Ripjar upgrades bank screening AI, but claims need scrutiny

Ripjar, a UK financial-crime technology firm, said on 24 September 2026 that it has added new AI models to ULTRA, its screening engine, and to Screening Assistant, its alert-triage tool. The updates cover sanctions, politically exposed person (PEP), watchlist and adverse-media screening — the checks UK banks run to work out who they are dealing with and whether that person or entity carries elevated financial-crime risk.

The announcement matters to UK compliance teams because it lands four months after the Financial Conduct Authority (FCA) published findings, on 28 May 2026, that deficiencies in screening and alert management were the most common causes of suspected sanctions breaches reported by firms. Ripjar's pitch is that better entity matching and an "auditable system of record" can close exactly those gaps. Whether that claim holds up is a separate question, and one that the announcement itself does not settle.

Ripjar Ltd is an English private company (company number 08217339); Companies House records show Ripjar Holdings Ltd has held at least 75% of its shares and voting rights since 29 October 2024.

What Ripjar changed

Ripjar says the updated matching system is designed to distinguish parent organisations, subsidiaries and similarly named group entities, and to handle complex naming conventions, including Arabic names. Screening Assistant, Ripjar says, uses agentic AI — software that can take a sequence of actions towards a goal rather than simply flagging data — to triage alerts and decide whether a raised alert genuinely relates to the entity being screened. Separately, Ripjar's adverse-media classifiers are intended to determine whether a news article concerns a specified risk and to identify the person involved.

Ripjar describes explainability as linking each alert, decision and summary to the evidence and sources behind it, and retaining that trail in an auditable system of record.

The performance claims, and what they don't show

Ripjar says ULTRA attributes risk to the correct entity with 98.7% accuracy and reduces the volume of alerts reaching an analyst by 99%. Both figures come from the company. Ripjar has not disclosed the evaluation dataset, sample size, class balance, the definition of a "correct" result, or the precision, recall, false-negative rate or confidence interval behind either number, nor named an independent assessor.

That gap matters because a missed true match in sanctions screening can be more consequential than a false alarm, and neither figure states a false-negative rate. It is also unclear whether the 98.7% figure applies equally across sanctions, PEP, watchlist and adverse-media tasks, or reflects a single entity-resolution benchmark. No baseline alert volume, customer cohort or observation period is given for the 99% reduction, and the wording refers specifically to alerts reaching an analyst, not necessarily to a 99% cut in total alert volume or workload.

Ripjar also cites scale: it says ULTRA resolves more than 10 billion historic articles and more than 5 million new articles a day, across more than 150 languages. Its own screening product page, however, describes adverse-media screening as covering more than 6 billion articles in more than 22 languages, with approximately 6 million new articles a day. These figures may describe different things — ULTRA's whole data estate against its adverse-media subset — but Ripjar has not explained the difference in the material reviewed.

Ripjar's material also states that it serves institutions including 25% of global systemically important banks and more than 35 Fortune 500 companies. It does not name those customers or explain how the figures were calculated.

What explainable screening should actually mean

A trail of evidence behind a decision is a necessary part of explainability, but not the whole of it. For a bank's financial-crime function, understanding an automated screening tool means being able to interrogate the matching logic itself, not just view a generated summary of an outcome. It means testing accuracy and recall on data that represents the bank's own customer base, including non-Latin scripts and variant name spellings. It means governing the thresholds and exclusion rules that decide which alerts are suppressed automatically, and monitoring the system for drift as models change, while retaining a route for human escalation and quality assurance over alerts the system closes without a person looking at them.

The Bank of England and FCA's joint 2024 survey of 118 regulated firms, published 21 November 2024, found that 81% of firms already using AI employed at least one explainability method — so the practice is now mainstream. The same survey found that one-third of reported AI use cases across the industry were third-party implementations, rising to 64% in risk and compliance specifically, and that 46% of firms using or planning AI reported only a partial understanding of the technologies involved, against 34% reporting complete understanding.

