Adverse media screening that turns a pile of articles into one cited story, decided by two people
A pile of articles becomes one cited story, decided in minutes.
See one case, screen by screen ↓Adverse media screening is the check for negative news about a customer (fraud, corruption, money laundering, sanctions evasion, tax crime) before onboarding a higher-risk customer, at an EDD refresh or after a trigger event. Adverse Media Screening puts eight agents in front of that check: they search every language tied to the customer, group the copies into stories, try to prove each match wrong and cite the original sentence. A KYC analyst confirms and an EDD reviewer accepts.
AI use cases in banking: a buyer's guide to agentic AI for bank and credit union operations →
Updated
Adverse media screening comes before a bank onboards a higher-risk customer, at every enhanced due diligence (EDD) refresh and after a trigger event: a KYC analyst must find out whether the news says anything the bank needs to know — fraud, corruption, money laundering, sanctions evasion, tax crime. Today that means a negative news search string typed into a search engine or a news database, and dozens of results: most of them copies of one report, some about someone else with the same name, some in a language the analyst does not read.
Telling a namesake from the customer means reading for an age, a job and a city, and doing the arithmetic, because an article from 2023 gives the age the person was then. A Portuguese report gets pasted into a translator, losing the wording the file should cite, or it is skipped. An allegation is weighed like a conviction unless someone reads the whole story, and ongoing monitoring re-alerts on old news. Examiners want to see what was searched, in which languages, what was set aside and why, and who decided. A screenshot of a search page is not that.
Estimated split of analyst minutes for one review with a relevant story, by hand and with the solution; the 6-minute total is the design target.
Eight agents search every language, group the copies, try to prove each match wrong and draft the file; an analyst confirms and a reviewer accepts.
Corvane National Bank, Thursday, October 22, 2026. Review AMR-26-104387 is an EDD refresh for Rafael Monteiro Duarte: a mechanical engineer in Houston, a US and Brazilian citizen, director and 30% beneficial owner of a high-risk commercial client, Aldrava Process Systems, and director of its subsidiary in Goiânia. Here is what happens next, screen by screen, in the working solution.
1,186 adverse media reviews are open across the team, 46 of them Marisol's. Overnight the agents opened 548 reviews, read 18,416 articles and grouped them into 2,904 stories: 391 found nothing about the customer, 142 need a look and 15 are proposed relevant. Each review shows articles found, stories and relevant at a glance. Rafael Monteiro Duarte's reads 37 → 4 → 1.
“The adverse media service did not answer from 5:40 – 6:05 AM. 22 searches were held and run again at 6:06 AM; no review lost its clock, and one is still being grouped.”
The Search agent built the plan from the customer record: six forms of the name; English; Portuguese because of his citizenship, the subsidiary in Goiânia and 41 wires to Brazil; and Spanish because of 14 wires to a supplier in Monterrey. Risk terms are searched per category in each language, such as bid rigging, licitação and licitación amañada. Five sources, ten years back: 37 articles.
The Search agent never searches outside the bank's approved sources.
The review opens on 37 article tiles, each with its outlet, language and date. The Story agent collapses them into 4 stacks, one per story. The Different-person agent's result dims two stories about other men and one that is harmless company news, and 1 relevant story is left. Counters across the top move between the three levels, and the agents' steps are listed with their confidence and the hand-off to Marisol.
4 stories: 33 copies and follow-ups grouped. 1 relevant story: 2 other men, 1 not adverse.
Story A, in Portuguese, 14 copies, is about a procurement investigation. Story B, 11 copies, is about a footballer of 27 in Porto. Story C, 8 copies in English, is about a Miami real estate broker of 61 and escrow funds. Story D, 4 copies, is about Aldrava Process Systems winning a Gulf Coast water pump contract. Each shows the headline in English and as published, the category and stage, and the identity evidence.
The Story agent never drops an article: every copy stays listed under its story.
The customer record says born June 11, 1973, mechanical engineer, director of Aldrava do Brasil in Goiânia. Story A names “Rafael M. Duarte, 52”, published September 18, 2025: the customer was 52 that day. Age, occupation, company, role and city agree; the name is partial. Story B's footballer is 27 when the customer was 53, and story C's broker is 61 when he was 49.
