A Queue That Arrives Sorted
Your analysts open a queue that is already sorted. Every new alert the pipeline handles carries a verdict, a confidence and a proposed next step before anyone picks it up, so the team starts with the alerts that matter.
AI that does the gathering.
People who make the call.
AI agents that work your alert queue alongside your analysts. For each alert, an agent gathers the match details and the client record, compares them field by field, and proposes a verdict with its reasoning written out. Then it stops. Nothing is closed, cleared or escalated until a person approves it, and your institution sets where the agents work and how far they go.
What is an agentic workflow?
A sequence of steps an AI agent carries out on its own, within limits a person sets.
In DetectX®, it prepares the case inside AlertViewer, on the same queue, lists and audit trail as every other alert. A person decides.
- 1Alert Raised
- ACAgentic ComplianceBefore the alert is opened
- 2Analyst Opens
- AIAI CopilotAt the open alert
- 3A Person Decides
A Second Pair of Eyes
With the AI Copilot, your analyst does not work an alert alone. Beside the open alert, it sets out what matches, what differs and what is missing, and suggests an outcome. The analyst stays in charge of the alert from start to finish.
A Person Decides
3The analyst accepts, overrides or declines every verdict, and closures go to a second person.
The alert workload
Seven steps stand between an alert and a decision.
Today an analyst does all seven by hand, one source at a time. Switch views to see which steps move to the agent, and which always stay with people.
Today: 0 steps with AI, 7 with people.
Nothing prepared. Every step is done by hand.
- Load the match details
- Look up the client
- Compare each match, attribute by attribute
- Propose the outcome and next step
- Write the reasoning
- Decide: confirm, override or decline
- Second check under four-eye control
How it works
Five steps, the same for every alert.
The steps the agent takes on, shown on one alert. For each one, your institution sets the limits, so the agent works only where and how you allow it.
Every match, gathered first.
The agent pulls each list match on the alert, with the list it came from, the entry and its similarity score. Nothing is judged at this step.
You setWhich alerts the agent works on.
For example, PEP alerts for private-banking clients.
- enrich
- client
- classify
- verdict
- policy
PEP list · entry 90210004411BIANCHI, Marco
Similarity 0.8700- EntryDate of birth12.02.1962Loaded
- EntryNationalityITALoaded
- EntryPlace of birthNot on the list entryMissing
The person you know, not just the name.
It loads the client record your institution holds, so each match is weighed against the client. A missing field stays missing: the agent records the gap, it does not guess.
You setWhich client fields the agent may read.
- enrich
- client
- classify
- verdict
- policy
Client 418.229 · relationship approvedMarco Bianchi
From your records- ClientDate of birth04.11.1979Read
- ClientNationalityCHERead
- ClientAccount and transaction dataNot permitted for this pipelineNot read
Field by field, with a confidence.
The client is compared with each list entry, field by field, and the match gets a verdict with a confidence from 1 to 5. When the deciding fields are missing, the verdict is “Needs clarification”, not a guess.
You setHow differences are weighed.
For example, whether a different date of birth alone rules a match out.
- enrich
- client
- classify
- verdict
- policy
Match 1 · PEP listClient against list entry
| Field | Client | List entry | |
|---|---|---|---|
| Date of birth | 04.11.1979 | 12.02.1962 | Differs |
| Nationality | CHE | ITA | Differs |
| Place of birth | Lugano | – | Missing |
Match verdictProbable false positive
4/5One verdict, every reason written down.
The match verdicts are combined into one verdict for the alert, with the reasoning written out line by line, so a reviewer can check each line against the evidence.
You setHow several match verdicts combine into one.
- enrich
- client
- classify
- verdict
- policy
Alert AL-ONLINE-00312744Probable false positive
4/5- Line 1The given name and surname matchNameMatches
- Line 2Date of birth differs by 17 years04.11.1979 against 12.02.1962Differs
- Line 3Nationality differsCHE against ITADiffers
Proposed by the agent. Approved by a person.
The verdict leads to the action your policy sets, and the agent stops there. Nothing is closed, whitelisted or escalated until a person approves it.
You setThe action for each verdict, and who must approve it.
- enrich
- client
- classify
- verdict
- policy
Proposed for a probable false positiveWhitelist, then four-eye confirmation
Proposed- 1Approve the actione.marti, reviewerWaiting
- 2Confirm the closurej.feldmann, second personWaiting
- –Close the alert automaticallyNot allowed by your policyOff
Illustrative product views. All data synthetic.
