Lloyds Banking Group is one of the UK’s largest financial services groups, serving around 27 million customers across retail, commercial, insurance and wealth management. The group reported pre-tax statutory profit in 2025 of £6.7bn on total income of £19.4bn, with capital gains of ~£3.9bn.
Lloyds Banking Group is transforming its operational architecture by incorporating AI as a core strategic lever. The company moved from an experimental pilot to a large-scale deployment.
AI is now a board-level priority for Lloyds. The group appointed former AWS data and AI leader Rohit Dhawan as group director of AI and advanced analytics in August 2024 to run a centralized AI center of excellence that brings together data science, ML engineering, behavioral science, and AI ethics under a single mandate.
Management revealed that more than 50 generative AI solutions will be in production in 2025, contributing around £50m of value, and the group aims to achieve more than £100m of AI value in 2026. The common technology pillar is the Google Cloud Vertex AI platform, which the group transitioned to in 2024 and currently supports over 300 data scientists and at least 18 GenAI systems in production.
This article examines two internal AI use cases that illustrate how Lloyds applies AI to its operations.
- Generative AI at scale for cutting-edge knowledge retrieval: Modernizing information access with GenAI reduces manual search latency from nearly a minute to seconds, empowering frontline staff to resolve customer inquiries on the first touch, and reducing total operational processing time.
- Real-time machine learning for debit card fraud: The move from a rules-based engine to adaptive ML-based scoring enables sub-second transaction decisions, enabling the Group to outperform evolving fraud typologies while minimizing friction for valid customer payments.
Lloyds Client Operations supports 27 million customers across banking, insurance and wealth brands. Previously, frontline staff were viewing 13,000 internal articles during live calls, creating both operational friction and FCA compliance risks. Lloyds has publicly stated that fixing this inefficiency is one of the key reasons it invested in generative AI in 2025.
The relevance is both operational and regulatory. The FCA’s AI guidance requires explainability and auditability, so tools used during customer interactions must rely on approved internal sources. At the same time, OECD research shows that generative AI delivers the greatest productivity gains for short-tenure knowledge workers, the precise profile of front-line customer relations staff.
Lloyd’s implemented Athena to address this issue. Athena runs on the group’s Vertex AI-based ML and GenAI platform and pulls answers from approximately 13,000 certified internal knowledge articles rather than from the open web.
Although Lloyds has not disclosed the specific underlying model behind Athena, the group confirmed that its platform supports RAG (Search Augmentation Generation) for internal content stores, with centralized logging and guardrails enforced at the platform layer.
How Lloyds can meet the FCA’s expectations around explainability and data location by grounding Athena’s responses on approved internal content. The operating rules for regulated institutions are simple. GenAI assistants should never reference customer information from sources that the company cannot audit line by line.
Athena transforms frontline workflows in four practical ways:
- Instead of searching for the title of a document, colleagues ask questions in natural language during the call and receive synthesized answers.
- Answers surface as evidence-based information, so colleagues can see authorized sources before speaking with customers.
- Decisions that previously required escalation to product or policy experts can now be resolved with the first touch.
- Usage and outcome signals are collected centrally, allowing the AI Center of Excellence to prioritize which knowledge domains to expand next.
Athena is Lloyd’s first large-scale GenAI deployment and is already past the pilot stage. About the group have disclosed Specific result data:
- By mid-2025, 21,000 employees will be using Athena in active workflows, and rollout will continue across customer operations.
- There will be 2.1 million searches in the first half of 2025, and the group expects around 40 million searches to occur by the end of the year.
- Average search time decreased from 59 seconds to 20 seconds (66% reduction).
- The phone banking team alone saves an estimated 4,000 hours a year, which directly translates into reduced wait times for customers.
Lloyds believes a significant share of the £50m of GenAI value in 2025 will come from tools comparable to Athena, and has confirmed that its AI-powered financial assistant for retail customers will launch on a mobile app in 2026, with the same platform underpinnings extended to the customer-facing side.
Card and payment fraud remains a major cost and control challenge for UK retail banking. According to UK Finance, criminals stole £1.17 billion through authorized and unauthorized fraud in 2024. UK-issued card fraud totaled £572.6m, with fraud incidents increasing by 14% to 3.13m.
Rule-based fraud systems amplify the second problem. Wedge and his colleagues used real bank data to demonstrate that only about one in five transactions flagged as fraudulent was actually fraudulent, and that approximately one in six customers rejected valid transactions in the previous year.
A 2025 systematic review of ML for digital banking fraud detection confirms that imbalance-aware, cost-focused ML approaches consistently outperform static rules in both recall and false positive reduction, making ML-based scoring the operational standard at Lloyds scale.
The Group operates a proprietary machine learning platform, the Dynamic Risk Engine (DRE), which scores all debit card authorizations in real time.
Writing in Lloyds Banking Group’s engineering publication on AI, Lloyds engineers explain that DRE consumes signals from past transactions, devices, and behaviors, has response times as short as 0.01 seconds per transaction, and is invisible to the customer at the point of sale.
DRE features a dynamic risk assessment layer co-built with Google that examines approximately 900 million indicators of financial crime each month, voice fraud detection on incoming calls, and a global correlation engine for cross-channel cybersecurity analysis. Best practice for publishers of similar size: Treat the rules-based engine as a complement to refinement rather than the primary decision-making layer.
For fraud analysts and the customers they protect, DRE creates three operational shifts.
- All approvals are scored in real-time and routed to approval, appeal (step-up authorization or out-of-band contact), or rejection, ultimately eliminating the latency of manual review from the approval path.
- New fraud typologies are learned and deployed through retraining cycles, rather than by human analysts writing new rules, reducing the lag between the emergence of new fraud and banks’ detection coverage.
- Analyst decisions and customer dispute outcomes are fed back into the training data, so the model continually improves rather than declines as fraud tactics change.
DRE is the most mature AI deployment in the Lloyd’s fraud stack and has been deployed at UK scale in production. Based on Lloyd’s own engineering disclosures and sector benchmarks.
- According to the group’s engineering team, DRE processes more debit card transactions every day than any other bank in the UK.
- Inference latency of approximately 0.01 seconds per transaction enables real-time approval decisions with no visible customer friction.
- UK Treasury estimates that across the industry, banks prevented a combined £1.45bn of fraud in 2024. This means the critical operating margin lies in the layer Lloyds has built: real-time detection and scoring.
- Lloyd’s is extending its stack into next-generation detection. In April 2026, the group completed a nine-month experiment with IBM that applied quantum algorithms to money mule identification in transaction graphs using anonymized data on a 156-qubit quantum system.
This article highlights some strategic insights from Lloyds Banking Group’s AI initiatives.
- Centralize your platform and decentralize your use cases: By consolidating into a single ML and GenAI platform (Vertex AI) and allowing business units to own their individual use cases, Lloyds was able to move over 50 GenAI solutions and 80 ML use cases into production in less than a year without increasing vendor sprawl or governance debt.
- Manage not only models but also sources: Athena’s value depends less on model selection and more on all answers based on an approved corpus of 13,000 articles. For regulated entities, it is control over the source material that makes GenAI accountable and auditable under the FCA’s AI approach.
- Conflicts at the authorization layer. As fraud prevention in the UK now outweighs total fraud losses, marginal benefits are shifting from ex-post review to sub-second decision-making at the time of authorization. Dynamic Risk Engine is built for that layer, so Lloyds is prioritizing investment there and is already piloting next-generation (quantum) enhancements.
