The requirement hiding inside the word ‘automatically’

An operations manager at a field-service company told me he has learned to flinch at one word in an AI brief. Automatically. As in, “the system will automatically schedule the technician” or “it will automatically flag the urgent tickets.” The word sounds like a feature. He treats it as a warning that three requirements are being hidden inside a single adverb, and that nobody has agreed to them yet.

So when someone says a feature will do something automatically, he stops and asks them to unpack the word. It is not pedantry. Each thing the word hides is a real decision that will surface later, usually at the worst time.

What the word is actually hiding

He walked me through what “automatically” almost always smuggles in. There are three assumptions, and they are rarely true at the same time.

No human checks it. Automatic implies the action happens without anyone confirming. That is a claim about how good the model has to be, because now a wrong answer acts on its own.
It happens in real time. Automatic usually implies immediate. That is a claim about latency and infrastructure, and it is very different from a job that runs once an hour.
It is always on. Automatic implies the thing works nights, weekends, and holidays without a person nearby. That is a claim about monitoring and what happens when it breaks at two in the morning.

Say “the system automatically dispatches the nearest tech” and you have quietly promised all three. No dispatcher confirms the assignment. It happens the instant the ticket lands. It works at three on a Sunday when the on-call list is thin. Each of those is a design decision with a cost, and the word buried all of them.

He gave me a concrete one. A team wanted urgent tickets to be automatically escalated to a supervisor. Fine, until he unpacked it. Automatically meant no human triage, so the model’s false alarms would page a supervisor at night directly. It meant real time, so a slow classifier was not acceptable. It meant always on, so someone had to own it when the escalation pipeline died over a holiday weekend. Once those three requirements were on the table, the team decided a human should confirm escalations during the day and only truly critical patterns should page at night. The word had been promising a system nobody actually wanted.

Unpack it into requirements you can see

His fix is simple and a little annoying, which is why it works. Every time “automatically” appears in a brief, he replaces it with the three answers. Who checks it, or no one. How fast, in a real unit of time. What happens when it breaks, and who gets the call.

A friend who leads engineering at a utility does the same with the word “effortlessly,” which she says hides even more. Her point was the same as his. These soft words feel like they describe a quality of the experience, but they are actually unpriced requirements. Once you make the team write down the concrete version, half the time they realize the automatic version is more machine than the problem needs, and a suggestion a person confirms would be cheaper, safer, and just as useful.

The deeper habit is to distrust any adverb that describes how the AI behaves rather than what it does. “Intelligently,” “effortlessly,” “automatically,” “proactively.” Each one tends to hide a requirement that someone will pay for later if nobody names it now.

How we approach it at Density Labs

In the AI Opportunity Assessment, our fixed two-week engagement at $2,500, part of the work is translating soft language into hard requirements. When a brief says something happens automatically, we make it explicit. We write down whether a human confirms, how fast the answer has to arrive, and who owns the failure. That translation often changes the design before a line of code exists, and it usually makes the build smaller and safer than the original word implied.

When someone tells you an AI feature will do something automatically, do not nod. Ask them to say the word without saying the word, and watch three requirements appear.