Adverse media screening fails in two ways: drowning analysts in irrelevant articles, and missing the one report that matters. The agent searches across sources, discards articles about other people, groups duplicates, classifies each finding by allegation and credibility, and drafts a summary. An analyst decides whether a finding is about the customer and what it means for the relationship.
What is adverse media screening, and who owns it?
Adverse media screening searches news, public records and other sources for negative information about customers and connected parties: allegations of financial crime, fraud, corruption, sanctions evasion, serious regulatory action or other conduct relevant to risk. Under a risk-based approach, firms screen at onboarding and periodically, especially for higher-risk customers, and alongside PEP and sanctions screening.
KYC and EDD analysts review the findings, under the MLRO's policy. The work is hard because search results are noisy: common names, articles in many languages, reprints of the same story, opinion pieces, and stories where the customer is the victim rather than the subject. Whether an AI agent can do adverse media screening sets out the boundary.
What does the agent prepare, and what does a person approve?
- Person, MLRO: approves the screening policy. Which customers, which sources, which categories of allegation, and how far back.
- Agent, depth 1: searches. For the customer and connected parties, such as directors and beneficial owners, across approved sources and languages, using names, aliases and identifiers.
- Agent, depth 1: matches identity. Discards articles clearly about someone else, using age, location, occupation, company names and other identifiers. Ambiguous matches are kept and marked.
- Agent, depth 1: groups duplicates. The same story across many outlets becomes one finding with its sources listed.
- Agent, depth 1: classifies. Type of allegation, the customer's role (subject, witness, victim, mentioned), stage (allegation, charge, conviction, acquittal, settlement) and date.
- Agent, depth 1: grades credibility. Established news outlets, court and regulator records, blogs and anonymous posts are weighted differently. All are marked as external.
- Agent, depth 2: drafts the summary. Relevant findings, ambiguous matches, and what the policy requires next.
- Person, analyst: decides whether each finding relates to the customer and whether it changes the risk view.
- Person, senior approver: approves continuing a relationship with material adverse media, where policy requires.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Policy | Nothing | None | MLRO |
| Search | Customer and connected parties | 1 | None |
| Identity | Discards clear non-matches | 1 | Ambiguous cases |
| Duplicates | Groups reprints | 1 | None |
| Classify | Allegation, role, stage, date | 1 | None |
| Credibility | Grades sources, marks external | 1 | Weighing them |
| Summary | Drafts with next steps | 2 | Analyst decides |
| Approval | Nothing | None | Senior approver |
Which controls apply, and what evidence does each leave?
Discards explained. Every article discarded as a different person has a reason, such as a different age or country. Evidence: the discard log, which QA samples, because wrong discards are the dangerous error.
Role and stage recorded. Being named as a victim, or acquitted, is different from being convicted. Evidence: classification on each finding.
Sources marked external. A news report is an external claim, not a verified fact. Evidence: provenance on each finding.
Search parameters kept. Names, aliases, languages and sources searched. Evidence: reproducible searches, which show the screening was adequate.
Decision by a person. Evidence: the analyst's decision on each relevant finding.
What should you measure?
- Articles per customer before and after the agent's filtering.
- Wrong discards found by QA sampling, which should be rare and each reviewed.
- Findings that changed a risk rating.
- Analyst time per screening.
- Languages and sources covered against policy, reviewed whenever the customer base extends into new markets.
Where does it usually go wrong?
Name-only matching. Common names produce many false matches, and aggressive filtering produces missed true ones. Use identifiers wherever possible and keep ambiguous cases for the analyst.
Old stories treated as current. An allegation from many years ago that ended in acquittal should not weigh like a fresh charge. Stage and date matter, and the summary should show both for every finding.
Search terms too narrow. Searching only for the customer's name plus words like "fraud" misses stories that use other terms, such as "investigation", "charged" or "struck off", or that are written in another language. The search term set should be approved by the MLRO, cover the languages of the customer's countries, and be reviewed when QA finds a missed story.
Connected parties forgotten. Adverse media about a company's director or owner can matter more than media about the company. Screen them.
Overlap with PEP screening. PEPs often attract media coverage; adverse media is part of assessing their risk. See what PEP screening is and whether an agent can do PEP screening.
Personal data overcollected. Storing whole articles about people who turned out not to be the customer holds personal data with no purpose. Keep the discard reason, not the article. Where screening processes personal data in new ways, consider DPIA screening before scaling up.
The overall stance on compliance agents is in AI compliance agents.
Where Pryme fits
In Pryme, every article reaches the agent as an external fact through the Pryme Context Protocol, and court or regulator records carry their issuer, so the summary shows evidential strength by source. An external claim cannot by itself change what any agent may do for a customer; the analyst's decision does. Changing a risk rating is outside the screening agent's role scope.
Questions people ask
Should the agent screen continuously or on a schedule?
Continuous screening catches new stories sooner, but produces more findings to review. Many firms screen high-risk customers continuously and others on a schedule set by risk.
Can the agent read sources in other languages?
Yes, if your policy includes them. Translations can lose nuance, so analysts should be able to see the original and, for material findings, get a reliable translation.
What if a finding is serious but the customer denies it?
Record the customer's account as declared alongside the finding. The analyst weighs both, and serious unresolved findings go to the senior approver.