I think the greatest help right now would be creating tutorials/code snippets for efficient use of OSM data in external apps, which are easily findable and applicable by AI agents. Ideally they would push users towards hosting their own data but I understand that this will be very difficult in the age of why-spend-effort-on-hosting-when-things-are-free-in-the-internet age.
There are currently no heavy users of Overpass that really stand out (and they would quickly get banned anyway). The servers are seeing a very, very long tail of small time apps and websites. On top of that we have the notorious scrapers. They are also mostly small load and keep requests to an acceptable level but they cause a base level of load that is completely unnecessary. They would be significantly faster in their work if they would just download the darn planet and filter it with osmium. These are unfortunately also the users that are the most difficult to reach.
Finally, the biggest offender right now is osmNX, or to be more precise, one or more users abusing the library through botnets to scrape OSM data at a massive scale. If somebody wants to help bring the PR for reading from planet dumps over the finish line, then this would be a candidate for reducing traffic on the servers.