KYC and AML are usually run by different teams with different systems. Onboarding checks who the customer is. Monitoring and investigations look at what the customer does afterwards. The gap between them is where most of the pain sits: an account opened on a thin file carries that gap into every alert and review that follows.
AI agents can now do much of the checking on both sides. They read documents, run identity and screening checks, trace who owns a company and put the results in one file. The question compliance leaders ask us is less about whether that works and more about where the agent has to stop and a person has to decide.
We rewrote this page on November 8, 2026. The earlier version said AI agents could approve customers without manual intervention, which contradicts how we build FluxForce and a coming EU rule. It presented an HSBC result from a Google Cloud transaction-monitoring model as an agentic onboarding outcome, and it repeated a Commonwealth Bank figure we couldn't trace to any source, so that case is gone. Several speed and accuracy figures with no source were removed as well. This version is written for AML, KYC and onboarding leads at banks and fintechs.
The complete-file test is an editorial test for judging whether an AI-prepared onboarding file is ready for a person to decide on: Identity, Ownership, Screening, Risk and Decision. It is not a regulatory standard or a product certification.
Yes, most of the checks, but not the decision. An AI platform can collect and read identity documents, run document and liveness checks, pull company registry data, trace beneficial owners through layers of companies and trusts, screen every connected person against sanctions, PEP and adverse-media sources, and propose a risk rating on the factors in your policy. What it should hand over is one file with the evidence and the gaps marked. Whether the account opens, is declined or goes to enhanced due diligence is a decision for a person with the authority to make it. From 10 July 2027, the EU's AML Regulation requires that kind of decision to be subject to meaningful human intervention when automated processes or AI are involved.
KYC is the front of an AML program. The customer due diligence done at onboarding decides what "normal" looks like for that customer, and every later alert is judged against it. If the file is wrong or thin, monitoring rules fire on the wrong things and investigators start each case by redoing onboarding work.
| Stage | What gets checked | What an agent prepares | Who decides |
|---|---|---|---|
| Identity | ID documents, selfie and liveness, consistency across sources | Verification results with mismatches flagged | Onboarding analyst |
| Business and ownership | Registry data, directors, licences, ownership chain | Ownership chart with controllers and the threshold applied | Onboarding analyst or KYB reviewer |
| Screening | Applicant, directors, owners and signatories against sanctions, PEP and adverse media | Hits with match reasoning; namesakes ruled out with verified data | Analyst for every true or unclear match |
| Risk rating | Your policy's risk factors | Proposed rating, the factors behind it, EDD trigger if any | Compliance team approves the rating |
| Account decision | The whole file | A complete file with gaps listed | Authorized approver |
| After onboarding | Trigger events, alerts on activity | A review or case with the onboarding file attached | Analyst, then MLRO for escalations |
Our page on customer due diligence sets out what examiners look for in each element. The point of the table is the last column. Agents can do most of the work in the middle columns, and none of the decisions in the right one.
Ask a vendor to show you an onboarding file where one check failed or a source returned nothing. If it looks the same as a clean file, your approver can't tell a complete file from a thin one.
The earlier version of this page compared three kinds of AI. The comparison is still useful, with one correction: none of them should be the one deciding.
| Type | What it does in KYC and AML | Main limit |
|---|---|---|
| Rule-based systems | Check IDs, documents and transactions against fixed rules and flag matches | Produce many false positives and don't adapt when typologies change |
| Generative AI | Reads documents and drafts summaries, reports and explanations for a reviewer | Doesn't act on a case and can state things confidently that aren't in the evidence |
| AI agents | Run multi-step checks across sources, gather evidence, propose an outcome and record each step | Need good data and clear governance, and act only within the authority you give them |
Agents are the right fit for onboarding because the work is multi-step: read the document, check it, look up the company, find the owners, screen each one, then rate the risk. Each step depends on the last. That's also why the record of each step matters as much as the result.
The checks a KYC AI agent can run fall into six groups. Each produces evidence for the file rather than a decision.
Ownership is where onboarding most often stalls. Under the US customer due diligence rule a beneficial owner is each individual who owns 25 percent or more of a legal entity customer's equity, plus one individual with significant responsibility to control it. Other markets set different thresholds, and a cross-border group can need more than one. An agent that applies the threshold you configure for each market, and shows its working, saves the reviewer the slowest part of the job.
Speed depends on your products, your markets and your data sources. We don't quote onboarding times from demos. Measure turnaround on your own applications, side by side with your current process, before you change anything.
Search interest in KYC onboarding fraud detection with AI is about whether the software can catch a fake applicant at the door. It can help with several signals:
What AI can't do is decide intent. A mismatch might be fraud or a data-entry error, and a shell indicator might be a new holding company with a sound reason to exist. The agent's job is to surface the signal with its evidence and make the reviewer's decision faster. For the identity side in more depth, see our guide to KYC and digital identity verification.
