PP&A drafted the interview guide, refined it with the client's research team, and completed ten hour-long interviews between June 25 and June 28, 2024. Six interviewees were technology decision-makers, from a chief data and analytics officer to an AI researcher and an engineering founder. Four were legal leaders, including a deputy general counsel, a chief legal and compliance officer and a privacy attorney. Their organizations spanned AI start-ups, developer software, semiconductors, digital health and electronics manufacturing.

Each conversation moved from current model use to how organizations evaluate and approve models, then to license terms. Two screen-share exercises anchored the discussion. In the first, interviewees picked the two most and two least important of nine selection criteria, from cost and model quality to parent-company reputation and licensing terms. In the second, they compared three anonymized model licenses, one of them the client's, across five terms: indemnity in each direction, commercial restrictions, attribution and acceptable use, then reviewed a sample acceptable use policy. Probes covered pre-approved licenses, limits on commercial use and whether a license that mirrored a standard open-source license would make a model more attractive.

PP&A delivered the results one week after the first draft of the guide as a structured notes workbook rather than a slide deck, with one column per interview and one row per section of the guide, so the client's research team could compare technology and legal views clause by clause.

License terms rarely decided which model won. Eight of the ten interviewees placed model quality or capability for the use case among their top two criteria, and six named cost. Only two, both legal leaders, ranked licensing terms in their top two, one of them only for content generation. Brand reputation mattered less than the client expected: most interviewees placed it or customer support at the bottom.

Licensing worked instead as a gate. At most organizations legal teams reviewed model licenses before use, while several pre-approved standard permissive open-source licenses. Approving a new model took one to four weeks. Individual terms could still rule a model out. Provider indemnity split the panel by function. Legal leaders treated it as the most consequential term, and one had steered clients away from a leading model for content work because its indemnity did not cover outputs. The same leader had approved an open model for back-office work only, since its license lacked the enterprise protections that content generation needs. Technology leaders called narrow indemnities a gimmick or did not expect them.

Attribution divided product teams: one engineering leader called prominent branding a no-go because it would disrupt the product experience, while others saw no issue. Limits on commercial use at scale worried few interviewees, though some asked how usage would be measured and two preferred usage-based limits. Acceptable use policies were broadly accepted, but two interviewees flagged the clause on disclosing known dangers as too broad. Several expected licensing to matter more as AI regulation and litigation mature. The workbook showed the client which clauses create friction for each function and which do not.

• --