Record/replay might be valuable to record all steps necessary to make a bug reproducible. But what if running recorded sequence the next day you need to enter next day's date, and your page has no "today", but concrete date was entered? All your pre-recorded steps are useless now.
And for regression testing, you want to use PageObject design pattern, so business logic, locators and actions are centralized, so if you need to change them later, you don't have to change all instances of record/replay which use them.
I can see use-case for record/replay, when manual tester records performing the operations and later can review if interacted elements have good locators (easier that inspecting locator for every element being interacted with).
Other than that, record/replay is IMHO of limited use, and only for pages which are designed with testing using record/replay in mind ("click here to insert current date", etc).
Bit more about our experience.
We have a (huge) suite of tests similar to record/replay (FitNesse) and (slowly growing) suite of new test which use PageObject patern. FitNesse are fragile, and if GUI/business logic change is made, many test break and needs fixing in many places. As a result of this fragility, test developers are reluctant to stray away from "happy path", and most tests are repetitive. We are working on retiring the whole suite (and replacing them with PageObject tests).
If PagoObject tests break, they likely need fixing in a single place, so as a result we are more adventurous and more willing to implement functionality straying away from "happy path", inspecting data and selecting some data in random, creating more "fuzzy" tests - and it is also much more fun for test developers to develop these tests (because developer can use Python debugger to inspect located elements), and such tests are much easier to debug - especially later, in maintenance phase when something changed and test feailed.
So (in my experience), record/replay tests are of a limited use as a starting point for good PageObject-related tests. Hard part is to design good page methods for each page (useful for tests), separating the interaction in proper size "chunks" where group of widgets are acted upon in a single action which makes sense for a user/test. Getting locators and order of interaction (which is generated by record/replay) is the trivial part.