Before anything is proposed, the Different-person agent argues the other side. For the footballer it worked out the customer's age on the day of publication, compared occupation and looked for any tie to Porto: no address, travel or payment to Portugal on file. Five identifiers conflict, so story B is set aside, with every test kept. On story A it ran four tests, including the Brazilian registry extract in the customer file.
“Could not show a different person: 5 identifiers agree, 1 is partial, none conflict.”
Diário do Vale do Cerrado, September 18, 2025: state prosecutors are investigating whether a R$38.6 million tender for industrial pumps was steered to one bidder; the company denies wrongdoing; no one has been charged. Pointing at a sentence lights it in both languages, at a translation confidence of 0.97. The Risk agent proposes relevant: corruption and bribery, public procurement, under investigation, at 0.93 against the 0.85 threshold.
“Entre os investigados está o engenheiro Rafael M. Duarte, 52 anos, diretor da Aldrava do Brasil, empresa com sede em Goiânia.”
Marisol confirms story A as relevant with reason AM-05: relevant, under investigation, no charge. Confirm stays disabled until a code is chosen, and rejecting the proposal asks for a note and records AM-07. She is the maker, so a different person must accept the story into the customer file.
The EDD writer drafts only from a story a person has confirmed.
The EDD writer drafts the paragraph from the confirmed story only, citing sentences 1, 2, 4 and 5 of the Portuguese report. Signed in as Marisol, Accept is locked and the file is handed to Desmond Hale. He accepts it at 10:05 AM and proposes the rating input: corruption weight 23 × 0.7 for an investigation = +16, Elevated 62 to High 78.
What the relationship manager sees: “Compliance review in progress. No action needed.”
Grace approves High 78, retain with conditions. Only then is the rating written to the client lifecycle system and risk code H to the core banking system, and a 90-day review of wires to Brazil referred to case management, visible to the BSA unit only. At 10:32 AM monitoring starts watching story A for a charging decision.
Signatures: Marisol Ocampo 9:16 AM (AM-05) · Desmond Hale 10:05 AM · Grace Whitlock 10:31 AM (High 78, retain with conditions).
The same review, seen by the four people who carry it: what their day looked like, and what it looks like now.
Eight agents search every language, group the copies, try to prove each match wrong and draft the file; an analyst confirms and a reviewer accepts.
Opens each review and runs the other agents in order (search, translation, grouping, matching, the challenge, classification), keeps the review clock and hands the sorted review to a named analyst.
Builds the search plan from the customer record: every form of the name, every language tied to the customer by citizenship, address, subsidiary or wire country, plus English, and risk terms for each category in each language.
Detects the language and translates each article sentence by sentence, aligned to the original, so a finding cites the sentence as published.
Groups syndicated, rewritten and translated copies and follow-ups into one story per event, keeping the earliest original as the source of record.
Compares six identifiers in each story (name, age, occupation, company, role and city) with the customer record, working out the customer's age on the day each article was published.
Tries to prove each matched story is about somebody else, using age on the day, occupation, company, city and the registry extracts in the customer file, and writes down every test.
Classifies each story by predicate offense and by stage, from allegation to conviction or acquittal, and proposes relevant or not with its confidence.
Drafts the EDD paragraph and the risk rating input from confirmed stories only, citing each sentence of the original.
Confirms with a reason; a reviewer accepts. The agents propose; a named person decides.
Ask about any customer, story or rule. The answers point to the sentence or the setting they come from.
Why is the Portuguese story about this customer?
Five of six identifiers agree. The article says engineer Rafael M. Duarte, 52, director of Aldrava do Brasil, in Goiânia. The customer was born on June 11, 1973, so he was 52 on September 18, 2025; he is a mechanical engineer and a director of that subsidiary. Only the middle name is partial. The Different-person agent ran four tests and could not show it is someone else.
Show me what the story actually alleges, in Portuguese
Prosecutors are investigating whether a R$38.6 million pump tender at the state water company was steered to one bidder, possibly using inside information. The company denies it and nobody has been charged, so the agents propose corruption and bribery, under investigation, not a conviction. Here is the passage beside its translation.
Draft the EDD paragraph for Aldrava Process Systems
The draft uses the confirmed story only, and every sentence cites the original: the R$38.6 million tender investigation, the director named, the company's denial and that no one is charged. The rating input is corruption, weight 23 × 0.7 for an investigation = +16, taking Elevated 62 to High 78. Nothing is written until Desmond Hale accepts it and Grace Whitlock approves the rating.
