How our AI works

Full transparency, in line with Article 22 GDPR: our scores are indicative and contestable — never binding automated decisions.

Where our data comes from

Every analysis relies exclusively on the following sources — nothing more, nothing less:

AgentPro internal database

Player and club statistics, contracts and market history entered into or imported to the platform.

Anthropic Claude (AI analysis)

The AI model that produces our analyses. It only receives pseudonymized inputs — see the "Pseudonymization & Article 22" section below.

Market values

A public, community-run proxy of Transfermarkt data. An unofficial source, based on publicly available data.

Psychological profiles only

Web search (SerpAPI/Google), Wikipedia and Google News RSS — exclusively covering public facts about professional players, with an evidence chain (verifiable URLs).

Planned integrations, not yet active

Integrations with official providers (Opta, Wyscout, FIFA) are planned but not active to date: no current analysis relies on these sources.

The mental dimensions we analyze

Psychological profiles assess 12 core dimensions and 4 bonus dimensions, each scored from 1 to 10 by the AI.

Every score is grounded in sourced events: the analysis builds an evidence chain with the URLs of the sources consulted — never on mere impressions.

12 core dimensions

  • Discipline
  • Leadership
  • Resilience
  • Stress management
  • Adaptability
  • Emotional stability
  • Confidence
  • Team spirit
  • Motivation
  • Concentration
  • Winner mentality
  • Professionalism

4 bonus dimensions

  • Clutch factor (decisive moments)
  • Media management
  • Dressing-room influence
  • Longevity potential

Confidence level

Every profile carries a confidence level, determined by the number of distinct verified events found during the analysis:

  • High — at least 8 distinct verified events
  • Medium — 4 to 7 distinct events
  • Low — 1 to 3 distinct events
  • None — no verifiable data found

This level depends on the number of distinct events, not on article volume: 50 articles about the same red card count as a single event.

Pseudonymization & Article 22 GDPR

Four safeguards govern every AI analysis:

Anonymous tokens before any AI call

Player names are replaced with anonymous tokens (a salted hash in the format "p_" followed by 16 hexadecimal characters) before anything is sent to the model. Real names are restored server-side, in the response only.

Complete audit log

Every AI call is logged — service involved, cost, timestamp — for audit purposes.

Indicative, contestable scores

Every score carries the "Indicative score — not a decision" badge and a "Contest this score" button (limited to 10 contests per day), which triggers a human review.

No training on your data

No model is ever trained on our clients' data: we use an inference-only API, with zero fine-tuning.

Accuracy & backtesting

We publish no accuracy figure to date: we refuse to display a percentage that is not scientifically substantiated.

A backtest with honest error intervals is in preparation; its results will be published as soon as they are available.

Until then, valuations should be read as indicative ranges — never as guarantees.

Known limitations

Our analyses have limits, and we prefer to name them:

  • Possible biases in public sources: uneven media coverage, dominant languages.
  • Uneven coverage across leagues: less-covered competitions yield less usable data.
  • Market data freshness depends on the public proxy we use.
  • Psychological scores are a reading of public events, not a clinical diagnosis.

Going further

These pages complement this methodology:

A question about our methodology? Our data protection officer will answer you.

Contact our DPO