TCAML-S001: AML Screening Skill

Maintainer: Tom Custos
Category: Human-supervised AML operations

1. Purpose

This skill defines how an AML agent should assist with screening a subject against sanctions, PEP, watchlist, adverse risk, and internal risk sources.

The skill is designed for human-supervised workflows.

The agent may collect, compare, structure, and summarize information.

The agent must not approve, reject, clear, block, or legally classify a subject.

2. When to use

Use TCAML-S001 when reviewing any of the following:

individual customer;
company customer;
director;
ultimate beneficial owner;
authorized representative;
counterparty;
supplier;
merchant;
payment beneficiary;
wallet owner;
related entity;
name returned by another AML workflow.

3. Agent may

The agent may:

collect available identifiers;
request missing identifiers;
run or request screening checks;
compare possible matches;
identify similarities and differences;
summarize source results;
document uncertainty;
prepare an evidence log;
prepare a false positive assessment;
recommend human review or escalation.

4. Agent must not

The agent must not:

approve a customer;
reject a customer;
clear a sanctions hit as final;
determine that a person or entity is legally sanctioned;
determine regulatory reporting obligations;
provide legal advice;
ignore missing identifiers;
invent missing facts;
hide conflicts between sources;
remove findings because they are inconvenient;
override internal policy;
replace the MLRO or compliance officer.

5. Required input

The agent must collect or mark as missing:

For individuals

full name;
aliases or alternative spellings, if available;
date of birth, if available;
nationality, if available;
country of residence, if available;
document country, if available;
customer or case ID, if available.

For entities

legal name;
trading name, if available;
registration number, if available;
jurisdiction of incorporation, if available;
registered address, if available;
directors, if available;
UBOs, if available;
website or domain, if available;
customer or case ID, if available.

If required input is missing, the agent must continue only if permitted by the workflow and must state that confidence is limited.

6. Recommended sources

The agent should use source categories appropriate to the implementation:

sanctions lists;
PEP datasets;
regulator notices;
law enforcement notices;
government company registries;
court or insolvency records;
internal watchlists;
licensed commercial data providers;
public adverse media sources;
prior internal case history.

The agent must identify which sources were checked.

The agent must not imply that unchecked sources were checked.

7. Workflow

Step 1: Normalize subject data

The agent should organize identifiers into a consistent structure.

For names, the agent should preserve the original spelling and record known variants.

For dates and jurisdictions, the agent should preserve source-provided format and add normalized format where possible.

Step 2: Identify missing data

The agent must list missing identifiers that materially affect screening quality.

Examples:

missing date of birth;
missing nationality;
missing registration number;
missing jurisdiction;
unknown UBOs;
unknown source freshness.

Step 3: Run or request screening

The agent may use one or more screening systems.

The agent should record:

source name;
source type;
date/time checked;
query used;
result summary;
evidence quality grade.

Step 4: Compare possible matches

For each possible match, compare available identifiers:

name similarity;
aliases;
date of birth;
nationality;
residence;
document country;
registration number;
jurisdiction;
address;
role or relationship;
source program or category.

Step 5: Classify match assessment for human review

The agent may use the following operational assessment labels:

No relevant result found
Possible false positive
Possible match - requires review
Strong possible match - escalate
Insufficient data - requires additional information

These labels are not final legal determinations.

Step 6: Assign confidence

The agent must assign confidence:

Low;
Medium;
High.

The agent must explain the reason.

Confidence should be reduced when key identifiers are missing, sources conflict, or only name-only screening is available.

Step 7: Prepare output

The agent must produce a structured output containing:

subject summary;
sources checked;
possible matches;
match assessment;
confidence;
missing information;
evidence log;
recommended next human action.

8. Evidence quality

Use the following evidence quality labels:

Grade	Evidence type
A	Official government, regulator, court, registry, sanctions authority
B	Official company or institution source
C	Trusted commercial or licensed data provider
D	Reputable media, public records, NGO or research source
E	User-provided or unverified source

Where evidence quality is unclear, mark it as Unknown and explain why.

9. Confidence guidance

High confidence

Use only when:

key identifiers are available;
multiple relevant sources were checked;
the result is supported by strong identifier comparison;
no material conflicts remain unresolved.

Medium confidence

Use when:

some identifiers are available;
at least one reliable source was checked;
the assessment is plausible but not complete;
some uncertainty remains.

Low confidence

Use when:

screening is based mainly on name only;
date of birth or jurisdiction is missing;
source freshness is unclear;
only one weak source was checked;
source results conflict;
identity cannot be reliably compared.

10. Escalation triggers

The agent must recommend escalation when any of the following occur:

possible sanctions match with insufficient disambiguation;
strong name and identifier overlap with a sanctioned or listed subject;
PEP match requiring policy review;
adverse source from regulator, court, law enforcement, or official notice;
conflicting source data that cannot be resolved;
subject refuses or fails to provide key identifiers;
UBO or ownership data is unavailable for an entity where required;
internal policy threshold is met;
the agent cannot determine whether a match is likely false positive.

11. Output requirement

The output must be suitable for human review.

It should not be written as a final approval or rejection.

It should make clear:

what was checked;
what was found;
what remains unknown;
what evidence supports each finding;
what human action is recommended.

12. Completion criteria

A TCAML-S001 review is complete only when:

subject identifiers are recorded;
missing identifiers are documented;
sources checked are listed;
possible matches are compared;
confidence is assigned and explained;
evidence is logged;
next human action is stated.

If these conditions are not met, the output should be marked incomplete.

tcaml-s001-aml-screening

SKILL.mdskills/core/TCAML-S001-aml-screening/SKILL.md