AI policy & ethics — Labelflow

Labelflow uses artificial intelligence in specific, narrow places where it earns its keep. We think AI in music tools is currently being marketed badly — oversold as the headline, underexplained when it actually matters — so this page exists to be specific.

The principle

AI in Labelflow is a draft layer, not a decision layer. It proposes; the human approves. Anything that touches an artist relationship, gets sent on the label's behalf, or affects the public version of a release passes through your hands before it ships. We do not run AI autonomously on your data without your explicit permission.

This is partly an ethical commitment and partly a craft argument: a label's voice is the thing that distinguishes one label from another. Outsourcing the voice to an AI defeats the purpose of having the label. Labelflow uses AI to take the friction out of repetitive work; the label still does the label part.

Where AI is used

1 · Communication drafting

Onboarding emails, premaster reminders, mastering coordination, status updates — Labelflow drafts the text in your label's voice based on templates you control. You approve, edit, or fully automate sending.

You can override the AI's tone at any time. You can edit any draft before sending. You can disable AI drafting per template or globally. The drafts default to "manual approve" for new label workspaces.

2 · Drift detection

When you edit a communication template, Labelflow compares your edits against the baseline model and surfaces meaningful changes (drift detection). This helps you catch unintended divergence over time. The drift detection itself is structural, not generative — it doesn't write new content, only flags differences.

3 · Visual asset generation

Cover art, Spotify canvases, event flyers, social-post graphics — the operational visual workload that surrounds every release. Labelflow can generate these in your label's aesthetic: drawing on your established palette, type, and image direction, and the artwork brief for each specific release. The AI proposes; you approve, edit, or skip the AI route entirely and commission a human designer. The label's visual identity is your call; the AI is a draft layer for the asset workload, not a replacement for art direction.

What we generate: cover artwork, Spotify canvases, banner and social-post graphics, event flyers tied to a release. What we don't generate: anything that imitates a specific artist's prior work without permission, anything that uses copyrighted source material as input, and — never — any audio output (see below).

4 · Future features (transparency about what's coming)

We're exploring AI assistance for: artist demo intake (auto-extracting structured data from submission emails), release readiness summaries, anomaly detection in release timelines. None of these are shipped at the time of this writing. When they ship, we will update this page.

What we don't use AI for

A&R decisions. Whether to sign an artist, whether to release a record, what to do with a demo — these are human decisions. Labelflow will never recommend signing or rejecting an artist.
Autonomous communication on your behalf without your approval flow. No "AI agent emails the artist for you" without an explicit human-in-the-loop opt-in.
Profiling artists. We do not generate AI profiles, predictive scores, or performance forecasts about artists, fans, or competitors.
Generative audio. Labelflow is not a generative-music tool. We do not generate masters, remixes, stems, or any audio output. The music itself remains the artist's.
Training third-party models on your data. See below.

Training data

We do not use your data to train AI models, and we do not allow our AI providers to use your data to train their models.

Specifically: API requests we make to upstream AI providers (currently Anthropic via the Claude API) are configured to opt out of training data collection where the provider supports it. We do not run our own model training on customer data. We do not sell, share, or otherwise hand your data to third parties for AI training purposes.

You can review the upstream providers we use in the Privacy Policy.

Disclosure

Communications drafted with AI assistance are not marked as such on the artist-facing side, because the human (you) approves and signs every draft — the message is yours. If a beta participant requests a system to mark AI-assisted communications, we will consider it as a feature.

Bias, errors, and accountability

AI models can produce errors, hallucinations, or subtle bias. Labelflow's drafts are based on templates and structured data you provide, which reduces but does not eliminate the risk. The human approval step is the safeguard. If an AI draft produces something inappropriate, please report it to hello@labelflow.ai — we treat these as bugs and prioritize fixing them.

The label manager remains responsible for what gets sent under the label's name. Labelflow is a tool, not a delegate.

EU AI Act

Labelflow's AI features fall under the limited-risk and minimal-risk categories of the EU AI Act. We voluntarily provide this transparency page in line with the spirit of the Act's transparency obligations, even where strict compliance does not yet require it. As the Act's implementing regulations evolve, we will update this page.

Questions or concerns

If you have questions about how AI is used in Labelflow, want to opt out of specific AI features, or want to raise an ethical concern about something the system did, email hello@labelflow.ai. We respond personally.

Last updated: 2026-04-29.