Who decides
Who decides is your choice, not ours.
The agent gives every alert a verdict. For each verdict, your pipeline sets whether the agent carries out the action or waits for a person first.
Definite false positive
Whitelist + four-eyeClosed as a false positive, a second person confirms
Probable false positive
Whitelist + four-eyeClosed as a false positive, a second person confirms
Needs clarification
Inquiry to relationship managerMissing information requested before anyone decides
Possible true positive
Send to four-eye reviewA reviewer checks the case before anything happens
Highly probable true positive
Escalate to ComplianceStrong indication of a true match
The agent carries out 2 of 5 verdicts. The other 3 wait for a person, with the reasoning attached.
AI functionsOn. Verdicts and reasoning on every alert.
Each closure as a false positive still needs a second person to confirm it under four-eye control. Settings shown are illustrative.
The review
The alert arrives prepared. You decide.
One alert the agent has worked, as a reviewer opens it. Every step is yours to drive: nothing is approved on its own.
Illustrative exampleExample data; names and identifiers are fictional.
Alerts › Online › AL-ONLINE-00312744e.marti
91
Alert ID
AL-ONLINE-0031274414.10.2026 09:42
Matches
Name check [0.9104] · Global sanctions listCustomer Nr.
418.229Status
OPENAgent assessmentAI · reasoning attached
Verdict
Definite false positive
High confidenceThe verdict comes first.The agent has already placed this alert: definite false positive, with high confidence.Reasoning · list entry 90210004411
- Date of birth differs by 22 years.
- Nationality differs: client Swiss, list entry Austrian.
- Full name matches; no other identifier matches.
Proposed action
Whitelist + four-eyeWaiting for a person
Nothing runs until you decide.Approve the proposed action or change it. A closure then goes to a second person to confirm.FieldList entry 90210004411Client 418.229
NameOberholzer, DanielOberholzer, Daniel
Date of birth17.11.195803.05.1981
NationalityAustriaSwitzerland
Client since–2019
The deciding fields are marked.List entry and client record side by side. The red fields are the ones the reasoning rests on.Explore the DetectX® Platform
Rules work with every capability.
The Agentic Workflows decides what happens. The other capabilities give it more to decide on. Pick one to see what it does.
Agentic Workflows FAQs
What compliance teams ask first, starting with what is available today.
Something we have not covered?Talk to our team
An agentic workflow is a sequence of steps an AI agent carries out on its own, within limits a person sets. In DetectX®, the agent prepares a screening alert before an analyst opens it. It loads the match details and the client record, compares them field by field, and proposes a verdict, a confidence from 1 to 5 and the next step.
Yes, for the parts that are live. The AI Copilot recommends an outcome on the open alert, and Agentic Compliance prepares alerts in the pipeline before anyone opens them. In both, a person decides. Automate, where the agent would carry out an action on its own, is on the roadmap.
No. The agent recommends and a person decides. Every proposed action waits for an analyst to approve it, and an alert closed as a false positive always goes to a second person under four-eye control.
Name Screening alerts in AlertViewer. Your pipeline selects them by alert type, subtype and client, for example PEP alerts on private-banking clients. Alerts the pipeline does not handle are worked as usual, in the same queue.
It says so. If the fields that would decide a match are missing, the verdict is “Needs clarification” and the proposed action is an inquiry to the relationship manager. A missing client field is recorded as missing; the agent does not guess a value to fill the gap.
Both are kept on the alert. On accept, the agent’s reasoning is written as the comment. On override, the comment records the action the agent suggested, the action chosen and the analyst’s reason, and the agent’s original reasoning stays on file. The case then runs through your Workflows procedure as before.
Go Deeper
All resourcesInsightHow LLMs Can Reduce the Burden on Compliance Teams
Where large language models can take on routine compliance work, and where a person still has to decide.
Read the articleInsightIs Human Input Becoming Redundant in AI-Driven Financial Crime Prevention?Why judgement and accountability still sit with people, and what that means for reviewing an AI recommendation.
Read the articleSolutionWorkflowsThe case management the agent hands over to: four-eye control, escalation and investigations, all on one record.
See WorkflowsSee Agentic Workflows in a demo
Book a working session on your own workflow policy. Bring your questions about what can be automated under your governance model and what must stay with an analyst.