Confirm the 391 reviews where you found nothing
I cannot confirm them: confirming is a named analyst's decision. In all 391, every story was another person or not adverse, and each review keeps the tests that set it aside. I have opened the bulk confirmation with this program's 5% second look: 20 go to Priya Natarajan first.
13 screens from the working solution, on its sample data. Pick one to see it large.
What the program decides for itself is a setting: versioned, changed only by the administrator or the compliance manager, and on the audit trail.
| Setting | Default | Choose from |
|---|---|---|
| Who accepts a story into the customer file | Any EDD reviewer, never the person who confirmed it | Any EDD reviewer · EDD reviewers and the compliance manager |
| Who approves a risk rating change | The compliance manager | The compliance manager · The BSA officer |
| Second look on reviews confirmed in bulk | 5% of “nothing relevant” reviews | 5% to 20%, in steps of 5 |
| Confidence a match needs before it is proposed | 0.85 | 0.70 to 0.95, in steps of 0.05 |
| Review clock for an EDD refresh | 5 business days | 5 · 10 business days |
| Review clock for onboarding | 2 business days | 1 · 2 · 3 business days |
| Languages searched | From the customer's footprint, plus English | From the customer's footprint, plus English · English and Spanish only |
| Look back | 10 years | 5 · 7 · 10 years |
| Keep reviews and their evidence for | 5 years | 5 · 7 · 10 years |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample data · “target” = the design goal, measured in the live solution · “estimated” = our estimate of where an analyst's minutes go · “typical” = published figures: a 2024 industry survey of investigators who adopted an AI adverse media tool (61%), and vendor and trade sources for the commonly reported 90–95% screening false-positive range and the 15–30 minutes per hit, directional only. Rules: 31 CFR 1010.230 and 1020.210 (customer due diligence), 31 CFR 1010.430 (five-year retention), 31 CFR 1020.320(e) (SAR confidentiality). Corvane National Bank, its customers, the companies, the news outlets, the articles and the people named on this page are fictional — characters and sample data in the working solution.
Adverse media screening is the check for negative news about a customer (fraud, corruption, money laundering, sanctions evasion, tax crime) before onboarding a higher-risk customer, at an EDD refresh or after a trigger event. Adverse Media Screening puts eight agents in front of that check: they search every language tied to the customer, group the copies into stories, try to prove each match wrong and cite the original sentence. A KYC analyst confirms and an EDD reviewer accepts.
The Search agent builds the plan from the customer record, the Translation agent translates sentence by sentence and keeps the original, and the Story agent groups copies into one story per event. The Match agent compares six identifiers, the Different-person agent tries to prove each match wrong, and the Risk agent proposes a category and stage. People then confirm, accept and approve.
It searches every language tied to the customer by citizenship, address, subsidiary or wire country, plus English, with risk terms for each category in each language. Each article is translated sentence by sentence, and the file cites the sentence as published, with the English beside it.
The Match agent compares name, age, occupation, company, role and city with the customer record, working out the customer's age on the day each article was published. The Different-person agent then tries to prove the match wrong and records every test. On the sample review, two stories about other men were set aside with five conflicting identifiers each.
Every night the agents search again for high-risk customers. A new article about a story already on file joins that story instead of opening a new alert; only a moved stage, such as investigation to charged, or a new story reaches a person.
Yes. The agents propose and never confirm a story, accept it into a file, change a rating or write to a customer record. A KYC analyst confirms with a reason code, a different person accepts it into the customer file, and the compliance manager approves any rating change.
No. Relationship managers see status only, such as “Compliance review in progress. No action needed.” Whether a SAR exists never appears in this solution or on any customer-facing view.
The Agentic Solution Engine builds and deploys it from your requirements and documents (your categories, review clocks, sources and a sample of past reviews), and it goes live once every quality gate has passed. We will walk you through it on your own reviews first.
Every PEP hit settled with its sources, and PEP treatment ends the day the rules allow.
1-min demoSanctions Ownership & ControlEvery blocked share traced to its source; the 50% call made while the payment is held.
1-min demoOFAC Sanctions SearchOne verdict line, the license's conditions matched to your facts, every source dated.
1-min demoWe'll run Adverse Media Screening on a sample of your own past reviews.