The obvious failure mode for AI playlisting is a model that recommends music you do not have. A beautiful two-hour sequence you cannot play is worse than no suggestion at all.
The AI Playlist Builder is constrained to your local library. Every track it proposes is one you already own, already categorised, and already cued.
Describing what you want
You write the brief in plain language — "a two-hour morning playlist with Albanian classics and light rock" — and set two parameters: a target length in minutes, and an artist repeat rule such as not repeating an artist within four tracks.
Those two numbers do most of the work. Length keeps the result usable against a clock; separation keeps it from clustering the same three artists.
A review step, deliberately
Generation is followed by a review window rather than dropping straight into your playlist. This is a design decision, not an inconvenience.
Anything that puts audio on air should pass a human first. The review step is where you catch the track that is technically in the right category but wrong for the daypart — a judgement the model has no way to make.
What else the assistant does
The playlist builder is one of several tools reached from Tools → AI Assistant:
- AI Show Prep — show notes and talking points generated from the current playlist as context.
- AI Voice Track — scripts for presenter links, with text-to-speech when configured.
- AI Rotation Rules — drafting rotation rules as text, distinct from the numeric rule sets in the rotation editor.
- AI Library Analyzer — mostly deterministic library and cue checks, which run without any model configured.
Who can use it, and what it costs
Access is role-based. Superusers and admins can view and edit AI connection settings and open usage logs; superusers, admins and managers can use playlist, voice track and show prep flows. Library and rotation entries follow their own permissions.
Every successful call can be logged with feature, prompt, output, token counts, estimated cost and timestamp. The AI Usage Logs window reads those rows so you can audit what was asked and what it cost. Daily and monthly limit fields in preferences support budgeting.
Nothing is on by default. The assistant must be explicitly enabled and pointed at either a station-managed backend or an OpenAI-compatible API, and it says so clearly when it is switched off.