow, bounded loops that can hand off to each other is better suited to this texture. Each loop stays small enough to be reliable. Together they cover ground that a single general assistant cannot hold without becoming fragile.
What this changes in practice
Moving from “which app to open?” to “which recurring obligation is being handled?” sounds like a small shift. In practice it changes what you build, what you measure, and what counts as success.
The right entry point is a small set of narrow loops with hard stopping rules — less showy than a general AI agent (a system that handles multiple tasks autonomously), more reliably useful. The harder question, the one worth sitting with, is what whole process you are willing to hand over and actually trust. That is where the leverage lives, and we are only at the beginning of learning how to answer it well.
Text summarized and optimized using Anthropic’s models and reviewed by a human.