Adverse media screening is a volume problem. A common name returns hundreds of articles, and an analyst reads them one by one to find the few that describe the right person.

In one case from our demo data, a name search returned 212 articles. Two were about the customer. The other 210 were about namesakes, other companies with similar names, or the customer mentioned in passing in stories with no risk content at all.

Why the volume keeps growing

News search is wide by design. It indexes local papers, trade press, court listings and aggregators that republish the same story under a new headline. A single local story can appear twenty times. A name shared with a public figure can return thousands of results that have nothing to do with the customer.

Narrowing the search with keywords such as fraud, laundering or sanctions helps, but it also drops articles that describe risk in other words. Most teams end up reading more, not less.

Resolve the person first

Date of birth, location, employer and known associates rule out most results before anyone reads them. An article about a namesake in another country is a collision, not a risk.

The Adverse Media agent extracts what each article says about the person it describes: age or birth year, city, employer, role, family members, companies. It compares those facts with the KYC file. An article that disagrees on two or more facts is set aside. An article that agrees, or that gives too little to tell, stays in.

Keep the reason for every discard

The Adverse Media agent records why each article was set aside: different age, different city, different company. When an examiner asks how 210 results were cleared, the answer is already written.

• Different person: age, city or employer does not match the file.

• Duplicate: the same story republished by another source.

• No risk content: the customer is named, but nothing adverse is reported.

• Outdated and resolved: a matter closed with no finding, with the date.

• Insufficient detail: kept open for the analyst, with the reason.

The last category matters most. An article that cannot be resolved is not discarded to make the number smaller. It goes to the analyst with what is missing.

Thresholds for what counts as a disagreement are set by the team, not by the agent. A bank that wants every article with a matching name and city read by a person can say so, and the agent keeps those articles in the analyst's pile.

Read what remains

The two articles that match are summarized with their source and date, and linked to the case. The analyst reads two stories instead of 212.

In the Halvard Trading case, the relevant story reported that the company's owner was linked to a customs fraud probe opened in 2025. The summary states what the article reports, who published it and when, and does not go further. A probe is not a conviction, and the summary says so.

That finding becomes one evidence item in the case. If the analyst escalates, the SAR draft cites it as public reporting, with the source, not as a fact the bank established.

Sources and language

Not every source carries the same weight. A court record or a regulator's public notice is a different kind of evidence from a blog post or an anonymous forum. The agent records the type of source for every article it keeps, and the summary says what kind of publication made the claim.

Language matters too. A customer with business abroad may appear in the local press in another language, under a transliterated name. The agent searches in the languages relevant to the customer's countries of business and translates what it keeps, with the original linked. The analyst reads the summary in English and can always check the original text. An article that only appears in translation is flagged as such, so nobody mistakes a summary for a quotation.

When to search again

Adverse media is not a one-time check at onboarding. Stories appear after the account opens. The agent reruns the search when the customer's risk rating changes, when another trigger reopens the file, and on a schedule for high-risk customers. Each run keeps the previous discards, so a rerun only reads what is new.

That makes the second search much cheaper than the first. Articles already resolved stay resolved, with their reasons, unless the KYC file changes in a way that could reopen them.

What to do on Monday

Pick five recent adverse media reviews and check how each discarded article was recorded. If the file says only "no relevant results", ask the analyst how they knew. Then make that answer the minimum standard for every discard, whoever or whatever does the reading.