Bloomberg has expanded its Vault communications surveillance platform with AI policies for insider dealing and personal trading. The launch follows the integration of BSpeech, which converts recorded calls in more than 50 languages into searchable transcripts.
The products extend Vault’s coverage across voice, email and chat. They also address the pressure that rising communications volumes and expanding channel coverage place on compliance teams.
Surveillance Under Pressure
“Trade surveillance is being stretched across all three areas: coverage, alert quality and investigative capacity,” said Mike Googe, Product Manager for Bloomberg Trade Compliance Analytics.
Firms trade across more venues and asset classes, often using separate surveillance systems for different desks and regulatory requirements. They must reconcile data across these systems, maintain data quality and confirm that their controls provide adequate coverage.Googe said alert quality was probably where pressure was most visible. Trading behaviour can change faster than established rules. Firms also continue to refine their control calibration and definitions of a false positive.
Historical analysis and backtesting can improve calibration. New trading patterns will still create a lag between changes in behaviour and updates to surveillance rules.
AI is starting to change how firms approach this work. Bloomberg has seen some large institutions move agentic surveillance tools from testing into production, particularly for initial alert reviews. Googe said these tools require oversight, documentation and model-risk controls.
Searchable Voice Surveillance
Voice has remained one of the largest gaps in communications surveillance because reviewing hours of recorded calls at scale has proved difficult. Transcription converts that audio into a data source that firms can search and monitor.
BSpeech uses Bloomberg’s financial data corpus and domain-specific machine-learning models to transcribe calls. Bloomberg refines the models to reflect financial terminology and changes in language.
Within Vault, transcripts enter the same search and surveillance environment as chat and email. Firms can apply lexicon and AI policies. Analysts can return to the original recording when they need more context.
Perry Goetz, Global Head of Compliance Solutions at Bloomberg, said accents, tone, context and ambiguity still require human judgement. “The goal is not to treat voice as identical to written communications,” he said. “It is to make systematic surveillance of voice possible at a scale that was historically difficult to achieve.”
Risk-Specific AI Policies
The Insider Dealing AI Policy identifies communications that may indicate the improper acquisition, sharing or misuse of material non-public information (MNPI). It analyses content and context for signs that someone could gain an unfair market advantage.The Personal Trading AI Policy covers communications about personal trades and investment accounts. It also identifies possible breaches of a firm’s internal personal-trading policies.
Vault offers configurable lexicon-based and AI-based surveillance policies. Its AI policies use machine learning and natural language processing to analyse communications against defined compliance risks. Bloomberg trains the models on permissibly licensed data and documents the intended scope and risk coverage of each policy. Large language models enter the process after initial model inference. Bloomberg says this additional stage improves alert relevance and reduces false positives.
Commenting on the launch, Jotham Banyikidde, Investment Compliance Senior Associate at Impax Asset Management, said: “Bloomberg’s AI-powered surveillance models have substantially improved alert quality and reduced false positives. This allows our team to focus on meaningful review and oversight. Just as importantly, the models are transparent and well documented, which supports our internal AI impact assessment and gives us confidence in how alerts are generated”.
AI-Assisted Investigations
Goetz said AI can take on more analytical work by surfacing relevant information, connecting signals and helping analysts understand why a communication requires attention. Vault presents surveillance events for human review, and clients determine whether further action is appropriate.
“AI should enhance investigators’ capabilities, not replace human judgment or accountability,” he said adding that AI should support human decision-making when a material conclusion is required.
“Where a material decision or conclusion is required, a person should understand the supporting evidence and remain accountable for the outcome.”
Goetz described human-in-the-loop as the appropriate model for regulatory compliance today. That approach allows AI to support evidence assembly, summarisation and investigation while leaving material decisions with compliance staff.
Controlled Model Lifecycle
Bloomberg develops, fine-tunes, tests and validates each AI policy through a controlled release process. It assesses performance using measures that include precision and recall. The policies do not learn or adapt after deployment. Their behaviour remains fixed until Bloomberg changes them through its development, training, evaluation and release process.
Bloomberg may use gap analysis, client feedback and controlled fine-tuning to address communication patterns that lack enough representation in the training data. Each policy includes documentation covering its scope, target risks and role within the workflow.
These controls provide information that clients can use to support AI governance, supervision and regulatory obligations. Clients remain responsible for reviewing each surveillance event and deciding whether it warrants further action.
Subscribe to our newsletter
