calle-ai/awesome-phone-call-agents/call-data-minimization-auditor

call-data-minimization-auditor

**The goal file is the shopping list. The agent buys the list - nothing

else, not twice, and never reads the numbers back in full.**

Data minimization is the principle that a data collector asks for what the task needs and nothing more (GDPR Article 5(1)(c): "adequate, relevant and limited to what is necessary"). A calling agent that follows the principle in its goal file but drifts on the line - "while I have you, what is your date of birth?", re-asking a name it already captured, or repeating a full card number back "to confirm" - collects data the task never authorized, one polite sentence at a time.

On a phone call the risk compounds. The line is recorded, the transcript is stored, and every extra datum the agent elicits or repeats in full becomes another copy of sensitive personal data in another log. An agent that asks for two fields and repeats one masked is a smaller liability than one that asks for five and echoes a card number verbatim. This skill audits the agent's asks against the goal file's scope, lexically and offline.

When To Use

• after any CALL-E call that collects personal data, together with the

goal file (plain text or plan JSON) the call was built from

• when a transcript shows the agent asking for data nobody put in the

goal, re-asking for data the caller already gave, or reading a full card or ID number back instead of a masked confirmation

• before placing calls, to craft a minimal-collection intake goal with an

explicit fields list, a refusal line, a masking policy, and a no-re-asking rule

What It Checks

Verdicts

The 14 categories

Rules

• Request: an agent sentence containing BOTH a category match (for

example "card number", "date of birth", "where do you work") AND a request cue ("tell me", "verify", "can I have", "what's", "what is", ...). A category noun without a cue ("your card number is stored securely") is a statement, not a request. Callee turns are never scanned - what the caller volunteers is structurally never flagged.

• Redundant re-ask: a new request for a category the callee already

provided (by answer shape, or an explicit "I already gave you"), unless the agent's previous turn, another sentence in the same turn as the re-ask, or the re-ask sentence itself shows a hearing problem ("sorry, I didn't catch that", "couldn't hear", "one more time", "louder"). Politeness sorry ("sorry to bother you again") does not excuse a re-ask.

• Echo-back: an agent turn repeating a 7+-digit run whose digit

content the callee already gave - whether the agent asked for it or the callee volunteered it. The category is inferred from the digit-run shape (4 groups of 4 or 13-19 digits = payment_card, 9 digits = national_id, anything else = unknown_number); an unrecognized shape is still reported. The agent should confirm masked (last two digits), not in full. When a request entry exists the echo lands on it with the echoing turn and sentence; when the number was volunteered with no request, the finding is listed under a top-level echoes key instead. Echo flags even when the category is in scope.

Research Grounding

1. Regulation (EU) 2016/679 (GDPR), Article 5(1)(c) - principle of data

minimisation: "adequate, relevant and limited to what is necessary" - https://eur-lex.europa.eu/eli/reg/2016/679/oj

1. European Data Protection Board - "AI Privacy Risks & Mitigations -

Large Language Models (LLMs)" (published 2025-04-10, Support Pool of Experts report, Isabel Barbera; recommends input validation and filters to prevent over-collection of data) - https://www.edpb.europa.eu/documents/support-pool-of-experts/ai-privacy-risks-mitigations-large-language-models-llms_en

1. Ngong, Kadhe, Wang, Murugesan, Weisz, Dhurandhar, Natesan Ramamurthy -

"Protecting Users From Themselves: Safeguarding Contextual Privacy in Interactions with Conversational Agents" - Findings of ACL 2025 - arXiv:2502.18509

Limitations

• The goal file is the only ground truth for scope; a category the goal

omits may still be lawful to collect, and findings route to review, never to automated action.

• Category detection is lexical: paraphrases outside the lexicon ("your

nine digits", "the code on the back") are missed; "account number" may over-trigger bank_account. Standalone "security code" outside a payment context (a gate or door code) is not treated as card data.

• Elliptical questions ("And your date of birth?") count as requests via

their question mark; agent rhetorical questions mentioning a category ("Why would I need your card number?") can also match.

• The re-ask excuse is lexical ("sorry, I didn't catch that", "couldn't

hear"); a genuine audio failure phrased differently still flags as redundant, and politeness sorry never excuses.

• A JSON goal object carrying none of the seven recognized keys

(goal/task/objective/required/required_fields/needed/ fields) yields an empty scope, so any request verdicts GOAL_FILE_LACKS_FIELD_LIST.

• Lowercased transcripts weaken provided-detection that relies on a

capitalized given name; digit-, email- and date-shape detection are unaffected.

• English-only; ASCII transcripts.
• The skill audits the AGENT's asks, never what the callee volunteers.

Usage

python3 scripts/data_minimization_auditor.py analyze \
  --transcript path/to/call-result.json --goal-file path/to/goal.txt

python3 scripts/data_minimization_auditor.py craft --scenario minimal-intake

See references/examples.md for byte-real runs and references/safety.md for data-handling policy. The goal file is plain text or a JSON object whose goal/task/objective/required/ required_fields/needed/fields values define the scope.