Fraud Detection AI at Barclays, HSBC and Standard Chartered: Only One of Them Has Actually Proven It Works

Key Takeaways

• Barclays, HSBC and Standard Chartered, three of the UK’s largest banking groups, have all built AI-driven fraud and financial-crime detection systems, but the quality and independence of what each has actually disclosed about those systems varies enormously, and a UK C-suite comparing them should notice the gap in evidence as much as the technology itself.
• HSBC’s is the best-documented of the three. A dated, on-the-record press release from Google Cloud, quoting HSBC’s own Group Head of Financial Crime Risk and Compliance, states that its AML AI system detects 2 to 4 times more true positive risk than HSBC’s previous approach, while cutting alert volumes by more than 60% and reducing processing time for billions of transactions from several weeks to a few days. HSBC also won Celent’s Model Risk Manager of the Year award in 2023 after deploying it, a genuine third-party validation rather than a vendor’s own claim.
• Barclays’ account, published as a named customer story by database vendor Aerospike, is about infrastructure performance rather than AI accuracy specifically: its fraud platform now handles more than 10 million transactions a day, four times its previous throughput, at under 100 milliseconds latency for 99.99% of requests, an 80% latency reduction, while the underlying fraud dataset grew from 3 terabytes to more than 30 terabytes in three years without a system redesign. Those are real, named-executive-attributed figures about the data layer that makes Barclays’ fraud models workable at scale, not a disclosed accuracy or false-positive figure for the models themselves.
• Standard Chartered was the earliest mover of the three, becoming the first global bank to deploy AI vendor Silent Eight’s platform in its sanctions compliance operations in July 2018, confirmed in Standard Chartered’s own press release. But that release contains no performance figures at all. The widely cited claim that Silent Eight’s platform cuts manual investigation time by up to 70% is the vendor’s own general marketing figure, not one Standard Chartered has itself confirmed applies to its deployment, and it’s worth reading with that distinction in mind. Standard Chartered’s venture arm, SC Ventures, also invested in Silent Eight’s 2019 funding round, making the bank both a customer and a financial backer of its own AI supplier.
• None of this means Standard Chartered’s or Barclays’ systems work less well than HSBC’s, only that HSBC has published independently verifiable outcomes and the other two, so far, largely haven’t. For a UK C-suite benchmarking its own fraud AI investment, the more useful comparison isn’t which bank’s number sounds most impressive. It’s which bank’s claims are actually checkable, and HSBC is currently the only one of the three that clears that bar.

Ask three UK banking groups how their AI fraud detection performs, and you’ll get three very different kinds of answer:

One with a dated press release, a named executive and an independent industry award behind it.

One told through a database vendor’s customer story, rich with named-executive infrastructure numbers but silent on the fraud model’s actual accuracy, and one confirmed only in outline, its most-quoted performance figure belonging to the vendor rather than the bank.

All three are real deployments. Only one of them is genuinely easy to independently verify.

HSBC: the best-documented of the three

HSBC’s system, built on Google Cloud’s AML AI product and internally named Dynamic Risk Assessment, is the most thoroughly disclosed of the three banks’ AI financial-crime systems.

According to a press release Google Cloud published on 21 June 2023, the system detects 2 to 4 times more true positive risk than HSBC’s prior transaction-monitoring approach, while cutting overall alert volumes, the flagged transactions that then require a human analyst’s time, by more than 60%.

HSBC monitors around 4 billion transactions a year across its customer base, and Google Cloud’s release says the system cut the batch processing time needed to analyse billions of transactions across millions of accounts from several weeks down to a few days. “Google’s models are already demonstrating the tremendous potential of machine learning to transform anti-financial crime efforts in the industry at large,” said Jennifer Calvery, HSBC’s Group Head of Financial Crime Risk and Compliance, in the release.

HSBC went on to receive Celent’s Model Risk Manager of the Year award in 2023 after adopting the system as its primary transaction-monitoring approach in key markets, an independent industry recognition that sits alongside, rather than merely repeating, the bank’s own figures.

