What problem does it solve?
Due diligence requires banks to know whether a customer, a director or a beneficial owner is a politically exposed person or has been linked to crime, corruption or other serious wrongdoing. Curated databases cover only part of the world's news, and keyword searches on the open web return pages of irrelevant hits: people with the same name, old stories, opinion pieces.
Analysts read article after article to rule out namesakes, often in languages they do not speak, and the result is inconsistent. Real risk gets missed in the noise, while onboarding and periodic reviews slow down. The quality of the written conclusion, why a hit was or was not relevant, is what auditors check, and it often varies between analysts.
How does it work?
- Search broadly. For each subject the system queries curated risk databases, news archives, court and regulatory records and the open web, in the languages that match the subject's footprint.
- Disambiguate. Entity resolution compares each article's person or company with the subject's known attributes (age, location, occupation, associated companies) and discards look alikes with a stated reason.
- Classify the risk. Relevant articles are classified by risk category (fraud, corruption, sanctions evasion, organised crime) and by credibility and recency of the source.
- Summarise with citations. The system writes a short summary of the credible findings, translated where needed, with a link to every source article.
- Analyst decides. The analyst confirms relevance and source reliability, records the disposition and decides whether it changes the customer's risk rating; monitoring continues between reviews.
- 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 | at least 60% Not pooled: up to 50% | 1plus 1 up to | 1 vendor |
| False positive reduction | Too few to pool | 95% | 1 | 1 vendor |
Value drivers: Compliance quality, Employee productivity, Speed and cycle time, Risk and loss reduction.
Indicative value
A bank running 40,000 adverse media reviews a year across onboarding and periodic reviews
USD 105,000 to USD 800,000
Analyst capacity released per year
How this is calculated
Formula: reviews * minutesPerReview / 60 * timeSaved * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Adverse media reviews per year reviews, reviews per year | 40,000 | 40,000 | The reference bank. |
| Analyst minutes per review today minutesPerReview, minutes per review | 15 | 40 | Editorial assumption. Replace with your own time study. |
| Share of review time saved timeSaved, fraction of review time | 0.3 | 0.5 | Conservative against the benchmark on this page (Xapien reports that Save the Children cut donor due diligence review times by over 60% with its AI due diligence tool). Replace with results from your own pilot. |
| Fully loaded analyst cost per hour costPerHour, USD per hour | 35 | 60 | Editorial assumption. Replace with your own. |
What it leaves out: Counts analyst time only. It leaves out faster onboarding, risk found that manual searches missed, data licence costs and the cost of the platform.
Who already uses it?
7 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
OCBC
Singapore · Banking · 2026
OCBC launched HELIOS in July 2026, an agentic AI platform that gathers intelligence on prospective private banking clients and completes most of the customer due diligence before a relationship manager engages them. OCBC says private banking accounts can now be opened in 15 business days, against an industry median of about six weeks, while relationship managers and review teams keep accountability for judgment and decisions. OCBC plans to extend HELIOS to ongoing monitoring of customer activity to detect changes in risk profiles. Bank of Singapore relationship managers use it in Singapore, Hong Kong and Dubai, with the rollout due to finish in the third quarter of 2026.
No outcome disclosed.
Save the Children
United Kingdom · Cross industry · 2026
Save the Children uses Xapien, an AI supported due diligence platform that produces a report on a prospective donor and surfaces areas of concern early in the report, to vet corporate donors for alignment with its values and for reputational risk. The vendor reports that review times fell by more than 60%, with reports completed in as little as 15 minutes rather than over an afternoon, so the team can vet more donors. The platform is one part of a wider, human led review. It shows the same adverse media job outside banking.
- Handling time reduction: at least 60%, analyst review time per corporate donor
"Save the Children uses Xapien to accelerate corporate donor due diligence, cutting review times by over 60%."
Claimed by: vendor
HSBC
United Kingdom · Banking · 2024
Silent Eight has supplied HSBC with automation for name screening and adverse media alerts, and in February 2024 the two expanded the partnership to automated alert closure for transaction screening, which investigates and resolves payment screening alerts in real time. No outcome figures are disclosed.
No outcome disclosed.
Mashreq
United Arab Emirates · Banking · 2024
Mashreq selected Silent Eight in May 2024 to automate the adjudication of name screening and adverse media alerts related to sanctions and anti money laundering requirements. Under the plan, false positives are to be investigated and closed quickly and potential true positives escalated to Mashreq analysts. The announcement is a multi year partnership; no results are disclosed.
No outcome disclosed.
Deutsche Bank
Germany · Banking · 2020
Deutsche Bank used WorkFusion's AI automation for screening work in anti money laundering, including adverse media monitoring and PEP checks for new accounts and refresh screenings, which had required large teams to scan news reports manually. For the KYC programme as a whole, the vendor reports shorter handling times, about 25,000 cases handled per quarter and tens of thousands of hours saved each year.
