What problem does it solve?
Banks, brokers and trading venues must detect and report suspicious orders and transactions. Most surveillance still runs on rules: an order that exceeds a size, moves a price or matches a pattern raises an alert, and an analyst has to reconstruct what happened before deciding whether it is worth a closer look. That reconstruction is the expensive part: pulling the order book, related trades, the issuer's filings, news around the event and, for conduct cases, the trader's emails and chats.
Rules produce large volumes of false positives, and subtle manipulation that spans venues, instruments or time zones does not match a single rule. Supervisors also expect firms to prove that their surveillance works: in Market Watch 79 the FCA described alert scenarios that failed unnoticed, in one case for over three years, because of faulty alert logic or data that was never ingested. The result is a team that spends much of its time closing noise, with the hard cases getting less attention than they deserve.
- 1LoD's 2026 Surveillance Benchmarking Survey found that 89% of banks want AI enhanced trade surveillance but only 11% have it, and that 78% want generative AI assistants for analysts while 7% have deployed one.Banks want AI surveillance but lack the data to run it (2026)
- The same 1LoD survey reports that 93% of banks rate false positives a meaningful drag on surveillance and 52% call them a high challenge, which the survey attributes to fragmented data capture and ageing platforms upstream of the alert stage.Banks want AI surveillance but lack the data to run it (2026)
How does it work?
- Alert in. The existing surveillance system (rules or models) raises an alert on an order pattern, a trade ahead of a price move or a flagged message.
- Assemble the context. An agent pulls the relevant orders and trades, the instrument's price and volume around the event, the issuer's filings and news, the trader's history and prior alerts, and for conduct cases the linked communications.
- Explain the trigger. The AI states in plain language which behaviour set off the alert and which facts make it more or less suspicious, with a link to each underlying record.
- Score and route. Alerts are ranked by likely risk; clear false positives are proposed for closure with a reason, and the rest go to an analyst queue.
- Draft the case. For alerts that go further, the AI drafts the investigation narrative and, where needed, the first version of a suspicious transaction and order report.
- Human disposition. An analyst reviews, edits and decides every alert. Their decisions and reasons are logged and feed back into tuning.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Early adopters
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Handling time reduction | Too few to pool | about 33% | 1 | 1 organization |
Value drivers: Compliance quality, Employee productivity, Risk and loss reduction, Speed and cycle time.
Indicative value
A bank with a markets business raising 40,000 surveillance alerts a year
USD 180,000 to USD 1.2 million
Analyst time released from first line alert review per year
How this is calculated
Formula: alerts * hoursPerAlert * timeSaved * hourlyCost. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Surveillance alerts reviewed per year alerts, alerts per year | 40,000 | 40,000 | The reference bank. Replace with your own alert volume across trade and communications surveillance. |
| Analyst hours per alert at first review hoursPerAlert, hours per alert | 0.5 | 1 | Editorial assumption for gathering evidence and writing the first assessment. Replace with your own time study. |
| Share of review time saved timeSaved, fraction of review time | 0.15 | 0.3 | Conservative against the benchmark on this page (in Nasdaq's proof of concept testing, surveillance analysts estimated a 33% reduction in investigation time). |
| Fully loaded cost of a surveillance analyst hourlyCost, USD per hour | 60 | 100 | Editorial assumption, replace with your own. |
What it leaves out: Counts analyst time only. It leaves out the cost of the AI and data work, any change in the number of alerts, and the value of detecting abuse that rules miss, which is the larger prize but hard to price.
Who already uses it?
5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Nasdaq
United States · Capital markets · 2024
Nasdaq added a generative AI feature, built on Amazon Bedrock, to the market surveillance technology it runs for regulators and marketplaces. When an alert fires, the feature gathers and condenses the evidence an analyst needs for the initial assessment, such as a table of the company's regulatory filings, news summaries and sentiment and other mitigating or aggravating factors. The reported gain comes from proof of concept testing, in which analysts estimated the investigation time saved; Nasdaq said it planned to use the feature for its own US equity market surveillance.
