IndustriesFinTech

Compliance is not the problem. The false positives are

You cannot loosen the controls and you should not want to. What you can do is stop spending most of your compliance capacity on alerts that were never going to be anything, and stop losing good customers to an onboarding flow that treats them like suspects.

70%

of financial institutions have lost prospective clients to slow onboarding

Fenergo, up from 48% in 2023

85-95%

typical share of AML alerts that turn out to be false positives

Facctum

1 in 5

onboarding applications abandoned over KYC documentation friction

Industry reporting

3 min

verification time beyond which abandonment rises sharply

Industry reporting

Where the time and revenue actually go

These are the specific, measured problems in fintech and financial services. Not generic digital transformation.

01

Onboarding loses the customers you wanted most

70 percent of financial institutions report losing prospective clients to slow or complex onboarding, up from 67 percent the year before and 48 percent two years earlier. The trend is going the wrong way while everyone talks about digital onboarding.

02

The drop off is concentrated in document handling

Around one in five applications are abandoned specifically over KYC and AML documentation, and a request to re-upload a document sharply increases the chance the applicant simply leaves. The control is necessary. The friction in how it is collected mostly is not.

03

Compliance analysts spend their week clearing noise

False positive rates on AML alerting typically run between 85 and 95 percent. That is a team of trained people, expensively recruited, spending the overwhelming majority of their time confirming that nothing happened.

04

Alert triage is inconsistent between analysts

Two analysts given the same alert will gather different evidence and write different narratives. That is a supervision problem and a regulatory one, because consistency of process is exactly what an examiner asks to see.

05

Disputes and chargebacks arrive as unstructured text

A customer describes the problem in their own words across chat, email and phone. Someone reads it, classifies the reason code, pulls the transaction, checks the merchant, and starts the case. The classification drives the deadline, and getting it wrong costs the case.

06

Support cannot answer without the account state

Why was I charged this, did my transfer go through, why was my card declined, when will my payout land. None of these are in a help centre. They are in the ledger, the payment processor and the risk engine, so a documentation only assistant is useless for the questions that actually arrive.

07

Everything has to be explainable afterwards

In this sector a system that produces a good outcome by means nobody can reconstruct is a liability rather than an asset. Whatever gets automated has to leave a record an examiner can follow, which rules out most of what gets sold as AI automation.

What we build for fintech teams

01

Onboarding document handling that fails gracefully

Extracts and validates what was submitted, tells the applicant precisely what is wrong while they are still in the flow rather than by email two days later, and escalates genuine edge cases to a human instead of rejecting them.

02

Alert triage with a written evidence trail

Gathers the supporting evidence an analyst would gather, assembles a consistent narrative, and presents a recommendation with its reasoning. The analyst decides. The system does the retrieval and the writing, and every step is logged.

03

Dispute intake and classification

Reads the customer's own description, classifies the reason code, pulls the transaction and merchant history, and opens the case with the deadline already calculated and the evidence pack started.

04

Support grounded in live account state

Reads the ledger, the processor and the risk engine before answering, so the routine questions resolve properly and anything touching a control escalates rather than being guessed at.

05

Audit trail as a first class feature

Every retrieval, every tool call, every model and prompt version recorded per decision, so the reconstruction an examiner asks for is a query rather than an archaeology project.

Two workflows, end to end

Illustrative composites, drawn from how these workflows actually run in each industry. Not named client engagements.

01Getting the customer through onboarding
The situation

A business customer uploads incorporation documents at 11pm. One document is a scan with the registration number cut off at the page edge.

What usually happens

The application queues. Two days later a reviewer opens it, spots the problem, and sends a templated email asking for the document again. The applicant, who has by now started an application elsewhere, does not respond. The case is closed as abandoned and counted as a fraud control working.

What our system does

Extraction runs while the applicant is still in the flow. The system detects the truncated field, shows them exactly which corner is missing on their own upload, and accepts a corrected image immediately. The remaining checks run in parallel, the registry lookup confirms the entity, and the case reaches a reviewer complete rather than incomplete. Anything genuinely ambiguous, including anything that touches a sanctions or PEP match, goes to a human untouched.

Steps the system took
01Extraction and validation while the applicant is still present
02Precise, specific correction requests rather than a generic rejection
03Registry and screening checks run in parallel
04Reviewer receives a complete case, not a partial one
05Sanctions and PEP matches routed to a human, never auto cleared
02Clearing the alert queue honestly
The situation

A transaction monitoring rule fires on a customer whose incoming payment pattern changed this month.

What usually happens

An analyst opens the alert, pulls twelve months of history from one system, the KYC file from another and the prior alert history from a third, reads it all, writes a narrative, and closes it as a false positive. Forty minutes for a case that was never going to be anything, repeated across a queue where nine in ten end the same way.

What our system does

The system assembles the evidence pack before the analyst opens it: the history, the KYC file, prior alerts on the same customer, and the specific pattern change that triggered the rule, with a drafted narrative citing each source. The analyst reads, checks and decides. The decision, the evidence and the reasoning are recorded together. Nothing is auto closed, because in this sector the value is in making the analyst faster rather than in removing them.

Steps the system took
01Evidence assembled from every relevant system before review
02A consistent narrative drafted, with each claim cited
03The analyst decides, always, on every alert
04Decision, evidence and reasoning recorded as one record
05Reconstruction for an examiner is a query, not a project

Wired into the fintech stack you already run

Core and ledger

  • Stripe
  • Adyen
  • Marqeta
  • Modern Treasury
  • Core banking APIs

Identity and screening

  • Persona
  • Onfido
  • Alloy
  • ComplyAdvantage
  • Sanctions and PEP lists

Monitoring and cases

  • Unit21
  • Sardine
  • Case management systems
  • Transaction monitoring rules

Support and CRM

  • Zendesk
  • Intercom
  • Salesforce Financial Services
  • HubSpot
  • Twilio

Start with one fintech workflow

A 30 minute audit, no sales pitch. We map where automation can cut manual work and create measurable ROI across the tools you already use.