Employee communications surveillance (eComms & voice) for conduct risk

Banks, brokers, asset managers and energy traders must record staff communications about transactions. They must also detect and report suspected market abuse. Classic lexicon alerting ("keep this between us", "off the record") floods compliance teams with false positives. It also misses coded language, other languages and voice. AI surveillance transcribes calls, normalises every channel and classifies conversations by behaviour (insider dealing, collusion, front-running, mis-selling, moving the conversation off-channel). Reviewers get fewer alerts, each with its evidence. In the EU this is a regulated duty. It is also monitoring of employees, so GDPR, works-council co-determination and, very likely, the AI Act's high-risk regime apply.

Problem & value

• Who: investment firms and credit institutions under MiFID II, persons "professionally arranging or executing transactions" under MAR Art. 16(2), commodity and energy traders (MAR/REMIT), and insurance distributors with advice-recording duties. In practice the target group is enterprise. Mid-sized asset managers and brokers buy the service. SMBs outside regulated finance have no legal driver, and the legal limits on monitoring employees make the use case hard to justify for them.
• Today: communications are captured into a WORM archive. Lexicon/keyword rules generate alerts, and level-1 reviewers clear them, often offshore. Most alerts are noise from newsletters, disclaimers and banter. Voice is sampled, not reviewed, and foreign-language chat is barely covered.
• Regulatory pressure: under SYSC 10A the FCA requires firms to record relevant calls and electronic communications, keep them for five years (up to seven on request), and take reasonable steps to stop relevant communication on private devices that cannot be recorded [3]. In the US, the SEC charged 26 firms in September 2024 alone over off-channel communications recordkeeping failures, with combined penalties of more than USD 390 million [1]. That fine series made detecting "let's move to WhatsApp" a standard detection scenario.
• AI value: behaviour-based classifiers and LLM triage cut alert volumes. Vendor-reported figures include a large French bank reporting a 97 % reduction in false positives and a 0.15 % alert rate at an energy trading firm. Both are Shield customer testimonials and have not been independently verified [5]. AI also brings multilingual and voice coverage that lexicons cannot provide [4][5].

Variants

• Market-abuse eComms surveillance (core): insider dealing, unlawful disclosure, collusion and benchmark manipulation, front-running, spoofing chatter. Alerts are correlated with orders from ANOM-012.
• Voice surveillance: trader turrets, mobile recording, adviser calls. The pipeline runs ASR, diarization and code-switching handling before classification. It shares its stack with SPEECH-002.
• Off-channel and recordkeeping detection: spots attempts to move a conversation to unrecorded channels ("call my cell", "WhatsApp me") and gaps in capture coverage.
• Conduct and culture risk: mis-selling and suitability language in adviser calls (MiFID II/IDD, FCA Consumer Duty), and harassment or bullying in chat. Use this variant with great care in the EU; see the risks section.
• Conflicts of interest and information barriers: detecting communications across Chinese walls (private side to public side), linked to watch and restricted lists.
• Energy and commodities: REMIT (wholesale energy) and the commodity-derivative provisions of MAR, as used by utility trading desks.
• Not in scope: generic DLP and data exfiltration (SEC-017), insider-threat security analytics, and productivity or "engagement" monitoring of employees. That last one is not a lawful variant in most EU works-council settings and is deliberately excluded.

How it works

1. capture (email, chat, Teams, mobile, voice)
2. WORM archive (MiFID II / SEC 17a-4 retention)
3. normalisation: thread/conversation reconstruction, dedup, disclaimer & newsletter removal
4. voice: ASR + diarization + language ID (+ translation where needed)
5. policy scoping: who is in scope (role, desk, instruments, watch/restricted lists)
6. detection: encoder classifiers per risk + LLM reasoning over whole conversation (+ legacy lexicon as safety net)
7. correlation with trade/order alerts (ANOM-012)
8. alert with evidence spans & explanation
9. L1 reviewer
10. L2 investigation
11. STOR / disciplinary / close
12. feedback loop: reviewer decisions
13. model tuning & periodic model validation

• Classifiers: small fine-tuned encoders or small LLMs score each message or conversation per risk typology at high volume (millions of messages a day). A larger LLM then re-reads only the flagged conversations with context and writes a rationale. Behavox describes this split as agentic AI for open-ended investigation and deterministic, validated models for high-stakes decisions [4].
• Voice: domain vocabulary (tickers, desk slang, numbers, "yards", "bps") is the main source of ASR error. Custom vocabulary or fine-tuned ASR matters more here than the choice of LLM.
• Human-in-the-loop: every alert is decided by a trained compliance reviewer. The AI never makes a STOR filing or a disciplinary decision on its own. Escalations are confirmed by a second reviewer.