The UK rules that already apply

The FCA does not run a separate rulebook for AI in financial services. It applies its existing, outcome-focused principles and requires senior managers to remain accountable for the outcomes AI systems produce, whatever the underlying technology. That means a bank cannot transfer its financial-crime obligations to a vendor by adopting a new screening tool.

Automated sanctions screening is common — 70% of firms submitting the FCA's 2024–25 Annual Financial Crime Report (REP-CRIM) said they used it, and 81% said they screened repeat customers — but the FCA is clear that using automated screening is not itself an explicit legal requirement. Firms remain responsible for effective controls and for the compliance outcomes those controls produce, whatever tool sits behind them.

Separately, under regulation 35 of the Money Laundering Regulations 2017, which came into force on 26 June 2017, relevant UK firms must maintain appropriate risk-management systems and procedures to identify PEPs, their family members and known close associates, and to manage the enhanced risk those relationships carry. From 10 January 2024, firms have had to treat domestic PEPs as presenting lower risk than non-domestic PEPs unless specific enhanced-risk factors are present — a distinction that any automated triage tool needs to reflect correctly.

On the sanctions side, the UK Sanctions List became the sole source for all UK sanctions designations on 28 January 2026, following the closure of the Office of Financial Sanctions Implementation's (OFSI) Consolidated List. OFSI's guidance draws a distinction between a name match and a target match: sharing a name with a listed person does not, by itself, mean a customer is the designated person, and firms need to check the disambiguating attributes — such as date of birth or address — that OFSI's guidance sets out.

Adverse-media screening can support customer due diligence and a firm's own risk assessment, but the sources reviewed for this article do not establish a general, standalone UK legal requirement for every bank to use it. Its use depends on each firm's risk-based financial-crime framework.

The FCA's reality check on screening and alerts

The FCA's 28 May 2026 findings on sanctions systems and controls give a useful baseline against which to read any vendor's claims. In its own testing, 90% of alerts raised for exact-name matches correctly identified the sanctioned party, against 75% where the name appeared in a slightly different form — a gap that speaks directly to the variant-name and non-Latin-script matching that Ripjar's update targets, though the FCA's figures describe its own testing programme, not Ripjar's product.

The regulator also found wide variation in how quickly firms resolve alerts: roughly 44%–47% of firms resolved name- and payment-screening alerts within one working day on average, while around a fifth to a quarter of firms took three to five days to close them. The FCA said stronger alert-management frameworks tend to feature documented investigation rationales, quality assurance, clear escalation arrangements and periodic testing, and it warned explicitly that reliance on an external screening provider without sufficient internal oversight or assurance can cause delays or failures in escalating potential matches — a finding that applies squarely to any bank adopting a new AI-driven screening tool, Ripjar's included.

None of this amounts to a regulatory assessment of Ripjar's product. No evidence was found in the material reviewed that any UK regulator has tested or endorsed the specific updates announced on 24 September 2026.

Questions for bank buyers

Before relying on an automated screening upgrade, a bank compliance function would reasonably want to know: what the precision, recall and false-negative rate are, not just an aggregate accuracy figure; whether performance has been tested on non-Latin scripts and variant names representative of its own customer base; which decisions the system can make without an analyst, and which require human sign-off; how thresholds and automatic alert closures are governed and monitored for drift; whether closed alerts are sampled for quality assurance; and whether an independent party, rather than the vendor, has validated any of the claimed figures.

A wrongly closed or wrongly escalated alert is not a purely technical matter. Investigating and controlling the outcome of a sanctions or PEP match sits with the bank's own systems and procedures, whatever screening technology sits behind them.

What to watch next

Ripjar has not disclosed named launch customers, pricing, a deployment timetable or migration requirements for the updated ULTRA and Screening Assistant. No independent test results, model documentation or regulatory assessment of the specific 24 September 2026 updates were found in the material reviewed. Readers wanting the primary regulatory detail behind this article should go to the FCA's sanctions systems-and-controls findings and OFSI's financial sanctions guidance directly, rather than to any commercial intermediary.