Read AI outputs like a smart intern's draft.

Personas, audience matches, channel splits, budget allocations, tactics and recommendations inside Medusa are AI-generated suggestions. They're a strong starting point, not a finished plan. Review, edit and apply your judgment before pushing anything live.

The product uses AI in a few specific places:

• Brief parsing. Extracting objective, audience, market and KPI from documents you upload.
• Persona generation. Drafting audience archetypes from your strategy.
• Audience matching. Mapping personas to targetable segments on Meta, Google, TikTok and LinkedIn.
• Channel mix & budget recommendations. Suggesting splits based on your goal, historic performance and benchmarks.
• Tactics & copy ideas. Prompts for creative angles, bidding strategies and targeting approaches.
• Optimisation recommendations. Surfacing budget reallocations and creative tests based on pattern detection over time.

Large language models occasionally generate output that sounds authoritative but isn't supported by facts. This is called hallucination. In Medusa, that might look like:

• A persona attributed with a behaviour that isn't in your brief.
• A platform audience that doesn't actually exist on that platform.
• A confident benchmark number that wasn't in the source data.
• A tactic that's outdated or inappropriate for your market.

We've tuned prompts, grounded outputs in your data via RAG and added confidence indicators where we can. None of that fully removes the risk. Treat AI output like a thoughtful first draft from a smart intern: useful, fast, sometimes wrong.

When Medusa shows a Predicted ROAS, expected lift, projected conversions, diminishing-returns threshold or any other forward- looking number, that is an estimate from a statistical model.

It can be wrong because:

• The underlying data is thin (new accounts, short history).
• Market conditions shift (seasonality, competitor moves, macro events).
• Platform algorithms change without notice.
• Attribution windows truncate real impact.

Wherever we show a forecast we try to show a confidence band or quality score next to it. Use those signals. Wide bands mean "test small first".

Recommendations are designed to be reviewed and approved by a human before they affect spend. The Approve button exists for a reason. If you enable Auto-Pilot for specific accounts, you set the guardrails (max budget shift, blackout dates, channels in scope) and you carry final responsibility for what runs.

Medusa won't autonomously launch campaigns or change creative. Auto-Pilot, where you opt in, is restricted to budget shifts within ranges you define, and every action is logged with a reversible audit trail. Manual review remains available everywhere.

The optimisation engine uses attribution-aware machine learning that adapts to your account history. "Learning" here means weighting historical patterns; it does not mean the model retrains end-to-end on your data and pushes shared improvements to other customers. Models are scoped per tenant unless you explicitly opt into benchmarking.

Medusa is a planning tool. It does not provide legal, financial, medical, regulatory, brand-safety or compliance advice. For regulated verticals (finance, healthcare, gambling, alcohol, political ads) your own compliance review applies, not ours.

Outputs from Medusa are provided as-is. We're not liable for decisions you take based on AI recommendations, predicted performance, or persona/audience suggestions. Read this alongside the Terms of Service for the full picture.

If an AI output is clearly wrong, biased, unsafe or misleading, tell us. Use the thumbs-down on any recommendation inside the product, or email hello@getmedusa.ai. We read every report and tune accordingly.