- Handling time reduction: up to 50%, range of 25 to 50%, across the whole KYC programme (screening and document processing)
"25–50% reduction in handling time"
Claimed by: vendor
Scotiabank
Canada · Banking · 2020
Scotiabank automated its adverse media monitoring (negative news search) for anti money laundering with WorkFusion, applying the vendor's intelligent automation to the analysis and disposition of adverse media. The vendor reports a sharp fall in false positives, wider media search coverage (30 articles per name instead of 20) and the equivalent of more than a hundred compliance analysts freed for other work.
- False positive reduction: 95%
"95% reduction in false positives"
Claimed by: vendor
Santander UK
United Kingdom · Banking · 2019
Santander UK used ComplyAdvantage's adverse media screening, delivered through an API, as part of a digital onboarding proposition for corporate and SME customers, and screens every entity linked to an onboarding case. The vendor reports that the onboarding cycle fell from 12 days to 2 days on average; the figure covers the whole onboarding process, not the screening step alone.
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
- Subject attributes for disambiguation (date of birth, nationality, addresses, related companies)
- Licensed news and risk data sources, plus rules for which open web sources count as credible
- The bank's adverse media risk taxonomy and materiality criteria
- Historical dispositions to measure false positive rates
Systems to integrate
- KYC and customer due diligence system
- Screening engine and PEP database
- News and risk data providers
- Case management and customer risk rating
Complexity: Medium
Retrieval and summarisation are mature. The hard parts are reliable disambiguation of common names, licensed access to news content, and keeping the analyst accountable for the conclusion.
- 1
Define what counts as adverse
Write down the risk categories, how old a story may be, and which sources count as credible, with examples. The model can only be as consistent as the policy.
- 2
Get disambiguation right
Measure how often the system wrongly matches or wrongly discards a namesake on a labelled sample, per language and naming culture, before analysts rely on it.
- 3
Summaries with sources, never without
Require a link to every source in the summary and reject any claim that is not supported by a retrieved article.
- 4
Pilot on periodic reviews
Start with periodic reviews of existing customers, where time pressure is lower, then extend to onboarding and continuous monitoring.
Guardrails
- Every finding links to its source; unsupported statements are rejected
- Adverse media changes a risk rating only after an analyst confirms relevance and reliability
- Bias testing across names, nationalities and languages for both false hits and misses
- Source articles and dispositions retained for audit
- Licence terms respected for every news source
KPIs to instrument
- Hits per subject presented to analysts, before and after
- Share of analyst overturned discards and matches on a labelled sample
- Review time per subject
- Material findings per thousand reviews
- Miss rate on a known test set of adverse subjects
Human in the loop
The system searches, filters and summarises; the analyst decides whether a finding is about the subject, whether it is credible and whether it matters. Changes to risk rating or relationship decisions stay with named people.
Common failure modes
- Namesake contamination
- A common name links a customer to someone else's crimes. Require multiple matching attributes and show them in the summary.
- Language and culture bias
- Disambiguation works well for some naming conventions and badly for others. Measure performance per language and naming culture.
- Summary replaces reading
- Analysts stop opening sources. Sample decisions against the source articles and keep the analyst's conclusion in their own words.
What are the risks and rules?
EU AI Act
Minimal risk
Adverse media and PEP screening for due diligence is not listed in Annex III. It processes personal data, including data about alleged offences, so GDPR Article 10 and national AML law govern what may be collected and how long it is kept.
Rules that apply
Guidance
- Principles for Using Artificial Intelligence and Machine Learning in Financial Crime Compliance (Wolfsberg Group, Global). Industry principles for legitimate, proportionate and transparent use of AI in financial crime compliance.
- Notice 626 Prevention of Money Laundering and Countering the Financing of Terrorism, Banks (Monetary Authority of Singapore, Asia Pacific). Example of national rules on customer due diligence, PEP checks and ongoing monitoring.
Controls to put in place
- Written adverse media policy with risk categories, recency and source credibility rules
- Source links and analyst disposition retained for every finding
- Bias and accuracy testing per language and naming culture
- Human decision on any change to a risk rating or relationship
Frequently asked questions
- Is adverse media proof of risk?
- No. It is an input to a risk decision. The analyst checks that the article is about the customer, that the source is credible and that the allegation is material before it changes a rating.
- How does AI reduce adverse media false positives?
- Mostly through disambiguation: comparing ages, locations, occupations and related companies in the article with what the bank knows about the customer, and discarding namesakes with a stated reason. Measure it on a labelled sample in every language you screen.
- What about bias against certain names?
- It is a real risk. Names that are common in a community, or that are transliterated from another script in several ways, can produce more false hits and so more manual scrutiny for some customers. Test false hit and miss rates per naming culture and language, and fix the gaps before scaling.
How to cite this page
Blits.ai AI Use Case Library, "AI for PEP and adverse media screening", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/pep-and-adverse-media-screening. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published