Barclays: real numbers, but about the data layer, not the model

Barclays’ publicly available account of its fraud AI comes from a different kind of source: a named customer case study published by Aerospike, the real-time database vendor whose technology underpins Barclays’ fraud platform.

Dheeraj Mudgil, Barclays’ VP Enterprise Fraud Architect, is quoted directly: the platform gave Barclays “quick access to large datasets with one hop to the data from the client and the predictability we needed for our fraud platform.” The figures in that case study are concrete and specific: Barclays’ fraud platform now processes more than 10 million transactions a day, four times the throughput of its previous system, at under 100 milliseconds’ latency for 99.99% of transactions, an 80% reduction in latency compared with the prior architecture.

The underlying fraud dataset grew from 3 terabytes to more than 30 terabytes over three years without requiring a redesign of the system’s architecture or caching layer. Aerospike’s case study says the resulting consolidation reduced stand-in processing and improved fraud outcomes by lowering both false positives and false negatives, though it gives no specific percentage for either.

What’s worth being precise about is what this case study actually is: a real-time data infrastructure story, told by the database company that supplies that infrastructure, about the plumbing that lets Barclays’ fraud detection models run fast enough and on enough data to be useful at scale. It says nothing, in the material available, about how accurate Barclays’ actual fraud-detection models are, what algorithms they use, or how their outputs compare with the system they replaced. That’s a materially different, narrower kind of evidence than HSBC’s, even though both are, in their own ways, genuine and specific.

Standard Chartered: first to deploy, least to disclose

Standard Chartered has the longest track record of the three. On 9 July 2018, the bank announced it had become the first global financial institution to deploy AI vendor Silent Eight’s platform, which applies machine learning and natural language processing to sanctions and watchlist name-screening, matching customer names against watchlists using fuzzy logic and learning from the historical decisions of Standard Chartered’s own analysts. “I am very pleased with this investment which is contributing significantly to the Bank’s innovation agenda,” said Markus Schulz, Standard Chartered’s Global Head of FCC Controls, in the announcement, adding that the tool was meant to “empower our teams to make more effective and efficient decisions, complementing the core monitoring and screening platforms that we have already established.” Standard Chartered’s own press release contains no performance figures at all, no time saving, no accuracy improvement, no volume figure.

The number most often attached to this story, that Silent Eight’s platform cuts manual investigation time by up to 70%, comes from Silent Eight’s own marketing material, not from a Standard Chartered disclosure confirming that figure applies to its specific deployment. That distinction matters and is worth stating plainly rather than letting the two blur together, the way most retellings of this story do. There’s also a governance detail worth a UK C-suite’s attention: Standard Chartered’s corporate venture arm, SC Ventures, participated in Silent Eight’s $6.2 million Series A funding round in November 2019, more than a year after the original deployment. That makes Standard Chartered simultaneously a customer of and a financial stakeholder in the company supplying part of its financial-crime AI, a relationship that isn’t inherently improper, corporate venture arms invest in vendors they use fairly often, but is exactly the kind of detail a board should want surfaced and understood, not left implicit.

What a UK C-suite should actually take from this

All three of these are genuine, multi-year AI deployments at major UK banking groups, and none of the gaps described here are evidence that Barclays’ or Standard Chartered’s systems perform worse than HSBC’s. What differs, starkly, is the quality of publicly available evidence behind each one. HSBC’s account comes with a dated release, a named senior compliance executive, specific before-and-after figures, and an independent industry award. Barclays’ comes with real, named-executive-attributed numbers, but about data infrastructure throughput rather than fraud-model accuracy, published by the vendor supplying that infrastructure. Standard Chartered’s comes with a confirmed deployment date and a named executive’s general endorsement, but its most-cited performance figure belongs to the vendor’s marketing, not the bank’s own disclosure. For a board benchmarking its own AI fraud investment against any of the three, the useful exercise isn’t ranking whose numbers sound biggest. It’s asking, of your own programme, whether the evidence you’d eventually publish about it would meet the HSBC standard, an independently checkable, dated, specifically-attributed outcome, or whether it would look more like the other two: real, but resting on someone else’s word.