Policies · AI Training Data
AI Training Data Policy
By default we do NOT use your calls to train AI models. If we ever want to, we ask you first. If you opt in, we use only fully anonymised data, never voice characteristics, and you can withdraw at any time. We never sell or share your data with other AI companies.
Default position — no training without opt-in
UK GDPR's purpose limitation principle (Art 5(1)(b)) says we cannot use personal data collected for one purpose (handling your call) for an incompatible new purpose (training an AI model) without a fresh lawful basis. We treat AI model training as a separate purpose, requiring separate, informed, freely-given opt-in consent under Art 6(1)(a). No call is used for training without that consent. Stricter than what some other companies do — the right interpretation of UK ICO guidance and the safer answer.
What 'training' means
In scope (opt-in only): improving the AI model using transcripts or audio; fine-tuning the AI on Taurex-specific terminology (where personal data); vendor's own model improvement using our data (DEFAULT DISABLED); anonymised statistical analysis if it could re-identify. NOT in scope (other lawful bases apply): live audit, quality monitoring, sample review (service quality LIA already disclosed); specific human reviewing a call for a complaint or claim (investigation, separate basis); aggregated non-personal statistical analysis (e.g. 'average call length').
How we get consent
If we want to start using call data for training: publish clear opt-in notice on customer portal and in AI greeting. Opt-in SEPARATE from any other tick-box — not bundled with T&Cs acceptance. Plain language explanation: what is used, by whom, for how long, where processed, customer rights. Customer ticks positive opt-in — silence is not consent. Customer can withdraw at any time with same ease as giving (UK GDPR Art 7(3)). If not opted in, NONE of their call data enters training. No penalty — no degraded service, no extra cost, no obstruction.
What's used if opt-in given
Raw audio: NEVER directly used for model training. Voice characteristics (timbre, pitch, accent): NEVER extracted, never trained. Transcript: only if opt-in AND anonymised (names, account references, phone numbers, addresses, vessel / container IDs, monetary amounts, dates redacted). Conversation summary: only if opt-in AND anonymised. Intent labels: yes even without opt-in (no personal data). Sentiment markers: yes even without opt-in (no personal data). Aggregate analytics ('how often callers ask about lithium batteries'): yes — aggregate, no personal data.
Anonymisation standard
Opt-in training data is processed by an automated PII redaction pipeline (vendor's, audited by us) to remove: names (caller, third parties, beneficiaries); email addresses, phone numbers, postal addresses; customer account IDs, booking references, shipment numbers; vessel / container / AWB numbers; financial figures and bank details; dates of birth, ID document numbers, passport numbers; IP addresses, device identifiers. Output sample-reviewed by a human before joining training set. Any record that fails review is excluded. We follow the ICO's standard: data is 'anonymous' only when it cannot be re-identified by reasonably likely means. Where residual re-identification risk (e.g. a specific lane and date may still be identifying), we generalise further or exclude.
Vendor model improvement — no sharing
By default, our DPA with the vendor PROHIBITS the vendor from using Taurex calls to train the vendor's own model — outbound model improvement DISABLED. If the vendor offers a meaningful benefit for opt-in we evaluate, and where appropriate: re-negotiate the DPA with documented terms; obtain customer opt-in; use only anonymised transcripts; confirm vendor cannot leak data to other tenants; confirm trained model does not 'memorise' specific calls (representation + technical safeguards). We do NOT share call data with third-party AI / ML companies for their model training. OpenAI, Anthropic, Google, Microsoft, Amazon, Meta, and other AI labs do not receive Taurex call data unless they happen to be our chosen vendor's underlying model provider, in which case our DPA limits processing by them.
Retention and withdrawal
Anonymised transcripts in training set: 24 months from inclusion → reviewed for continued relevance, re-anonymisation check, deletion or retention extension with documented justification. Trained model parameters: per vendor lifecycle (typically 12–18 months) → replaced by next model. Audit log of training set composition: 6 years for accountability. Withdrawal of consent: email [email protected] or portal Profile → Privacy → 'Use my calls for AI training' → off. Future calls excluded. Past calls already in training set: where technically possible we remove from future training cycles; we CANNOT remove from already-trained models (openly explained at point of consent). No fee, no penalty, no degraded service. Special category data (Art. 9): CALLS EXCLUDED from training — anonymisation cannot reliably remove special-category content.
Last reviewed 2026-06-06.