The onboarding file pays off later. When a transaction-monitoring or screening alert fires on a customer six months after the account opened, the investigator's first questions are who this customer is, who owns them, what they said they'd do with the account and what risk rating they were given. If that file sits in a different system, or was never completed, each case starts with a rebuild.
Agentic AML investigation means an agent picks up the alert, pulls the onboarding file, the related transactions and fresh screening results, and assembles a case with the evidence attached. The analyst reviews it and decides whether to close, escalate or report. Where the case points to suspicion, the agent can draft the report content for the MLRO, who edits and approves it before anything is filed through your normal process.
Our AI KYC AML analyst, Rhea Ledger, is the agent profile for this work: screening, ownership tracing and the first risk rating, then draft reports for the MLRO to approve. Cases are worked in Case Management and Investigations, where every action is timestamped. Keeping the file current between reviews is covered in our piece on perpetual KYC.
The earlier version of this page cited HSBC as an agentic onboarding success. The result is real, but it comes from somewhere else.
HSBC used Google Cloud's Anti Money Laundering AI, a transaction-monitoring model. In a post on the Google Cloud blog, HSBC reports that "AML AI identifies two to four times as much suspicious activity as the previous system, while reducing the number of alerts by 60%". Google's June 2023 launch release explains the measure: it was based on the number of Suspicious Activity Reports filed with AML AI against a similar period under rules-based transaction monitoring at HSBC.
So it's a monitoring result at one bank, measured by reports filed. It isn't an onboarding result, it isn't an AI agent, and it says nothing about FluxForce. It does show that better detection models can cut alert volume while finding more of what matters, which is worth knowing when you weigh the monitoring side of an AML and KYC program.
The page also described a Commonwealth Bank of Australia case with a 60% alert reduction and faster onboarding. We couldn't find a source for those figures, so the case is removed.
In the US, the customer due diligence rule at 31 CFR 1010.230 says covered financial institutions "are required to establish and maintain written procedures that are reasonably designed to identify and verify beneficial owners of legal entity customers". Owners are identified when a new account is opened and verified using risk-based procedures. The section was last amended in September 2017. The rule doesn't say who presses approve, but the institution answers for the procedures and for each file, so the decision record matters.
In the EU, the AML Regulation, Regulation (EU) 2024/1624, addresses automation directly. Article 76(5) lets obliged entities adopt decisions resulting from automated processes or AI systems, on conditions. One is that any decision to enter, refuse or maintain a business relationship, or to increase or decrease customer due diligence measures, "is subject to meaningful human intervention to ensure the accuracy and appropriateness of such a decision". The regulation applies from 10 July 2027 for most obliged entities, so this is a rule to design for now, not one in force today.
International standards point the same way. Our summary of FATF Recommendation 10 covers its four core CDD measures and what to do when full CDD can't be completed. None of these frameworks bans automation. They expect the institution to stay accountable for each customer decision, which is easier to show when a person made it and the record says why. Our page on human-in-the-loop review sets out how to design that review so it is more than a checkbox.
Once you know platforms can automate customer onboarding checks, the next questions are which AI agent platforms have the fastest KYC turnaround, and who integrates KYC and AML workflows. Speed only counts on a complete file, so test both together. Ask each vendor to show the following on real files, ideally your own applications in a trial:
Then measure turnaround in shadow mode: time from application to a file a person can decide on, and how often files reach the approver complete the first time. A fast answer on an incomplete file is a slower answer later. Our CDD readiness assessment scores your current program across six CDD pillars and is a useful baseline before a trial.
Illustrative scenario, not a customer result.
This is an illustrative scenario, not a customer case. A fintech that issues business accounts receives an application from a trading company. The company is owned partly by another company and partly by a family trust, and one director shares a name with a person on a sanctions list.
The agents run the checks in parallel and assemble one file for the onboarding analyst.
The account opens on a person's decision, with a file an examiner can follow. The gap flag mattered most. Without it, a partial media search would have looked like a clean result.
AI agents can run the onboarding checks, gather evidence from the sources you connect, propose a rating with reasons and record every step. They don't own the decision. Your institution, its onboarding analysts, its MLRO and its compliance team stay accountable for every account decision, escalation, rating change and report.
| Component | Responsibility | Boundary |
|---|---|---|
| Identity agent | Checks documents, selfie and liveness, and matches details across sources | Flags mismatches; doesn't decide whether an applicant is genuine |
| Screening and ownership agent | Traces owners, screens every connected person and proposes the first rating | Rules out namesakes only with verified data and sends true matches to an analyst |
| Investigation agent | Builds a case from an alert with the onboarding file and fresh evidence attached | Recommends with reasons; doesn't close, escalate or file on its own authority |
| Onboarding analyst | Reviews the file, decides and records a reason | Acts within the authority your institution delegates |
| MLRO or compliance head | Decides escalations, approves reports and sets agent autonomy | Accountability cannot be delegated to an agent |
| Evidence record | Holds every check, source, gap, version, autonomy setting and decision | Shows what happened; it doesn't make a decision defensible on its own |
You decide how much each agent does without review. Any setting can be changed at any time, and every agent has a kill switch.
