Predictions say we will reach the yottabyte era (~1,000 zettabytes/year) by the early 2030s. Much of this data is in textual format, like websites, social media posts, emails, as well as academic publications. This information overload plagues both humans and machines alike. Automated summarization can reduce reading time, as well as the cost and bias of human summarizers, while also improving the effectiveness of indexing and other machine processing tasks. With the advent of the Transformer architecture, the task of automated summarization, like other NLP tasks, has witnessed enormous progress. Despite all this amount of work, several challenges obstruct the successful use of this technology to real-world applications that are slightly more complex than the typical, widely used in academia, news domain, where an article is given as input and a headline or two-sentence summary is requested in the output. The main goal of HARNESS is to develop novel neural abstractive summarization techniques to deal with three such key challenges, namely costly data acquisition, adaptation, and trustworthiness, paving the way for human-centered summarization systems that can deliver effective results in niche domains.