Suggested actions were becoming the make-or-break part of Goodword.
The product could store notes, support search, and help a person review relationships. Suggested actions were meant to do something more active. They needed to identify a useful next step without requiring the person to search for it first.
That could mean remembering a commitment, noticing that a conversation needed a reply, or presenting a relevant reason to reconnect. The suggestion had to be specific enough to act on and grounded in information the person had authorized Goodword to use.
The difficulty was timing. A recommendation system improves when it has more examples and feedback. A new product has to release before it has either in large quantities.
We expected the first versions to be uneven. The team still needed testers to see them so we could separate useful suggestions from generic or poorly timed ones.
The interface also had to make the source of a suggestion understandable. A person should be able to recognize why an action appeared and decide whether it matched the relationship. The application should not imply certainty where it only had a possible next step.
The private record contains individual examples and product data. None of that is included here. There are no accuracy figures, tester records, or internal model details in this article.
The public-safe event is the product decision. Suggested actions were no longer a secondary addition around search and notes. They were part of the main reason somebody might return to the application.
As the designs came together, that dependency became more visible. We could continue improving the screens, but retention would depend on whether the product regularly surfaced something worth doing.
The immediate work was to release the current version, observe the suggestions in use, and keep correcting the cases that were not useful.