Data preparation

The data is the archived communication of employees: email (Exchange, journaling archives), chat (Teams, Bloomberg, WhatsApp via capture tools), voice recordings from trading turrets and mobile recording, plus trade, order and employee master data to give context. Typical gaps are channels that are not captured (off-channel communication), inconsistent archives and formats, poor audio quality, many languages and slang, missing speaker identification, no labelled alerts from past cases, and access rights that are too wide.

Preparation steps: - Inventory all channels and close capture gaps (PREP-001, PREP-002, PREP-005). - Convert legacy mail archives and audio formats (PREP-007). - Harvest labels from past alerts and closed cases, and build a test set with compliance experts (PREP-020, PREP-018). - Pseudonymise content for review and model development (PREP-022). - Define data classes and access rules, and check purpose and legal basis (PREP-024, PREP-025, PREP-023).

Tools: existing archive and capture systems, speech-to-text such as Whisper, a pseudonymisation pipeline. Data work is roughly 40 to 50 percent of the effort, mainly capture, labelling and legal clearance. See GUIDE-003.

Implementation paths

Which path for whom

• Enterprise: Path B, because surveillance touches highly sensitive employee and client data, must be tailored to the firm's own channels and policies, and a shared EU-hosted AI platform with own models keeps data and models under control; buying a specialist suite (Path A) remains a valid alternative where validated vendor models are required.
• Developer: Path C, because vendors and in-house teams building surveillance components need open models and self-hosted classifiers to keep communications data inside their own perimeter.

Path A — Buy (valid alternative)

• Specialist vendors: Behavox Quantum (voice and text in 50+ languages, models validated against SR 11-7) [4], Shield (multilayered AI alert filtering) [5], Global Relay (archive plus AI surveillance), NICE Actimize SURVEIL-X (holistic trade and communications surveillance) [6], Smarsh (archive and surveillance), Theta Lake (collaboration and video-meeting compliance), and Verint / NICE for voice recording. EU/UK-relevant: Behavox and Shield both have large European operations. Check EU data residency, on-prem or private-cloud options, and whether the vendor supports a works-council configuration that disables everything outside the agreed scope.
• Microsoft Purview Communication Compliance: a good fit for firms on Microsoft 365. It provides trainable classifiers for Teams, Exchange and Viva Engage with pseudonymised review by default. It does not replace specialist market-abuse coverage on Bloomberg chat or voice.
• Choose A when: you are a regulated firm. Regulators expect validated, documented models and a mature case-management workflow, and both are expensive to build.

Path B — Build on cloud APIs (recommended)

• Models: EU-region deployments of Azure OpenAI, Amazon Bedrock (Frankfurt) or Google Vertex AI (EU), or Mistral (EU provider) for LLM triage of pre-filtered conversations. Use Azure AI Speech, AWS Transcribe or Google Chirp in EU regions for voice. Use a zero-data-retention configuration and contractual exclusion from training.
• Typical role: a second-stage triage layer on top of an existing lexicon or vendor alert stream. It explains and down-ranks alerts, and the archive, scoping and case management stay with the incumbent tool.
• Caveat: many banks forbid sending raw communications of in-scope staff to external LLM APIs. In that case Path B is limited to private endpoints in the firm's own cloud tenant.

Path C — Self-host (open source)

• ASR & diarization: Whisper large-v3 / large-v3-turbo via faster-whisper or WhisperX, pyannote for diarization, and NVIDIA Parakeet/Canary for speed.
• Classification: fine-tuned multilingual encoders (XLM-RoBERTa, ModernBERT-class) per typology. Open-weight LLMs such as Qwen3, Mistral Small or Llama 3.3 70B handle reasoning over flagged conversations and are served with vLLM. Llama Guard / Presidio cover PII handling and pseudonymisation.
• Search & review: OpenSearch/Elasticsearch with hybrid vector search for investigations. Case management is a custom build or an existing GRC tool.
• Hardware: HW3 (one 48–96 GB GPU) covers ASR plus a ~30B LLM for triage at mid-sized volumes. Large banks with continuous voice need several HW3 nodes or HW4.
• Ops effort: high. You need model validation (SR 11-7-style model risk management), drift monitoring and a regulator-ready audit trail. Self-hosting is attractive mainly for strict data-sovereignty or on-prem mandates.

