Going for operation use cases was a huge win. Once novel data existed in the system (rather than just transforms of existing data), it became a lot harder to rip out. That could be as simple as having someone merge records so you know that two companies are actually the same.
Foundry was a really interesting case because it was basically an enterprise ETL platform before those became very popular + a team of people who helped you get data into it. One of the genius things about the business model was that it operated like a consultancy, but built contracts like a product company. That allowed them to charge based on the value provided rather than hours worked, then pull the best lessons from the deployment back into the main product.