- Handling time reduction: about 33%, proof of concept testing, estimated by surveillance analysts
"During proof-of-concept testing, surveillance analysts estimated a 33% reduction in investigation time, with improved overall outcomes."
Claimed by: organization
Commodity Futures Trading Commission
United States · Government and public sector · 2023
The CFTC Division of Enforcement ran a pilot in 2023 that applied supervised and unsupervised machine learning to order message data to find spoofing patterns, trained with the division's existing expert based spoofing detection algorithms. The model gave a probability of spoofing behaviour for every trader in a given market on a given day. The 2024 federal AI inventory lists the project as retired in April 2024; no results were published.
No outcome disclosed.
Japan Exchange Group
Japan · Capital markets · 2018
Japan Exchange Regulation and the Tokyo Stock Exchange put two machine learning systems from NEC and Hitachi into their market surveillance operations on 19 March 2018. The systems were supplied with the knowledge surveillance staff had used to evaluate irregular trading, and help staff finish the preliminary investigation of orders flagged by the criteria based surveillance systems faster, so they can focus on detailed investigations. The decision whether to investigate further stays with surveillance personnel. No outcome figures were published.
No outcome disclosed.
U.S. Securities and Exchange Commission
United States · Government and public sector · 2018
The SEC Division of Enforcement uses a classical machine learning tool, operational since April 2018, to identify accounts whose trading came in advance of material equity price moves and that warrant further investigation. The output is a list of leads for enforcement staff, who decide what to investigate. The entry appears in the 2025 US federal AI use case inventory; no outcome figures are published.
No outcome disclosed.
Deutsche Bank
Germany · Banking · 2026
According to Bloomberg reporting relayed by trade press in February 2026, Deutsche Bank is working with Google Cloud on AI agents that monitor trading, spot anomalies in orders, trades and market moves and flag them to a human compliance officer, and plans to use the same approach on the communications of client facing staff such as traders and salespeople. AI News adds that Goldman Sachs is exploring agentic surveillance too and that human compliance staff remain responsible for reviewing flagged cases. No bank statement or outcome figure was found.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Complete order and trade records across venues and asset classes, reconciled to source
- Communications records (email, chat, voice transcripts) held in original form and linked to traders
- Reference data, issuer filings and a news feed with timestamps
- A labelled history of past alert dispositions with reasons
Systems to integrate
- Existing trade and communications surveillance platforms (alert source)
- Order management and execution systems, market data
- Communications archive and voice recording platform
- Case management for investigations and suspicious transaction and order reports
- Model inventory and model risk management tooling
Complexity: High
The AI layer is the easier part. The hard work is complete, reconciled trade, order and communications data with a clear chain of custody, and model governance that a supervisor will accept for a system that clears alerts.
- 1
Start with explanation, not auto closure
First give analysts a plain language summary and an evidence pack for every alert. Measure review time and analyst agreement before letting the system propose closures.
- 2
Fix the data before the model
Reconcile order and trade feeds to source and check that every business line and venue is actually monitored. An AI layer on incomplete data reprocesses the same gaps faster.
- 3
Define the closure policy
Write down which alert types may be proposed for closure, the evidence required, the sampling rate for quality checks and who signs off the policy.
- 4
Validate like any surveillance model
Put the triage model through model validation: back testing on past alerts including known true cases, stability over time and a documented explanation of its logic.
- 5
Add communications and cross product views
Once trade triage is trusted, link trade alerts to communications and to related instruments, where rules alone miss the most.
Guardrails
- A human analyst dispositions every alert; the AI proposes, it never closes on its own
- Every explanation links to the underlying orders, trades and messages it relies on
- Random quality sampling of alerts the AI proposed to close, with results reported to compliance
- Access to communications data limited by role and logged, with personal data minimised in prompts
- Change control and regression tests on known true positive cases for every model or prompt change
KPIs to instrument
- Median review time per alert, by alert type
- Share of alerts proposed for closure and the analyst agreement rate
- True positives found per period, including cases the rules did not flag first
- Quality sample findings on closed alerts
- Share of alerts with a complete evidence pack
Human in the loop
Surveillance analysts own every disposition and every escalation to a suspicious transaction and order report. Second line compliance approves the closure policy and reviews quality samples, and model validation signs off the triage model before it goes live and after changes.