Path D — Custom training

Worth it for large firms and vendors. Fine-tune classifiers on historical reviewer decisions, which can mean hundreds of thousands of labelled alerts, and adapt ASR to desk jargon and accents. True positives are extremely rare, so supplement them with synthetic scenarios written by compliance experts and red-team role-plays. Labels must come from reviewers, not from HR outcomes, to avoid encoding bias.

State of the art (2026)

• From lexicon to behavioural/LLM detection: leading vendors now combine high-recall classifiers with LLM-based conversation reasoning and generated rationales. Shield markets a GenAI toolkit (AmplifAI) for "agentic compliance" [5]. Behavox frames AI agents for investigation alongside deterministic validated models [4].
• Voice and multilinguality are the main differentiators. Vendors claim coverage of dozens of languages [4][5]. Code-switching (for example German-English trader slang) and noisy turret audio remain hard.
• No public benchmark exists for communications-surveillance accuracy. Claims are vendor-reported, so firms should run a parallel test on their own historic alerts with known outcomes.
• Failure modes: coded language that is genuinely novel, sarcasm and banter, images and screenshots in chats, encrypted or unmonitored channels (the off-channel problem moves rather than disappears), LLM rationales that sound convincing but are unfounded, and over-monitoring that exceeds the agreed legal scope.

Shared resources & combinations

• Speech stack (asr + diarization): shared with SPEECH-002 (contact-centre QA), SPEECH-001 (meeting transcription) and SPEECH-010 (speaker identification for attributing recorded speakers).
• Text-classifier + LLM triage stack: shared with LANG-012 (routing), LANG-014 (moderation), LANG-015 (entity extraction for tickers and counterparties) and LANG-009 (translation of foreign-language chats).
• Archive, search & review: the same corpus and review tooling as DOC-021 (eDiscovery, investigations, DSARs), and DOC-018 for pseudonymised review and redacted disclosure.
• Correlation: ANOM-012 (trade surveillance) is the natural partner. A holistic alert joins a suspicious order with the chat that preceded it. AGENT-018 (controls monitoring) and SEC-022 (AI governance inventory) cover model governance.
• Hardware: one HW3 GPU node can host ASR, a classifier and a mid-size LLM for this use case together with SPEECH-002 and DOC-021 workloads.

Effort, cost & KPIs

• Path A: one to three quarters including connectors, scoping, parallel run against the legacy system, works-council agreement and model validation. Licence costs depend on the number of monitored users and voice hours.
• Path B/C: six to twelve months to reach a regulator-defensible state. The main cost drivers are voice transcription volume, reviewer FTEs and model-validation effort.
• KPIs: alert volume per monitored user, false-positive rate, the share of alerts escalated to L2 or a STOR, recall on seeded or test scenarios, channel and voice coverage, review backlog and time-to-decision, and regulatory findings.

People & change

This is one of the most employee-sensitive use cases, so involve people early and in this order: management and the head of compliance (mandate), the works council and the data protection officer (before any pilot), IT-security and IT, legal/compliance and HR (disciplinary consequences, employment contracts). Employees are the monitored group and must be informed transparently.

Works council: yes. A system that analyses employee communication is objectively suited to monitor behaviour and performance, so §87(1) no. 6 BetrVG applies (BAG case law), and §90(1) no. 3 requires timely information on AI planning. Engaging an AI expert counts as necessary under §80(3). Conclude a works agreement on scope, retention, pseudonymised first review, access roles and consequences (GUIDE-004). Public bodies involve the staff council.

Data protection: the DPO is mandatory in practice, a DPIA under Art. 35 GDPR is needed, and the legal basis and proportionality must be documented (GUIDE-005). Under the AI Act this is an Annex III employment-related high-risk case, with duties from 2 Dec 2027 (GUIDE-014).