Covered financial institutions are required to establish and maintain written procedures that are reasonably designed to identify and verify beneficial owners of legal entity customers. Owners are identified at the time a new account is opened and verified according to risk-based procedures. A beneficial owner is each individual owning 25 percent or more of the equity interests, plus a single individual with significant responsibility to control, manage or direct the legal entity customer.
US covered financial institutions. The article does not say the rule prescribes who approves an account.
2026-11-08
Obliged entities may adopt decisions resulting from automated processes or AI systems provided that any decision to enter, refuse or maintain a business relationship, or to increase or decrease customer due diligence measures, is subject to meaningful human intervention to ensure the accuracy and appropriateness of such a decision. The regulation applies from 10 July 2027, and from 10 July 2029 for two categories of obliged entities.
EU obliged entities. Presented as an upcoming rule, not current law, and kept separate from US requirements.
2026-11-08
HSBC reports that AML AI identifies two to four times as much suspicious activity as the previous system, while reducing the number of alerts by 60%, based on 12 months after go-live compared with the preceding 12 months of rules-based monitoring.
One bank, transaction monitoring with Google Cloud AML AI. Not an onboarding, agentic or FluxForce result.
2026-11-08
States the HSBC figures were measured by the number of Suspicious Activity Reports filed with AML AI against a similar period under rules-based transaction monitoring at HSBC.
Vendor release; method statement for the HSBC figures only.
2026-11-08
| Metric | Definition | Decision guardrail |
|---|---|---|
| First-time complete files | Files that reached the approver with every required check and no unlisted gap, divided by all files. | A file with a hidden gap counts as incomplete. |
| Gap disclosure | Files where a source failed or returned nothing and the file shows it, divided by all files with a failed or empty source. | A silent source failure counts as a miss. |
| Time to decidable file | Time from application to a file a person can decide on, by customer segment. | Measure only on complete files, in shadow mode, against your current process. |
| Ownership resolution | Legal entity files where the owners at the configured threshold and the control person were identified, divided by all legal entity files. | Report by market, because thresholds and registry coverage differ. |
| Override rate | Decisions where the approver decided differently from the proposed rating or recommendation, divided by all decisions. | Review every override while an agent version is new. |
| Named approvals | Account decisions with a named approver and a recorded reason, divided by all account decisions. | Anything below all of them is a control gap. |
Agentic AI for KYC and AML works best when it's judged by the file it hands over, not by how few people touch it. A complete file shows the identity checks, the ownership chain, every screened person, the factors behind the rating and the gaps, and it reaches a person who decides and records why. The same file then sits behind every alert on that customer, which is where AML investigation time is won or lost.
The rules don't ban automation. They expect the institution to stay accountable, and from July 2027 the EU will expect a person in the loop on onboarding decisions made with AI. Build for that now and the audit trail comes with it. If you want to see that file built on your own applications, request a FluxForce demo.
AI agents can run most of the checks: document and liveness verification, registry lookups, ownership tracing, sanctions, PEP and adverse-media screening, and a proposed risk rating. They hand over a file with the evidence and gaps marked. The decision to open, decline or escalate an account stays with an authorized person.
It shouldn't. In our design the agent prepares the file and a named person approves, rejects or escalates the account. From 10 July 2027, the EU AML Regulation requires decisions to enter, refuse or maintain a business relationship made with automated processes or AI to be subject to meaningful human intervention.
KYC at onboarding establishes who the customer is, who owns them and what risk they carry. AML monitoring and investigations judge later activity against that profile. When the onboarding file is attached to each alert, investigators don't have to rebuild it before they start.
An agent picks up an alert, gathers the onboarding file, related transactions and fresh screening results, and builds a case with the evidence attached. The analyst reviews it and decides whether to close, escalate or report. Where a report is needed, the MLRO approves it before filing.
Turnaround depends on your products, markets and data sources, so vendor figures from demos tell you little. Measure it on your own applications in shadow mode: time to a file a person can decide on, and how often files reach the approver complete. A fast answer on an incomplete file costs time later.
It can flag altered or expired documents, failed liveness checks, synthetic or borrowed identities, details that disagree across sources and shell-company indicators. It can't decide intent. A reviewer looks at the evidence and decides whether a signal is fraud or an error.
No. HSBC reported the result for Google Cloud's AML AI, a transaction-monitoring model: two to four times as much suspicious activity identified and 60% fewer alerts, measured by reports filed against a similar rules-based period. It isn't an onboarding or agentic AI result.
31 CFR 1010.230 requires covered financial institutions to keep written procedures to identify and verify the beneficial owners of legal entity customers. Owners are identified when a new account opens: each individual with 25 percent or more of the equity, plus one individual with significant control.