Common failure modes
- Confident explanations on missing data
- The AI writes a fluent rationale while part of the order flow was never ingested. Check data completeness per venue and show gaps in the evidence pack.
- Automation bias
- Analysts accept the suggested disposition without reading the evidence. Track agreement rates, rotate blind reviews and sample closures.
- Tuning away real abuse
- Optimising for fewer alerts lowers detection. Always back test on known true cases and report detection alongside false positives.
- Unexplainable to the supervisor
- A model that cannot show why it cleared an alert fails regulatory scrutiny. Keep the reasoning and evidence for every alert.
What are the risks and rules?
EU AI Act
Depends on design
Surveillance of orders and transactions as such is not listed in Annex III. Where the system monitors and evaluates the behaviour of the firm's own staff, in their communications or their trading, it can fall under Annex III point 4(b) (AI used to monitor and evaluate the performance and behaviour of persons in work relationships), so the tier depends on whether the system scores individual employees. Inferring employees' emotions from biometric data such as voice recordings is prohibited in the workplace under Article 5(1)(f).
Rules that apply
Guidance
- Market Abuse Regulation (EU) No 596/2014 (European Union, Europe). Article 16 requires firms that arrange or execute transactions to have effective arrangements, systems and procedures to detect and report suspicious orders and transactions.
- Commission Delegated Regulation (EU) 2016/957 (European Union, Europe). Technical standards for detecting and reporting suspicious orders and transactions, including the duty to keep for five years the analysis of each examined order or transaction and the reasons for submitting or not submitting a STOR.
- Market Watch 79 (Financial Conduct Authority, Europe). FCA examples of surveillance failures caused by data ingestion and alert logic issues, and its 2023 peer review of how 9 investment banks test automated surveillance models under UK MAR.
- Artificial Intelligence (AI) Model Risk Management (Monetary Authority of Singapore, Asia Pacific). Good practices MAS observed in its 2024 thematic review of banks' AI and generative AI model risk management, covering governance, oversight, development and deployment; relevant when validating a triage model.
Controls to put in place
- Surveillance model and triage AI registered in the model inventory with an owner and validation status
- Documented closure policy approved by compliance, with sampling of AI assisted closures
- Data completeness checks per venue, asset class and communications channel
- Full audit trail of alert, evidence, AI output, analyst decision and reason
- Periodic back testing against known true positive cases
Frequently asked questions
- Can AI close market abuse alerts on its own?
- It should not. The defensible pattern is that the AI assembles evidence, explains the trigger and proposes a disposition, and a named analyst decides. In the EU, Commission Delegated Regulation 2016/957 requires firms to keep, for five years, the analysis of every examined order or transaction and the reasons for reporting it or not, so each closure needs a documented rationale.
- How much analyst time does AI triage save?
- Public figures are still few and early. Nasdaq reported that surveillance analysts estimated a 33% reduction in investigation time during proof of concept testing of its generative AI feature. Treat that as an estimate from a pilot and measure your own review times per alert type.
- Is AI surveillance high risk under the EU AI Act?
- Surveillance of client orders and transactions generally is not. Monitoring and evaluating the behaviour of the firm's own employees, in their communications or their trading, can fall under Annex III point 4(b), so a design that scores individual staff needs the high risk controls.
- What is the biggest obstacle?
- Data. In 1LoD's 2026 Surveillance Benchmarking Survey, 71% of answers on what most hinders surveillance pointed to fragmented, non standardised or poor quality data, and the report says this leaves many AI projects stuck at proof of concept.
How to cite this page
Blits.ai AI Use Case Library, "AI for market abuse surveillance alert triage", last verified 26 September 2026, https://www.blits.ai/ai-use-cases/market-abuse-surveillance-triage. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published