AI literacy: support compliance reviewers with training on false positives and bias (GUIDE-006). Pilot with one desk, keep human reviewers as decision makers, never automate sanctions, and communicate what is and is not analysed (GUIDE-008).

Risks, regulation & compliance

• EU AI Act: likely high-risk. Annex III point 4(b) covers AI intended "to monitor and evaluate the performance and behaviour of persons in [work-related] relationships" [2]. Conduct surveillance of employees fits this wording, so plan for provider and deployer obligations: risk management, logging, human oversight, and informing workers and their representatives before use (Art. 26(7)). Some firms argue that a narrow market-abuse filter that only routes content to human review is not "evaluating behaviour", but get a documented legal assessment before relying on that. Art. 5(1)(f) prohibits emotion recognition in the workplace. Voice-based stress or emotion scoring of traders therefore must not be deployed in the EU, so switch off such vendor features. Text-based risk classification is not emotion recognition based on biometric data.
• GDPR: a DPIA is mandatory (systematic monitoring). The legal basis is a legal obligation (MAR, MiFID II) for the regulated scope, and only that scope. Anything beyond it, such as culture or harassment analytics, needs a separate justification and a necessity and proportionality test. WP29 Opinion 2/2017 on data processing at work sets the baseline [8]. CJEU C-34/21 (2023) put national employee-data rules like §26 BDSG under scrutiny, so Art. 6 GDPR must carry the processing. Private content and special-category data will appear. Use filtering, pseudonymised L1 review, strict retention and access logging.
• Works council (DE): §87(1) Nr. 6 BetrVG gives the works council co-determination over technical systems suitable for monitoring employee behaviour or performance [7]. A works agreement (Betriebsvereinbarung) defining scope, typologies, retention, purpose limitation and a ban on using the data for performance management is effectively a prerequisite. Austria, France and the Netherlands have similar regimes.
• Sector law: MAR Art. 16 (detection and STORs) [9], MiFID II Art. 16(7) recording [10], FCA SYSC 10A [3], REMIT, DORA (ICT third-party risk for the surveillance vendor), and for US entities the SEC 17a-4 / FINRA rules.
• Mitigations: narrow and documented scenario scope, four-eyes review, no automated sanctions, regular bias and accuracy testing across languages and accents, explainable evidence spans, and employee transparency notices.

Getting started

1. Map the legal obligation: which staff, channels and instruments are in scope under MAR/MiFID II/REMIT, and which channels are not captured today.
2. Run a DPIA and an AI Act classification with the DPO and legal. Open works-council negotiations early, because they usually set the timeline.
3. Pull 6–12 months of historical alerts with known outcomes as an evaluation set. Run 2–3 vendors in parallel against it and your live lexicon.
4. Prioritise voice and foreign-language coverage and off-channel detection, which are the biggest current gaps.
5. Establish model validation, a reviewer feedback loop and MI on false positives and coverage before decommissioning lexicons.

Sources

1. SEC press release 2024-98 — Twenty-six firms to pay more than $390 million combined to settle off-channel recordkeeping charges — https://www.sec.gov/newsroom/press-releases/2024-98
2. EU AI Act, Annex III (high-risk use cases, point 4 employment) — https://artificialintelligenceact.eu/annex/3/
3. FCA Handbook SYSC 10A — Recording telephone conversations and electronic communications — https://www.handbook.fca.org.uk/handbook/SYSC/10A/1.html
4. Behavox — product overview (Quantum communication surveillance, 50+ languages, validated models) — https://www.behavox.com/
5. Shield — surveillance platform overview and customer testimonials (vendor-reported figures) — https://www.shieldfc.com/
6. NICE Actimize — SURVEIL-X Holistic Trade Compliance — https://www.niceactimize.com/
7. Betriebsverfassungsgesetz §87 (co-determination, Nr. 6 technical monitoring systems) — https://www.gesetze-im-internet.de/betrvg/__87.html
8. Article 29 Working Party — Opinion 2/2017 on data processing at work (WP249) — https://ec.europa.eu/newsroom/article29/items/610169
9. Regulation (EU) No 596/2014 (Market Abuse Regulation) — https://eur-lex.europa.eu/eli/reg/2014/596/oj
10. Directive 2014/65/EU (MiFID II) — https://eur-lex.europa.eu/eli/dir/2014/65/oj
11. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj