<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by AI at Lloyds Banking Group on Medium]]></title>
        <description><![CDATA[Stories by AI at Lloyds Banking Group on Medium]]></description>
        <link>https://medium.com/@CayleighODwyer?source=rss-4c2e7038b290------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*MjjwNnE8IWBXq-CXDJrQuQ.jpeg</url>
            <title>Stories by AI at Lloyds Banking Group on Medium</title>
            <link>https://medium.com/@CayleighODwyer?source=rss-4c2e7038b290------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 02:12:13 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@CayleighODwyer/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[VERGE: Building Better Multi-Hop RAG Evaluation for Enterprise Knowledge]]></title>
            <link>https://medium.com/ai-at-lloyds-banking-group/verge-building-better-multi-hop-rag-evaluation-for-enterprise-knowledge-de2ac7689adb?source=rss-4c2e7038b290------2</link>
            <guid isPermaLink="false">https://medium.com/p/de2ac7689adb</guid>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[retrieval-augmented-gen]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-evaluation]]></category>
            <dc:creator><![CDATA[AI at Lloyds Banking Group]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 11:56:01 GMT</pubDate>
            <atom:updated>2026-11-04T11:56:01.414Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Is your RAG really as good as you think it is?</em></p><h3><strong>The Summary</strong></h3><p>VERGE improves enterprise Retrieval-Augmented Generation (RAG) evaluation by verifying multi-hop questions during generation, producing harder, cleaner benchmarks.</p><p>Its pipeline combines evidence selection, distractor design, automated verification, and refinement or rejection.</p><ul><li>From 11,052 candidates, VERGE retained 2,854 verified questions, with a 26% acceptance rate across five domains.</li><li>Human evaluation reported 97.0% QA integrity and stronger distractor quality than the comparison method.</li><li>The resulting benchmarks better distinguish retrieval benefits, exposing performance gaps between closed-book and oracle settings that were 5.2 to 7.3 percentage points larger across tested models.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KVSCIcFa_RIBAfg0G9h-iQ.png" /><figcaption>Futuristic AI evaluation pipeline showing enterprise data flowing through retrieval, question generation, automated verification and refinement into a verified dataset, with quality and performance analytics.</figcaption></figure><h3>Setting the Scene</h3><p>Existing RAG test set generation methods such as RAGAs struggle to generate a high quality test set. We propose an alternative approach to generating a high quality test set to evaluate RAG systems on domain-specific knowledge bases through VERGE, our verification-enhanced test set generation pipeline. Instead of generating evaluation questions first and checking quality later like existing methods, VERGE validates each question during generation to confirm that it truly requires multi-hop reasoning and that the answer is supported by the provided context.</p><p>From 11,052 candidates across six domains, VERGE yields 2,854 verified multiple-choice items targeting aggregation- and comparison-style synthesis over topically coherent evidence. A human annotation study confirms strong alignment between the automated verifier and expert judgement. In practice, this method enables teams building enterprise RAG to benchmark their RAG applications against harder, cleaner, more discriminative tests that better reflect real-world knowledge bases.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/576/1*5SreIaibPWM7QmZ5BpJMZg.png" /><figcaption>Diagram showing the VERGE pipeline transforming source documents into candidate multi-hop questions, using an LLM verification agent to refine, reject, or approve questions for a verified evaluation dataset.</figcaption></figure><h3><strong>The Challenge</strong></h3><p>Enterprise RAG systems are rarely judged on the kind of questions they actually face in production. Public benchmarks are useful for comparing models, but they are often built on fixed document collections and do not reflect the structure, terminology, or reasoning patterns found in specialist knowledge bases. That creates an evaluation gap: a system can look strong on public benchmarks but still struggle when combining evidence across multiple internal documents.</p><p>The current technical problem is not just a question-answer (QA) pair generation but also quality control. Multi-hop evaluation only works if the question genuinely requires information from multiple chunks; and not if a model can guess the answer from a single passage, general world knowledge, or weak distractors.</p><p>In practice, that is hard to guarantee at scale:</p><ul><li>Existing generation methods often verify quality after the dataset has already been produced.</li><li>That means weak questions still consume compute and manual review effort.</li><li>Distractors are often too easy, which inflates benchmark scores.</li><li>Public benchmarks do not tell much about how a system will perform on a custom enterprise corpus.</li></ul><p>For organisations investing in AI-powered knowledge workflows, this matters. Better evaluation means better solution selection, better retriever tuning, and clearer evidence on whether a RAG system is actually using the right context.</p><h3><strong>Our Solution</strong></h3><p>We built VERGE: a dataset generation pipeline that puts verification inside the generation loop. The idea is: Generate a candidate question, verify whether it meets quality criteria, and if it does not, provide targeted feedback and refine where needed. If it still fails, discard it. This solution improves the quality of generated QA pairs compared to the existing methods, and the generated test set has better differentiability.</p><ul><li>QA-pair generation: create multiple-choice questions with one correct answer and three purpose-built distractors.</li><li>Verification: check ‘Complete Multi-Hop’ and ‘QA Integrity’ conditions.</li><li>Iterative refinement: if a question fails, provide feedback on unused chunks, weak reasoning structure, or unsupported answers.</li><li>Final filtering: discard malformed or low-quality items.</li></ul><p>The design principles behind the solution are practical:</p><ul><li>Verifier is integrated into the generation loop to improve the QA quality.</li><li>Optimise for evidence synthesis across documents.</li><li>Make distractors plausible enough to test reasoning.</li><li>Reduce verifier bias by using multiple generator models rather than a single model family.</li></ul><p>Tech stack and experimental setup:</p><ul><li>Generator models: Llama 3.2–3B, Gemma 2–9B, and Ministral-8B.</li><li>Verifier model: Gemma 2–9B, with pass-rate checks against Llama 3.2–2B and Ministral-8B showing less than 2% variance.</li><li>Retrieval configurations for downstream experiments: dense retrieval, sparse retrieval, hybrid reranking, oracle, and closed-book baselines.</li><li>Datasets spanning five domains: GovReport, HotpotQA, MultiFieldQA, 2WikiMultihopQA,SEC filings, and UK Covid-19 Inquiry document.</li></ul><p>This matters because it shifts evaluation from a one-off benchmark exercise to an adaptable capability. Teams can generate task-specific test sets for their own corpora and inspect where models fail, rather than relying on generic benchmarks that do not map well to enterprise use cases.</p><h3><strong>How It Works</strong></h3><p>VERGE starts by chunking a corpus into semantically coherent text segments of up to 4,000 characters. For each candidate test item, the system selects a seed chunk and retrieves related chunks to form a context set of between two and five chunks. That context becomes the evidence bundle from which a multi-hop question must be generated.</p><p>From there, the pipeline has four core stages:</p><ol><li>Candidate generation: the model proposes a question, four answer options, and the correct answer.</li><li>Distractor engineering: distractors are designed to exploit common failure modes such as partial information, contextual misalignment, and missed dependencies across chunks.</li><li>Verification: the verifier checks whether the answer truly needs multiple chunks and whether the correct answer is fully supported by the provided evidence.</li><li>Refinement or rejection: failed items are revised using verifier feedback or removed from the dataset.</li></ol><p>This architecture improves scalability because the verifier does more than label pass or fail. It produces actionable feedback, which turns quality assurance into a controlled loop rather than a static filter. In the final dataset, 267 questions were successfully rescued through refinement, and these tended to be more structurally complex than questions that passed on the first attempt.</p><p>The results shows that this approach helped produce more discriminative datasets: for example, the performance gap between closed-book and oracle settings was 5.2 to 7.3 percentage points in accuracy larger across tested models. That indicates that the benchmark can defferentiate whether a model was genuinely benefiting from retrieval.</p><h3><strong>Lessons Learned</strong></h3><ul><li>In-process verification materially improves dataset quality. Human review rated VERGE higher than the comparison method for both QA integrity and distractor quality.</li><li>Generating multi-hop questions is harder than preserving factual consistency. The Complete Multi-Hop criterion filtered out far more candidates than QA Integrity did.</li><li>Benchmark design shapes what you learn. A more discriminative dataset gives clearer signals about retriever quality, context use, and model limitations.</li></ul><h3><strong>Conclusion</strong></h3><p>We present VERGE, which integrates a quality control for multi-chunk RAG evaluation dataset creation inside the generation loop. An LLM-based verifier targets an operational multi-chunk criterion and QA integrity at generation time, enabling scalable curation with minimal human intervention. Verification yields a more retrieval-discriminative instrument than existing methods, and diagnosis-conditioned refinement adds yield in the structurally hardest stratum without sacrificing quality.</p><h3><strong>What’s Next</strong></h3><p>Next steps include:</p><ul><li>Package VERGE as a reusable internal evaluation toolkit for task-specific RAG assessments.</li><li>Connect benchmark outputs to retriever tuning and prompt evaluation workflows.</li></ul><h3><strong>Meet the Makers</strong></h3><p><a href="https://www.linkedin.com/in/kyosuke-morita-58329286/">Kyosuke Morita | LinkedIn</a>, Senior AI Reseacher at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/katie-adams-9267a445/">Katie Adams | LinkedIn</a>, Head of AI and Data Transformation at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/pedro-freiree/">Pedro Freire, Ph.D, MBA | LinkedIn</a>, Principle Lead Data and AI Scientist at Lloyds Banking Group</p><h3><strong>AI at Lloyds Banking Group</strong></h3><p>For more detail on our AI transformation, visit <a href="https://www.lloydsbankinggroup.com/who-we-are/group-overview/artificial-intelligence.html">Artificial Intelligence — Lloyds Banking Group</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=de2ac7689adb" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-at-lloyds-banking-group/verge-building-better-multi-hop-rag-evaluation-for-enterprise-knowledge-de2ac7689adb">VERGE: Building Better Multi-Hop RAG Evaluation for Enterprise Knowledge</a> was originally published in <a href="https://medium.com/ai-at-lloyds-banking-group">AI at Lloyds Banking Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building our Transaction Bedrock, an intelligent layer for every transaction]]></title>
            <link>https://medium.com/ai-at-lloyds-banking-group/building-our-transaction-bedrock-an-intelligent-layer-for-every-transaction-ce35a5fa1fcc?source=rss-4c2e7038b290------2</link>
            <guid isPermaLink="false">https://medium.com/p/ce35a5fa1fcc</guid>
            <category><![CDATA[responsible-ai]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[fintech]]></category>
            <dc:creator><![CDATA[AI at Lloyds Banking Group]]></dc:creator>
            <pubDate>Fri, 30 Oct 2026 11:17:23 GMT</pubDate>
            <atom:updated>2026-10-30T11:17:23.865Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Lloyds Banking Group is turning transaction data into a strategic intelligence layer through its Transaction Bedrock platform.</em></p><h3>The Summary</h3><p>Lloyds Banking Group is building Transaction Bedrock, an intelligence layer that transforms raw payment data into reliable, reusable insights. The platform combines machine learning, merchant intelligence, human oversight and continuous monitoring to support real-time banking and informed decisions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tZjloKhsOsolQbxSV8g4yQ.png" /><figcaption>Transaction data flowing through AI, machine learning and human oversight to generate banking insights.</figcaption></figure><h3>Setting the Scene</h3><p>Bedrock combines machine learning and merchant intelligence to make payment data more useful, reliable and actionable.</p><p>A custom Transformer forecasts roughly 2,200 time series using the previous 32 days, identifying spikes, dips, phase transitions, missed seasonality and other anomalies.</p><p>Automated prioritisation, reviewer feedback and monthly retraining create a continuous quality loop, detecting issues within a day and supporting classification of up to 22,000 transactions per second.</p><p>Bedrock delivers classified data through analytics, real-time streams, models and ready-made insights for customer experiences, fraud detection, financial decisions and business intelligence.</p><p>Every day, millions of payments flow through the financial system and Lloyds brands which accounts for around one in four card payments in UK, every single day.</p><p>A customer buys groceries. A small business pays a supplier. A family books a holiday. Someone starts a new gym membership, or receives their salary.</p><p>To most people, these appear as simple lines on a bank statement or their banking app.</p><p>But each transaction contains a story about customer behaviour, financial wellbeing, business activity, broader economic patterns.</p><p>The challenge is that these stories are often hidden within unstructured, inconsistent and incomplete data. Merchant names appear in multiple formats, new businesses emerge every day and payment references vary widely across channels and providers.</p><p>Without context, even the world’s richest transaction datasets are difficult to interpret.</p><h3><strong>Building a foundation for transaction intelligence</strong></h3><p>At Lloyds Banking Group, we’re building our “Transaction Bedrock” to solve that problem. How do analysts find all the transactions that relate to mortgages, or to Amazon? Was that Amazon payment a new chair being delivered to their home, or Amazon Prime? How do we verify a customer’s income if it comes from many sources? Is that income a salary, or a customer transferring themselves money from their savings account, or a friend paying them back for a holiday they booked? The raw data is not easy to translate into insight at first glance. We need a system to help us. Bedrock is Lloyds Banking Group’s strategic transaction intelligence platform.</p><p>Its mission is simple but powerful: Transform raw transaction data into meaningful, reusable intelligence that can power experiences, decisions and innovation across the organisation.</p><p>It’s crucial that we embark on this mission responsibly. From the outset, Bedrock has been designed with continuous monitoring, human oversight and feedback loops. Outputs can be reviewed, challenged and improved, helping maintain data quality and enabling appropriate use across the Group.</p><h3><strong>The Challenge with Transaction Data</strong></h3><p>Financial institutions like Lloyds generate vast quantities of transaction data every second. Historically, much of this has been used for operational purposes, moving money, reconciling accounts and supporting payment journeys.</p><p>As customer expectations have evolved, so has the opportunity to use this data more effectively.</p><p>Customers increasingly expect clearer insights into their spending and support to make informed financial decisions. Internal teams need deeper insight to support analytics, fraud detection and more personalised experiences.</p><p>None of this is possible without first understanding what individual transactions actually represent.</p><h3><strong>From Rules to Intelligence</strong></h3><p>Transaction classification at Lloyds has typically relied on a rules-based approach. These remain important.</p><p>However, the scale and complexity of payment ecosystems means that rules alone are no longer enough.</p><p>We combine established machine learning techniques with modern deep learning capabilities to improve transaction understanding, while keeping humans in the loop to maintain control and closely monitor our data.</p><h3><strong>Why data quality matters</strong></h3><p>Trust in any data product starts with data quality.</p><p>Misclassified or incomplete transaction data doesn’t just create technical issues, it directly shapes customer experiences and business outcomes. It can affect spending insights, subscription detection, financial wellbeing prompts, personalisation, and the confidence business teams have in the products they build.</p><p>At scale, these issues cannot be reliably identified through manual checks alone. Instead, we need a way to continuously monitor data and identify when something doesn’t look right.</p><p>This is where anomaly detection comes in.</p><h3><strong>Detecting the unexpected</strong></h3><p>Rather than manually checking millions of transactions, we monitor patterns in how transaction data behaves over time and flag where those patterns change unexpectedly.</p><p>To do this we aggregate classified transactions into a series of values over time, for example, daily transaction counts for major brands, categories and unique merchant patterns.</p><p>Each series has its own definition of “normal behaviour”.</p><p>We use machine learning trained to learn those patterns and predict what should happen next. When actual behaviour deviates from that expectation, it is flagged as a potential anomaly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/576/1*nzGUXAa3R8eQ4Kn5ABMhfw.png" /><figcaption>Transaction anomaly detection process showing ML forecasts identifying anomalies, followed by human review and bug resolution.</figcaption></figure><p>These anomalies can take several forms, including:</p><ul><li>Spikes: One-off increase in activity ,often reflecting real customer behaviour.</li><li>Dips: One-off drop, which indicates missing or delayed data.</li><li>Phase transitions: Sudden, permanent change in behaviour; like a step change in the transaction count or change in the seasonal behaviour. This is the anomaly we might see if a merchant stopped being classified entirely.</li><li>Missed seasonality: where expected seasonal patterns do not appear, like not seeing a spike in direct debits at the beginning of the month.</li><li>Each day, the self-learning model forecasts what every series should look like and compares those expectations with actual behaviour. Material deviations are identified as candidate anomalies and compared with previously detected events, to filter out expected variance. The model automatically prioritises higher-risk signals and suppresses known false positives such as bank holidays. The prioritised anomalies are surfaced for investigation by internal lab reviewers. Each anomaly is classified as a true or false positive, creating a continuous feedback loop that enables the model to learn and improve. Genuine issues are investigated and resolved.</li></ul><p>Without automated detection, bugs would go unnoticed until they impacted users, which might mean lying in wait for weeks or months. Now we can detect issues in just a single day.</p><h3><strong>Building for Real-Time Banking</strong></h3><p>Bedrock represents a shift from a batch-oriented capability to a platform that can support real-time transaction intelligence, allowing data to be used more dynamically. We see you Black Friday!</p><h3><strong>Thinking Like a Product</strong></h3><p>A key shift has been treating transaction intelligence as a product rather than a one off data-set.</p><p>Instead of a once and done classification data table, Bedrock provides a continuous improving data product that can be consumed in different ways: an analytics layer for insight generation, data streams for real-time decisioning, and for greater personalisation and modelling .</p><p>This approach helps deliver consistency while enabling flexibility for different teams and use cases.</p><h3><strong>Looking Ahead</strong></h3><p>Transaction data is one of the richest datasets available to financial institutions.</p><p>What began as a transaction classification platform is evolving into something much more valuable: an intelligence layer that helps us understand customers, merchants and financial behaviour at unprecedented depth and scale. The next step is integrating transaction insights and predicted features to enhance our agentic workflows.</p><p>The goal has never been simply to classify transactions, but on improving how they are understood and used, in ways that deliver meaningful value for our customers and colleagues.</p><h3>Meet the Makers</h3><p><a href="https://www.linkedin.com/in/david-garrick-024976225/">David Garrick | LinkedIn</a>, Senior Data Scientist at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/pratikbose/">Pratik Bose, CEng, MBA, MIET | LinkedIn</a>, Engineering Lead- Agentic &amp; AI at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/charlotte-nickels-7092139b/">Charlotte Nickels | LinkedIn</a>, AI Transformation at Lloyds Banking Group</p><h3>AI at Lloyds Banking Group</h3><p>For more detail on our AI transformation, visit <a href="https://www.lloydsbankinggroup.com/who-we-are/group-overview/artificial-intelligence.html">Artificial Intelligence — Lloyds Banking Group</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ce35a5fa1fcc" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-at-lloyds-banking-group/building-our-transaction-bedrock-an-intelligent-layer-for-every-transaction-ce35a5fa1fcc">Building our Transaction Bedrock, an intelligent layer for every transaction</a> was originally published in <a href="https://medium.com/ai-at-lloyds-banking-group">AI at Lloyds Banking Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Bridging Transformers and Rules-Based Logic for Scalable Document AI]]></title>
            <link>https://medium.com/ai-at-lloyds-banking-group/bridging-transformers-and-rules-based-logic-for-scalable-document-ai-4404b61d077a?source=rss-4c2e7038b290------2</link>
            <guid isPermaLink="false">https://medium.com/p/4404b61d077a</guid>
            <category><![CDATA[document-management]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[financial-services]]></category>
            <dc:creator><![CDATA[AI at Lloyds Banking Group]]></dc:creator>
            <pubDate>Fri, 23 Oct 2026 14:56:04 GMT</pubDate>
            <atom:updated>2026-10-23T14:56:04.158Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How a hybrid approach combined layout-aware transformers with deterministic spatial rules to extract structured data from millions of legacy finance documents</em></p><h3>The Summary</h3><p>A hybrid document AI system combined layout-aware transformers with deterministic spatial rules for scalable, auditable finance-document extraction. The approach targeted structured data recovery without expensive large language models or massive labeled datasets.</p><p>It paired a compact transformer with explicit validation and scalable CPU-based processing.</p><ul><li>Trained on 400 annotated documents, the system extracted 33 fields and achieved 99.2% key-value F1 and 98.8% tabular F1 on held-out data.</li><li>It processed around 3 million pages in 28 hours on CPU-only infrastructure, with 99.87% passing automated consistency checks; this was not an accuracy claim.</li><li>The hybrid approach cost about 250 times less per page than the compared frontier model and retained human review for failed validations.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rYmv2wTVR3k9DX59MgjJBw.png" /><figcaption>Illustration of finance documents flowing through an AI extraction system, with automated validation routing successful data to storage and failed checks to human review.</figcaption></figure><h3><strong>When bigger AI models aren’t necessarily the better answer</strong></h3><p>When tackling a difficult document-extraction problem, it can be tempting to reach for a bigger, more powerful AI model. But choosing the right approach means looking beyond model capability to cost, repeatability, auditability and the handling of sensitive data.</p><p>What if part of the intelligence you need is already in the document itself, in its layout, structure and the relationships between information on the page?</p><p>We explored that question while working with a large archive of historical motor finance records. Created by older mainframe systems, these fixed-format digital documents were designed for people to read, not machines to process. Repeated headers, multi-page tables and subtle variations in layout made reliable extraction challenging.</p><p>Simple templates were too brittle to handle those variations. A general-purpose generative approach offered greater flexibility, but at this scale it also brought trade-offs around cost, repeatability and validation. We wanted to see whether we could get the best of both worlds: learned understanding where interpretation was needed, and explicit rules where the documents themselves provided reliable structure.</p><p>Our answer was a hybrid approach. A compact, layout-aware AI model interpreted the text and its position on the page, while deterministic spatial rules captured relationships within the document structure. Accounting and syntactic checks then provided an additional validation layer.</p><p>The result was an approach designed for high-volume batch processing, with predictable infrastructure requirements and outputs that could be traced, tested and validated against explicit rules. Rather than asking how large a model we could use, we asked how much of the problem actually needed a large model in the first place.</p><h3><strong>The Challenge</strong></h3><p>Not all historical financial records sit neatly in modern databases. Some were created by older mainframe systems and stored as fixed-format digital text, designed to be read by people rather than processed as structured data. In this case, we were working with a large archive of historical motor finance closure and deletion statements.</p><p>These documents were designed for people to read, not machines to ingest. They included repeated headers, variable page counts, multi-page tables, and small layout shifts. That made simple template rules too brittle, but a pure generative AI solution would have been costly to run at scale and difficult to audit.</p><p>The challenge wasn’t simply turning old records into digital data. It was reconstructing account information reliably enough to support downstream compliance, remediation and analytics. A missing transaction row, incorrect field or broken relationship between a label and its value could undermine confidence in the result.</p><p>We therefore needed a solution that could scale to millions of pages, cope with moderate structural variation, and remain transparent enough for regulated environments.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/576/1*NI9zXZCxMp0dQOBo0fQqig.png" /><figcaption>Diagram showing a hybrid document AI pipeline that combines layout-aware extraction, spatial rules, validation and human oversight to turn finance documents into structured data.</figcaption></figure><h3><strong>Our Solution</strong></h3><p>We built a hybrid extraction pipeline that combines learned document understanding with explicit geometric reasoning.</p><p>At the centre of the solution is LiLT (Language-independent Layout Transformer), a compact layout-aware transformer fine-tuned on just 400 annotated documents. LiLT learns from both the text and the position of words on the page, which is essential when meaning depends on structure as much as language.</p><p>Alongside the model, we introduced a Spatially Aware Prior, or SAP. In plain terms, SAP is a rules layer that understands where a value should appear relative to a label or a table header. Instead of hard-coded zones, it uses soft spatial constraints so it can tolerate small layout drift while staying deterministic.</p><p>The final prediction combines both signals: the AI model contributes its understanding of the text and context, while SAP reinforces what the document’s structure tells us should be true.</p><h4>Core design principles included:</h4><ul><li>Use lightweight models where document structure is consistent and business rules are strong.</li><li>Preserve auditability by making structural constraints explicit.</li><li>Avoid unnecessary OCR when the source is already digital.</li><li>Optimise for throughput on commodity infrastructure rather than GPU-heavy deployment.</li><li>Outputs that failed the defined consistency checks were not treated as automatically validated. They could instead be separated for additional human oversight under the relevant operational controls.</li></ul><p>The broader delivery pattern mattered as well. Pre-processing, inference, rule execution, and output consolidation were decomposed into containerised stages orchestrated through Kubeflow on Kubernetes, making the workflow easier to scale and operate.</p><h3><strong>How It Works</strong></h3><p>At a high level, the approach is straightforward: understand what the text says, use where it appears on the page as an additional signal, combine those two sources of information, and then validate that the result makes sense.</p><p>LiLT provides the learned understanding of text and layout, while SAP adds explicit knowledge about the spatial relationships we expect to see in these documents. Their predictions are combined before a final set of deterministic checks tests the extracted data for consistency</p><h3><strong>The Technical Detail</strong></h3><p>Each digital text page is first converted into word-level elements with coordinates. Because the source documents use a monospaced format, those coordinates can be derived directly from the text stream rather than estimated using Optical Character Recognition (OCR), reducing both compute overhead and the potential for recognition noise.</p><p>From there, the pipeline follows two complementary paths:</p><ul><li>The LiLT path classifies each document element into one of the target schema labels using both textual and layout context.</li><li>The SAP path applies deterministic spatial constraints to model expected relationships within the document, such as a value appearing to the right of a field label or transaction cells aligning vertically beneath a table header.</li></ul><p>This combination is particularly useful for long transactional tables. Transformer context-window limits can require a table to be split into chunks, potentially separating rows from the header information that explains what each column represents. SAP compensates by following the spatial relationships between aligned cells and rows across the document, preserving structural information that may otherwise be lost when the content is chunked.</p><p>The two signals are then fused using a Product-of-Experts approach. The transformer contributes semantic and contextual confidence, while SAP contributes the spatial prior. In practical terms, strong structural evidence can help stabilise an uncertain model prediction, while the model’s learned understanding can constrain spatial rules in parts of the document where geometry alone could be ambiguous.</p><p>Finally, the extracted data passes through a deterministic validation layer. This checks:</p><ul><li>lexical and syntactic validity for fields such as identifiers and transaction codes</li><li>arithmetic consistency across extracted values</li><li>accounting integrity across running balances in transaction histories.</li></ul><p>These checks provide an additional form of evidence about whether an extraction holds together. Outputs that fail the defined consistency checks are not treated as automatically validated and can instead be separated for additional human oversight under the relevant operational controls</p><h3><strong>How we evaluated it</strong></h3><p>We assessed the approach in two ways. First, a held-out evaluation set allowed us to compare extraction quality across LiLT, SAP, the combined approach and an LLM baseline using F1 and exact-match measures. Second, we assessed how the hybrid approach performed when applied to the live use case at scale, including throughput and automated validation pass-through.</p><p>Keeping these two assessments separate matters: the first measures extraction performance against known answers, while the second tells us how the system behaved operationally at scale.</p><h3><strong>Results</strong></h3><p>The hybrid LiLT + SAP approach produced the strongest overall performance in the held-out evaluation, with its clearest advantage appearing in the more structurally complex tabular extraction task.</p><p><strong>Structured extraction results</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*T6RS3M7wbXARreZOC632pA.png" /><figcaption>Comparison table showing document extraction performance, with LiLT + SAP achieving the highest scores across key-value and tabular metrics.</figcaption></figure><p>These results come from the <strong>held-out evaluation set</strong> and show the value of combining learned document understanding with deterministic spatial reasoning.</p><h3><strong>From evaluation to a live use case</strong></h3><p>The approach was also applied at scale to the live business use case, processing around 3 million pages in 28 hours using CPU-only infrastructure. 99.87% of outputs passed the automated validation checks covering field validity, arithmetic consistency and transaction-table integrity.</p><p>That distinction matters: the extraction scores above measure performance on the controlled evaluation set, while the 99.87% figure is a validation pass-through rate for the live use case, not a claim of 99.87% extraction accuracy</p><h3><strong>Lessons Learned</strong></h3><p>First, structure matters more than many teams expect. For highly regular documents, alignment and page geometry can carry as much signal as the words themselves.</p><p>Second, the right approach depends on the characteristics of the workload. General-purpose multimodal models provide a flexible starting point for many document-extraction tasks. However, where documents are highly regular, volumes are large, fields are stable and validation rules are well understood, a specialist model combined with deterministic rules can still be a strong fit. In this case, the approach provided strong repeatability and auditability, while supporting controlled data handling and predictable processing costs.</p><p>There were also limitations. SAP uses tolerance thresholds to determine how much variation between the expected and extracted document structure can be accepted before it is treated as an error. In our evaluation, these thresholds were tuned against a relatively consistent set of digitally generated documents. More varied real-world scans, with differences in layout, typography and scan quality, would therefore require the thresholds to be revalidated and potentially adjusted.</p><h3><strong>What’s Next</strong></h3><p>The next step is to test the approach on noisier OCR-derived documents, where bounding boxes are less stable and character recognition can drift.</p><p>There is also room to make SAP more adaptive by learning or auto-tuning spatial tolerances from the corpus rather than configuring them manually.</p><p>Longer term, the same pattern could be explored across other structured document archives, such as regulatory filings, statements, forms and government-style records. Rather than positioning the hybrid approach as an alternative to general-purpose models, the focus is on identifying where it is the right fit, particularly for high-volume workloads with reasonably regular document structures, stable extraction requirements and well-understood validation rules.</p><p>The broader lesson isn’t that smaller models are better than bigger ones. It’s that the best AI architecture starts with the shape of the problem. Where documents are regular, volumes are high and the rules for a good result are well understood, combining learned intelligence with explicit structure can still be a powerful approach.</p><h3><strong>Meet the Makers</strong></h3><p><a href="https://www.linkedin.com/in/ishant-aggarwal/">Ishant Aggarwal | LinkedIn</a>, Document AI Researcher in the Agentic &amp; AI Platform at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/natasha-sarin-aa733478/">Natasha Sarin | LinkedIn</a>, Agentic and AI Services at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/kejiong-li-phd-79894b25/">Kejiong Li, PhD | LinkedIn</a>, Lead AI and Data Science at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/yawen-zhang-964247226/">Yawen Zhang | LinkedIn</a>, Senior AI Engineer at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/yue-fan/">Yue (Yuri) Fan | LinkedIn</a>, Data Scientist at Lloyds Banking Group</p><h3>AI at Lloyds Banking Group</h3><p>For more detail on our AI transformation, visit <a href="https://www.lloydsbankinggroup.com/who-we-are/group-overview/artificial-intelligence.html">Artificial Intelligence — Lloyds Banking Group</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4404b61d077a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-at-lloyds-banking-group/bridging-transformers-and-rules-based-logic-for-scalable-document-ai-4404b61d077a">Bridging Transformers and Rules-Based Logic for Scalable Document AI</a> was originally published in <a href="https://medium.com/ai-at-lloyds-banking-group">AI at Lloyds Banking Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gem AI: Transforming Help & Support in Gem Connect]]></title>
            <link>https://medium.com/ai-at-lloyds-banking-group/gem-ai-transforming-help-support-in-gem-connect-efb7e068ffc3?source=rss-4c2e7038b290------2</link>
            <guid isPermaLink="false">https://medium.com/p/efb7e068ffc3</guid>
            <category><![CDATA[ai-customer-support]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <category><![CDATA[retrieval-augmented-gen]]></category>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <category><![CDATA[conversational-ai]]></category>
            <dc:creator><![CDATA[AI at Lloyds Banking Group]]></dc:creator>
            <pubDate>Fri, 16 Oct 2026 11:28:52 GMT</pubDate>
            <atom:updated>2026-10-16T11:28:52.522Z</atom:updated>
            <content:encoded><![CDATA[<p><em>From question to answer, Gem AI is changing what great digital support can feel like.</em></p><h3>The Summary</h3><p>Gem AI transforms support in Gem Connect through conversational, knowledge-grounded assistance. The solution combines rapid delivery with governance and a reusable architecture.</p><ul><li>Eight-week delivery produced Commercial Banking’s first client-facing agentic AI use case, built by five engineers.</li><li>A retrieval-augmented generation architecture answers natural-language questions using approved support content, without exposing customer data to the model.</li><li>Planned expansion includes improved accuracy, broader automation, enhanced agentic capabilities, and additional platform use cases.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9E5Iw-mz15mJzfkTEPtQYQ.png" /><figcaption>Illustration of a person using an AI assistant that connects approved knowledge sources with secure retrieval, automation and scalable business support.</figcaption></figure><h3>Setting the Scene</h3><p>We built <strong>Gem AI - Smart Help &amp; Support, a conversational </strong>AI assistant embedded within Gem Connect.</p><p>This solution moves beyond static help pages by allowing clients to ask questions in plain English and receive responses grounded in approved Help &amp; Support content. It marks the <strong>first external Gen AI capability in Lloyds Banking Group Commercial Banking</strong>, built on GCP 2.0 and the Internal Developer Platform.</p><p>More than a new feature that makes the experience simpler, the project was delivered <strong>in just eight weeks</strong>, showing how we can combine speed, innovation and governance to unlock value at scale.</p><p>The result is a reusable foundation for future AI use cases across the platform.</p><h3><strong>The Challenge</strong></h3><p>Gem Connect supports Commercial Banking clients with their cash management and payment needs. Like many platforms, its Help &amp; Support experience relied heavily on:</p><ul><li>Static help pages</li><li>Navigation-heavy journeys</li><li>Manual support channels</li></ul><p>This created friction for clients who needed quick answers during critical workflows.</p><p>At the same time, internal support teams faced increasing demand driven by:</p><ul><li>Growing client volumes</li><li>Complex onboarding journeys</li><li>Rising expectations digital-first support experiences</li></ul><p>We wanted to answer a simple but powerful question:</p><p><em>“How do we move from static help content to intelligent, real-time support?”</em></p><p>The challenge aligned with broader objectives to improve the client experience, reduce operation effort and unlock value from AI capabilities already available in the Group.</p><h3><strong>Our Solution</strong></h3><p>We built<strong> Gem AI - Smart Help &amp; Support, </strong>a<strong> conversational AI assistant </strong>embedded directly within Gem Connect<strong>.</strong></p><p><strong>Core capabilities</strong></p><ul><li>Conversational, AI enabled Self-service support</li><li>Responses grounded in trusted knowledge sources</li></ul><p>The aim was to provide a more intuitive way for clients to access information already available within the platform, shifting from static pages to smart AI-driven answers.</p><h3><strong>Architecture and Design Principles</strong></h3><p>The solution was built using:</p><ul><li><strong>GCP 2.0 + Internal Developer Platform (IDP)</strong></li><li>A <strong>retrieval-augmented generation (RAG)</strong> architecture grounded in trusted Help &amp; Support content</li><li>A <strong>reusable multi-agent framework</strong> for future scalability</li></ul><h3><strong>Key Design Principles</strong></h3><ul><li><strong>Safe-by-design: </strong>No customer data exposed to the model</li><li><strong>Low-risk production MVP:</strong> Deliberately scoped to focus on approved knowledge sources</li><li><strong>Scalable foundation:</strong> Built for future agentic workflows</li><li><strong>Speed with governance:</strong> Rapid delivery without compromising controls</li></ul><h3><strong>Delivery Highlights</strong></h3><ul><li>Prototype delivered in around four weeks</li><li>Production-ready solution delivered in around eight weeks</li><li>Built by a <strong>small cross-functional team of five engineers</strong></li><li>Delivered as Commercial’s first client-facing GenAI RAG Agent, providing customers with AI-powered Help &amp; Support through intelligent retrieval and summarisation.</li></ul><p>This combination of speed, innovation and robust governance is a strong example of execution excellence from the IDEA Lab and BTB Gem team.</p><h3><strong>How It Works</strong></h3><p>At its core, Gem AI uses a <strong>retrieval-augmented AI approach</strong> to ensure responses are sourced from a knowledge base that’s:</p><ul><li>Accurate</li><li>Explainable</li><li>Grounded in trusted content</li></ul><h3><strong>High-Level Flow</strong></h3><ol><li><strong>User Query</strong></li></ol><ul><li>A client enters a natural language question in Gem Connect</li></ul><p><strong>2. Retrieval Layer</strong></p><ul><li>The system searches approved Help &amp; Support content and retrieves the most relevant information from trusted knowledge sources</li></ul><p><strong>3. AI Response Generation</strong></p><ul><li>The model generates a clear, concise, plain-English response using the retrieved content as context.</li></ul><p><strong>4. Response Delivery</strong></p><ul><li>The client receives an answer directly within the platform, reducing the need to search multiple static support pages.</li></ul><h3><strong>Why This Matters Technically</strong></h3><ul><li><strong>Low latency:</strong> In-memory retrieval enables fast responses</li><li><strong>Explainability:</strong> Answers are grounded in known knowledge sources</li><li><strong>Scalability:</strong> Multi-agent architecture supports future use cases</li><li><strong>Resilience:</strong> Built on strategic enterprise platforms (GCP 2.0 + IDP) This design ensures the solution is not just effective today, but <strong>future-ready</strong>.</li></ul><h3><strong>Lessons Learned</strong></h3><p>Delivering a first-of-its-kind LBG solution came with valuable insights:</p><h4><strong>What worked well</strong></h4><ul><li><strong>Lean teams accelerate delivery: </strong>small, focused teams can move fast</li><li><strong>Early alignment on governance is critical</strong>: embedding Responsible AI from the start reduced friction later</li><li><strong>Clear use case focus</strong> : narrowing the scope to Help &amp; Support enabled rapid MVP delivery</li></ul><h4><strong>Challenges we navigated</strong></h4><ul><li>Balancing <strong>speed vs governance requirements</strong></li><li>Navigating an evolving enterprise technology landscape while maintaining delivery momentum.</li><li>Ensuring accuracy and trust in AI-generated responses</li></ul><h3><strong>Key takeaway</strong></h3><p>Delivering AI at scale is not just about technology, it’s about <strong>combining engineering, governance, and collaboration effectively</strong>.</p><h3><strong>What’s Next</strong></h3><p>Gem AI is an important step in exploring how conversational AI can improve digital support experiences. Planned next steps include:</p><ul><li>Expanding agentic capabilities</li><li>Improving response accuracy and knowledge</li><li>Introducing automation across more support journeys</li><li>Scaling the framework to additional use cases across the platform</li></ul><p>Ultimately, Gem AI is evolving into a <strong>foundation for a broader AI-driven ecosystem within Gem</strong>.</p><h3><strong>Meet the Makers</strong></h3><p>This delivery was made possible through strong collaboration across teams:</p><p><a href="https://www.linkedin.com/in/meena-m-b06b3749/?isSelfProfile=false">Meena M | LinkedIn</a>, Senior Software Engineer at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/davidson-samuel-55442614/">Davidson Samuel | LinkedIn</a>, Head of AI, Business Transaction Banking at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/krishna-k-b-9b11bb235/">Krishna K B | LinkedIn</a>, Associate Software Engineer at Lloyds Technology Centre</p><p><a href="https://www.linkedin.com/in/akhil-sai-teja-reddy-90809110b/?isSelfProfile=false">Akhil Sai Teja Reddy | LinkedIn</a>, Business Transaction Banking at Lloyds Banking Group</p><h3><strong>AI at Lloyds Banking Group</strong></h3><p>For more detail on our AI transformation, visit <a href="https://www.lloydsbankinggroup.com/who-we-are/group-overview/artificial-intelligence.html">Artificial Intelligence — Lloyds Banking Group</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=efb7e068ffc3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-at-lloyds-banking-group/gem-ai-transforming-help-support-in-gem-connect-efb7e068ffc3">Gem AI: Transforming Help &amp; Support in Gem Connect</a> was originally published in <a href="https://medium.com/ai-at-lloyds-banking-group">AI at Lloyds Banking Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Referral to Real-Time Decisions: Using Responsible AI to Support Faster, More Explainable…]]></title>
            <link>https://medium.com/ai-at-lloyds-banking-group/from-referral-to-real-time-decisions-using-responsible-ai-to-support-faster-more-explainable-60f1d617f629?source=rss-4c2e7038b290------2</link>
            <guid isPermaLink="false">https://medium.com/p/60f1d617f629</guid>
            <category><![CDATA[insurance-underwriting]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <category><![CDATA[explainable-ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[AI at Lloyds Banking Group]]></dc:creator>
            <pubDate>Wed, 07 Oct 2026 15:02:25 GMT</pubDate>
            <atom:updated>2026-10-07T15:02:25.763Z</atom:updated>
            <content:encoded><![CDATA[<h3>From Referral to Real-Time Decisions: Using Responsible AI to Support Faster, More Explainable Underwriting</h3><p><em>What if AI could help customers get quicker insurance decisions without removing the human judgement that matters most?</em></p><h3>The Summary</h3><p>The Propensity to Approve model advances Lloyds Banking Group’s use of responsible AI to support faster, more explainable underwriting decisions for selected Scottish Widows protection applications. Its design combines historic application and underwriting data, CatBoost machine learning, and strong operational governance.</p><ul><li>The model produces a probability score indicating whether a referred application is likely to be accepted on standard terms, helping identify suitable cases for a faster pathway.</li><li>SHAP-based explainability, business guardrails, rigorous testing and human oversight ensure the model supports rather than replaces underwriting expertise.</li><li>The solution aims to improve customer journeys, reduce avoidable requests for medical evidence, increase consistency and allow underwriters to focus on more complex cases, while informing future responsible AI use across insurance.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cmf3ZojR7o5dGNS_O0WfgA.png" /><figcaption>A human expert oversees an AI system analysing insurance applications and guiding cases towards faster, protected underwriting decisions.</figcaption></figure><h3><strong>Setting the Scene</strong></h3><p>Life and health insurance underwriting helps customers access life and critical illness cover while ensuring risks are assessed fairly and consistently. While many applications receive an immediate decision through established underwriting rules, others require referral for more detailed review because of the information involved.</p><p>Within the Scottish Widows Protect adviser-led journey, these referrals rely on underwriting expertise to assess individual circumstances and support appropriate customer outcomes. However, they can also increase turnaround times and operational effort.</p><p>This created a clear opportunity: could data and machine learning help identify referred applications that are highly likely to be accepted at standard terms, while remaining explainable, governed and suitable for a regulated environment?</p><h3><strong>The Challenge: Speed, Confidence and Control</strong></h3><p>The business challenge was not simply to automate more decisions. In protection underwriting, speed only creates value if it is paired with confidence, consistency, and appropriate controls. We needed an approach that could help improve the customer journey for suitable cases, while respecting the role of underwriting expertise and the safeguards required to deliver fair customer outcomes in a highly regulated environment.</p><p>Several design questions shaped the work from the outset:</p><ul><li>How do we distinguish applications that are likely to be Standard Accept from those that genuinely require deeper review?</li><li>How do we introduce AI into an underwriting journey while maintaining appropriate governance, auditability, and accountability?</li><li>How do we test and monitor performance over time and ensure the model continues to support appropriate customer outcomes?</li></ul><p>These questions helped shape our model as a decision-support capability embedded within a controlled underwriting process, rather than a standalone algorithm operating in isolation.</p><h3><strong>Our Approach: A Propensity Model for Referred Applications</strong></h3><p>Our model focuses on Scottish Widows Protection applications submitted through adviser-only channels, specifically Life Insurance and Critical Illness products. The model is applied to referred applications, where the existing rules engine (which applies predefined underwriting rules) has not already produced an automatic decision and where there may be an opportunity to support a faster underwriting outcome.</p><p>At a high level, the model produces a probability score representing the likelihood that a referred application would be <em>Standard Accepted</em> (accepted at a standard premium rate with full scope coverage) by the underwriting team. It learns from historic application journeys, using patterns across policy, customer and application information alongside actual underwriting outcomes.</p><p>As the machine learning algorithm underpinning our model, we selected <em>CatBoost </em>[1], a tree-based learner well suited to datasets containing large amounts of categorical information. It uses a gradient-boosting approach, combining many decision trees, each optimised to improve the performance of the last, to distinguish Standard Accept outcomes from other underwriting decisions. CatBoost is designed to handle categorical inputs efficiently, such as responses to health and lifestyle questions, enabling us to capture signals from customer, policy and application data while minimising extensive manual data preparation.</p><p>To support robust evaluation, data was split into training, validation and test samples, preserving the proportions of Standard Accept outcomes and product groups across each dataset. The model was trained and tuned using the training and validation samples, while the test sample was reserved for independent performance assessment.</p><h3><strong>How It Works: In Practice</strong></h3><p>PTA fits into the existing underwriting journey by providing a propensity score for referred applications. Used within agreed operational guardrails, the score acts as a decision-support tool to help identify cases that may be suitable for a faster pathway while keeping interpretation and control within established business governance.</p><p>More complex scenarios, such as joint-life and multi-product applications, continue to be handled conservatively to reflect real underwriting practice.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O6hDI49IGlevX9cAgIvvCg.png" /><figcaption>Diagram showing referred applications scored by the PTA model, checked against governance guardrails, then routed to a faster pathway or standard underwriting journey before customer outcome.</figcaption></figure><h3><strong>Responsible AI by Design</strong></h3><p>Responsible AI shaped the solution from the beginning. Because the decisions can affect customer experience, business cost and operational workload, explainability, validation and control needed to be core design features.</p><p>We used SHAP-based explainability to help data scientists, underwriters and risk stakeholders understand model behaviour and support effective oversight. While explainability does not replace expert judgement or validation, it provides greater transparency into how the model operates.</p><p>We also introduced operational guardrails to avoid affecting cases with business-agreed high-risk disclosures. This ensures that the model is used only where appropriate and within its validated scope. Model risk review, data privacy, and ethical impact assessments supported the deployment of PTA as a governed decision-support capability within the existing underwriting process.</p><h3><strong>Impact Highlights: Better Journeys, Better Focus</strong></h3><p>The value of PTA is reflected across customers, colleagues, and the business. By helping identify suitable applications for a faster pathway, it has the potential to improve customer journeys, allow underwriters to focus on more complex cases, and support a more efficient underwriting process.</p><p>The objective is not to remove underwriters from decision-making, but to provide a controlled and explainable signal that supports faster, more consistent outcomes while maintaining human oversight.</p><h3><strong>Lessons Learned</strong></h3><p>Successful AI in underwriting depends on close collaboration between data science, underwriting, platform, risk, privacy and business teams. The model could not have been developed without the expertise of those who understand the customer journey, operational processes, and underwriting decisions.</p><p>Equally important is recognising that successful deployment extends beyond model performance. Explainability, governance and practical operational considerations all play a critical role in building confidence and ensuring AI is used responsibly over time.</p><h3><strong>What’s Next: New Use Cases &amp; Approaches</strong></h3><p>Building on the experience gained through PTA, we are continuing to explore how data and advanced analytics could responsibly support other parts of the insurance journey. Our focus remains on identifying opportunities that may improve customer and colleague outcomes while complementing established expertise and processes.</p><p>Any future applications will be developed and assessed within appropriate governance, testing and monitoring frameworks. Human judgement, explainability, and responsible use will remain central as we evaluate where these capabilities can add value.</p><h3><strong>References</strong></h3><p>[1] Prokhorenkova et al. (2018), <em>CatBoost: unbiased boosting with categorical features</em>. Advances in Neural Information Processing Systems, 31<em> </em>(NeurIPS 2018)<em>.</em></p><h3><strong>Meet the Makers</strong></h3><ul><li><a href="https://www.linkedin.com/in/omar-raf/">Omar Rafiq</a> — Head of AI at Lloyds Banking Group</li><li><a href="https://www.linkedin.com/in/sean-harkin-phd-22776674/">Sean Harkin</a> — Data Science &amp; AI Lead at Lloyds Banking Group</li><li><a href="https://www.linkedin.com/in/murdotraill/">Murdo Traill</a> — Senior Data &amp; AI Scientist at Lloyds Banking Group</li><li><a href="https://www.linkedin.com/in/richardsgzan/">Richards González</a> — Data &amp; AI Scientist at Lloyds Banking Group</li></ul><h3><strong>AI at Lloyds Banking Group</strong></h3><p>For more detail on our AI transformation, visit <a href="https://www.lloydsbankinggroup.com/who-we-are/group-overview/artificial-intelligence.html">Artificial Intelligence — Lloyds Banking Group</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=60f1d617f629" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-at-lloyds-banking-group/from-referral-to-real-time-decisions-using-responsible-ai-to-support-faster-more-explainable-60f1d617f629">From Referral to Real-Time Decisions: Using Responsible AI to Support Faster, More Explainable…</a> was originally published in <a href="https://medium.com/ai-at-lloyds-banking-group">AI at Lloyds Banking Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Platforms Don’t Scale Through Technology Alone]]></title>
            <link>https://medium.com/ai-at-lloyds-banking-group/ai-platforms-dont-scale-through-technology-alone-67e2566fdae9?source=rss-4c2e7038b290------2</link>
            <guid isPermaLink="false">https://medium.com/p/67e2566fdae9</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[platform-engineering]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[cloud-computing]]></category>
            <category><![CDATA[mlops]]></category>
            <dc:creator><![CDATA[AI at Lloyds Banking Group]]></dc:creator>
            <pubDate>Wed, 23 Sep 2026 16:01:02 GMT</pubDate>
            <atom:updated>2026-09-24T06:20:54.727Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Building an AI platform is one challenge. Scaling it so thousands of people can use it securely, reliably and with confidence is another entirely.</em></p><h3>The Summary</h3><p>Lloyds Banking Group launched Atlas in 2024 to give machine learning and AI teams a secure, governed route from development to production. By 2026, Atlas supported more than 1,400 users, over 140 use cases and more than 1,600 workbenches. As demand grew, the service model had to scale with the technology. Specialist ownership, automation and stronger operational controls reduced workbench provisioning from two days to one hour, project provisioning from three weeks to five minutes, and mean incident resolution time from 107 hours to eight hours. This article explains what changed and what comes next.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*91Ub6mENkkqj5o_4Kh1ynA.png" /><figcaption>Illustration of a scalable enterprise AI platform showing secure cloud infrastructure, specialist support teams, automation workflows, operational monitoring, governance controls and AI-powered insights working together to deliver reliable, trusted AI at scale.</figcaption></figure><h3>The Problem</h3><p>Atlas was created to make it easier for teams across Lloyds Banking Group to build and deploy machine learning and AI solutions within a secure and governed environment. As the platform matured, adoption accelerated rapidly across business areas, bringing more users, more use cases and greater operational complexity.</p><p>What began as a technology scaling challenge quickly became an operating model challenge. Every new workbench, project, release and support request increased demand on the teams responsible for running the platform. Processes that were effective at launch became increasingly difficult to sustain at enterprise scale.</p><p>By 2026, Atlas had grown into a critical capability supporting over 1,400 users and more than 140 active use cases and handled more than 700 service requests, over 300 changes and 178 user-raised incidents. At that volume, support could no longer sit at the edge of the product. It needed clear ownership, repeatable processes and automation that could grow with demand.</p><h3>What we changed</h3><p>We designed the support model around one goal: make Atlas simple for users, despite the complexity beneath it. The changes reduced friction across the user journey, accelerated issue resolution and improved service reliability. AI builders can focus on developing, testing and deploying solutions rather than navigating operational processes.</p><h3>What guided us</h3><p><strong>Start with the user:</strong> We tested each improvement against the user journey rather than internal convenience. The test was simple: does this make Atlas easier to use?</p><p><strong>Build for reliability:</strong> Dashboards, alerts and playbooks became core service capabilities. They helped teams identify issues earlier, act faster and recover with confidence.</p><p><strong>Trust by default: </strong>Integrate group security and governance controls into the platform with clear paths to human review and escalation.</p><h3>How it worked</h3><p>At the start of 2026, the Atlas support team reorganised around specialist disciplines. Infrastructure specialists owned cloud stability, capacity and security. API specialists supported the interfaces connecting Atlas services and user journeys. The evergreening team kept platform components current, supported and secure. Cloud Build specialists maintained automated environment creation and updates, while Route to Live specialists helped teams move changes safely into production. A dedicated helpdesk handled common requests and routed complex issues to the right specialists. Shared dashboards, alerts and playbooks connected these disciplines as one service without diluting expertise.</p><p>The model created clearer ownership, faster resolution and stronger operational control. Users had a simpler route through the service because requests reached the right specialist sooner.</p><p>The results were measurable. Project provisioning time fell from three weeks to five minutes. Mean incident resolution time fell from 107 hours to eight hours. Incident service performance rose from 58.5% to 86.3%, above the 80% target. These changes improved routine platform journeys and made support more reliable when users needed help.</p><p>The value extended beyond the support team. Faster provisioning helped users start sooner, fewer avoidable delays kept delivery moving, and faster recovery reduced time lost to incidents. The result was a more dependable foundation, creating the space for teams to spend more time developing and deploying AI solutions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/577/1*5Blrn_fO9QY1mm8u8N2AFg.png" /><figcaption>Diagram showing the Atlas support operating model, with service requests, incidents and support questions flowing through helpdesk and operations teams to specialist platform services and engineering experts.</figcaption></figure><h3>What we learnt</h3><p>Three lessons emerged.</p><p>First, support must be designed as part of the product from the start.</p><p>Second, user experience matters as much as technical performance because users judge the platform through every request, change and incident.</p><p>Third, automation creates value only when it removes real friction from the user journey.</p><h3>What’s next</h3><p>The next step is to use AI within operations. We plan to detect unusual patterns before they become incidents, identify emerging service risks, suggest likely causes and next actions, and automate selected routine support decisions where appropriate.</p><p>Earlier intervention and faster diagnosis will further reduce time spent on familiar issues and give users more relevant support. Specialists will remain accountable for decisions and outcomes. Human oversight and operational controls will remain wherever availability, security or user access could be affected.</p><p>Technology alone does not scale an AI platform. The operating model must make that technology dependable, usable and trusted. Atlas has reduced provisioning times, resolved incidents faster and removed avoidable friction, giving AI builders more time to create value. This service foundation helps more teams move from an idea to a governed, production-ready solution with greater speed and confidence. Atlas scales because the service around it scales too.</p><h3><strong>Meet the Authors</strong></h3><p><a href="https://www.linkedin.com/in/donald-buchanan-cmgr-mcmi-579101a5/">Donald Buchanan CMgr MCMI | LinkedIn</a>, Service Lead Machine Learning Platform at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/arshadahmed/">Arshad Ahmed | LinkedIn</a>, Director of Product, AI Platform at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/lucy-sweet/">Lucy Sweet | LinkedIn</a>, Head of Engineering, Machine Learning as a Service at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/richardabates/">Richard Bates | LinkedIn</a>, Director of AI and Analytics Platform at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/nick-l-7881b53/">Nick L. | LinkedIn</a>, Analytics and AI Engineering Director at Lloyds Banking Group</p><h3><strong>AI at Lloyds Banking Group</strong></h3><p>For more detail on our AI transformation, visit <a href="https://www.lloydsbankinggroup.com/who-we-are/group-overview/artificial-intelligence.html">Artificial Intelligence — Lloyds Banking Group</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=67e2566fdae9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-at-lloyds-banking-group/ai-platforms-dont-scale-through-technology-alone-67e2566fdae9">AI Platforms Don’t Scale Through Technology Alone</a> was originally published in <a href="https://medium.com/ai-at-lloyds-banking-group">AI at Lloyds Banking Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Inside Atum: Our Foundation Model for Banking]]></title>
            <link>https://medium.com/ai-at-lloyds-banking-group/inside-atum-our-foundation-model-for-banking-b10f0abb70bd?source=rss-4c2e7038b290------2</link>
            <guid isPermaLink="false">https://medium.com/p/b10f0abb70bd</guid>
            <category><![CDATA[financial-services]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[foundation-models]]></category>
            <dc:creator><![CDATA[AI at Lloyds Banking Group]]></dc:creator>
            <pubDate>Wed, 16 Sep 2026 08:41:41 GMT</pubDate>
            <atom:updated>2026-09-16T08:41:41.743Z</atom:updated>
            <content:encoded><![CDATA[<p><em>What if AI could learn from 18 billion transactions to better understand how people manage their financial lives?</em></p><h3>The Summary</h3><p>Atum 100M advances Lloyds Banking Group’s foundation model for learning reusable representations of customer financial behaviour. Its design combines transaction-specific data processing, hierarchical transformer architecture, and self-supervised learning at banking scale.</p><ul><li>The model tokenises transaction fields, represents elapsed time, and uses Event and History Encoders to capture patterns from individual events through long-term customer behaviour.</li><li>Training uses masked event prediction and next event prediction across roughly 18 billion transactions, supported by distributed tokenisation, 32 GPUs, and memory-saving techniques.</li><li>The release aims to improve downstream banking applications, reduce reliance on handcrafted features, and support future work toward Atum 1B.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cXVDR-u1pzMVU8oYGIVN6Q.jpeg" /><figcaption>Abstract illustration of a banking AI foundation model, showing financial transaction data flowing through layered neural network architectures into banking insights, analytics, and customer intelligence.</figcaption></figure><h3>The Problem</h3><p>In the first article in this series, <a href="https://medium.com/ai-at-lloyds-banking-group/introducing-atum-a-foundation-model-for-banking-4d5291577d50"><em>Introducing Atum: A Foundation Model for Banking</em></a><em>, </em>we set out why we are building a foundation model for financial services and our vision for applying foundation-model principles <strong>[1]</strong> to banking data. Atum is a shift away from models tailored to individual use cases and towards learning reusable representations of customer financial behaviour at scale.</p><p>Today we are announcing our next release, <strong>Atum 100M</strong>. It builds on the architecture of earlier versions and significantly expands the model’s capacity while keeping the same core objective: learning rich, reusable representations of customer behaviour from transaction history. This article takes you inside the model, <strong>along the three pillars that make it work: </strong>large-scale transactional data, a transformer architecture built for banking, and self-supervised learning.</p><p>This work has not been done in isolation. The idea of learning general representations from raw sequences and reusing them across tasks comes from natural language processing, where models like BERT<strong>[2]</strong> showed that a single pre-trained encoder could be adapted to many problems. A line of research has since carried that idea into transaction data: TabBERT<strong>[3]</strong> applied masked-token pre-training to tabular records, FATA-Trans<strong>[4]</strong> made it field- and time-aware for sequential transactions, and most recently Revolut’s PRAGMA<strong>[5]</strong>, built with NVIDIA, showed the approach working at production scale. Atum sits within this lineage, extending these ideas to the scale of Lloyds Banking Group and to our own governed data. We are also collaborating with NVIDIA teams experienced in accelerated computing for workloads of this scale.</p><h3><strong>Pillar One: Data</strong></h3><p><strong>From Transactions to Sequences</strong></p><p>Atum is designed around a simple observation: financial behaviour is inherently sequential. Just as a language model learns from sequences of words<strong>[6]</strong>, Atum learns from sequences of transactions. Language models can rely on an off-the-shelf tokenizer; transactions have no such standard, so we built a bespoke tokenisation pipeline for banking events.</p><p>Rather than treat a transaction as a single record, the pipeline maps each of its fields to a <em>prefixed token</em>, so the field a value belongs to is carried in the token itself: an amount band becomes amt_6, a country becomes cntry_3, a merchant category becomes mcc_5411, and so on. Continuous values like amounts are discretised into bands before being tokenised, and time is encoded separately with cyclical features that preserve daily and weekly rhythms.</p><p>Some fields have too many distinct values for each to be assigned its own token without making the vocabulary unwieldy. For these fields, we borrow an idea from language tokenisers and apply <em>BPE-style subword merges</em><strong>[9]</strong>, so common values get compact tokens while rare ones are composed from smaller pieces, keeping the vocabulary bounded. The result is a single transaction expressed as a short sequence of prefixed tokens:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/475/1*khLiWb9E0qGFclB9f9om3g.png" /><figcaption>Diagram showing how a banking transaction is converted into AI-readable tokens, transforming transaction details such as merchant, amount, category, and location into structured identifiers for model training.</figcaption></figure><h3><strong>Pillar Two: Architecture</strong></h3><p><strong>A transformer built for transactions</strong></p><p>The architecture adapts the Transformer<strong>[6]</strong> to the banking domain. At a high level the flow is: transaction stream → tokenisation → sequence → Transformer encoders → customer embedding. But unlike a sentence, a customer history can run to thousands of events across years. Processing all of it at once is impractical, so Atum uses a hierarchy of two encoders, an approach shared with PRAGMA<strong>[5] </strong>where local structure and long-range history are modelled at separate levels. Other approaches, such as FATA-Trans[4], also employ multiple encoders, although they partition the problem differently, modelling static and dynamic fields separately.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/490/1*d7EB6EQ2Qk2xOoPVNSuBRQ.png" /><figcaption>Diagram showing a two-stage AI architecture that learns customer financial behaviour by analysing individual transactions, short-term spending patterns, and long-term behavioural trends to create a customer representation.</figcaption></figure><h3><strong>Learning behaviour within a window</strong></h3><p>Histories are divided into windows of up to an optimal number of transactions per customer. Each window is processed by the <strong>Event Encoder</strong>, a Transformer that learns relationships between transactions in that period. Through self-attention, it identifies patterns such as recurring income, regular household spending, subscriptions and merchant relationships. A special <strong>USR</strong> token sits at the start of each window and learns to summarise everything in it, the same trick BERT<strong>[2]</strong> uses with its [CLS] token to pool a whole sequence into one vector, producing one compact representation of that stretch of behaviour.</p><p><strong>Learning behaviour across time</strong></p><p>Customers are characterised by patterns that emerge across months and years, rather than within a single window. The <strong>History Encoder</strong> takes the sequence of window summaries and learns the higher-level patterns: changing spending habits, income consistency, lifecycle transitions, long-term trends. An attention-based pooling step then combines these into a single customer embedding, one vector representing the customer’s overall financial behaviour.</p><p>Together the two encoders build behaviour up in layers, from a single payment to a whole financial life:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ya-Nhr0YR5LEcLk6Mc0iyA.png" /><figcaption>Table showing the three levels of learning in the Atum model: individual transactions, short-term behavioural patterns, and long-term customer financial behaviour.</figcaption></figure><p><em>Looking ahead, we are exploring richer context still: profile encoders and graph-based representations that capture the relationships between customers, accounts and merchants.</em></p><h3><strong>Modifications to the architecture</strong></h3><p><strong>Two design choices for longitudinal transaction data</strong></p><p>On top of the hierarchical backbone, two design choices are worth noting. Both are shaped by the nature of long banking histories rather than borrowed unchanged from language models.</p><p><strong>Representing the time between events</strong></p><p>Calendar features tell the model <em>when</em> a transaction occurred, including the hour, day and month. On their own they do not capture the spacing between transactions. That spacing is itself a behavioural signal: it separates a customer paying in quick bursts from one whose account is largely dormant, and it carries the rhythm of salary cycles and recurring payments. Encoder models for banking events, including PRAGMA<a href="https://ukc-word-edit.officeapps.live.com/we/wordeditorframe.aspx?ui=en-GB&amp;rs=en-US&amp;wopisrc=https%3A%2F%2Flloydsbanking-my.sharepoint.com%2Fpersonal%2Fsiyu_lin_lloydsbanking_com%2F_vti_bin%2Fwopi.ashx%2Ffiles%2Fd4bf1fb6a63b415d8138fadc949be0b4&amp;wdprevioussession=408e3561-03c4-ccc2-94ed-dc2624e7fb18&amp;wdnewandopenct=1788447713384&amp;wdo=4&amp;wdorigin=BrowserReload&amp;wdpreviouscorrelation=2e1de7a3-9825-498e-9156-72d19e52627c&amp;wdnd=1&amp;wdexp=TEAMS-TREATMENT&amp;wdenableroaming=1&amp;mscc=1&amp;wdodb=1&amp;hid=037F38A2-5007-3001-09FB-E94DA68E397C.0&amp;uih=sharepointcom&amp;wdlcid=en-GB&amp;jsapi=1&amp;jsapiver=v2&amp;corrid=c130867c-dfae-4c9e-494d-31617e1f1a39&amp;usid=c130867c-dfae-4c9e-494d-31617e1f1a39&amp;newsession=1&amp;sftc=1&amp;hfto=1788682131269.6&amp;uihit=docaspx&amp;muv=1&amp;ats=PairwiseBroker&amp;cac=1&amp;sams=1&amp;mtf=1&amp;sfp=1&amp;sdp=1&amp;hch=1&amp;hwfh=1&amp;dchat=1&amp;sc=%7B%22pmo%22%3A%22https%3A%2F%2Flloydsbanking-my.sharepoint.com%22%2C%22pmshare%22%3Atrue%7D&amp;ctp=LeastProtected&amp;rct=Normal&amp;afdflight=23&amp;csiro=1&amp;instantedit=1&amp;wopicomplete=1&amp;wdredirectionreason=Unified_SingleFlush#ref5"><strong>[5]</strong></a>, therefore encode elapsed time since the previous event alongside calendar time.</p><p>We represent this elapsed gap as a <em>learned embedding over log-scaled buckets</em> rather than a continuous transform. For each transaction we take the gap since the customer’s previous event and assign it to one of a set of intervals spanning minutes to months, then look up a dedicated embedding for that bucket. Writing the event representation as a sum of its parts:</p><p>The gap embedding <em>hgap</em> joins the existing field, calendar and position signals without displacing any of them; it is one more parallel feature the model can attend to. Discrete buckets let the model learn a distinct representation for, say, “within the hour” versus “over a month,” rather than assuming the effect of elapsed time varies smoothly.</p><p><strong>Keeping overlapping windows consistent</strong></p><p>Because long histories are sliced into overlapping windows, two adjacent windows from the same customer share most of their transactions. They are two partial views of the same history. However, a standard setup treats them as independent examples, meaning that nothing explicitly encourages their customer-level summaries to agree. The embedding can then shift depending on where an arbitrary window boundary fell.</p><p>A consistency objective addresses this<strong>[10]</strong>. For two overlapping windows from the same customer, we take their normalised window-level summaries and train them to agree, while treating windows from other customers in the batch as negatives. This helps preserve a more representative picture of a customer’s transaction history.</p><p>The effect is a customer embedding that is more stable and less sensitive to where a window boundary happens to fall, which matters directly for a representation meant to be reused across many downstream models.</p><h3><strong>Pillar Three: Self-Supervised Learning</strong></h3><p><strong>Learning without labels</strong></p><p>Traditional models are trained to predict a specific outcome and therefore usually require labelled data. Atum learns differently, directly from transaction histories through self-supervised learning, discovering structure without any manual labelling. Two complementary objectives teach it what financial behaviour looks like.</p><p><strong>Masked Event Prediction.</strong> During training we hide a subset of transactions and ask the model to reconstruct them from the surrounding history. This is the masked-language-modelling idea from BERT<strong>[2]</strong>, adapted to transactions: crucially, we mask whole transactions rather than individual fields, which stops the model taking shortcuts within a single record and forces it to learn behavioural context across the sequence.</p><p><strong>Next Event Prediction.</strong> The model also learns to link each transaction’s representation to the one that follows it, so its understanding of a customer’s history is organised by the order in which events happened in without relying on the reconstruction of its contents. This teaches it not only what has happened, but how behaviour is sequenced over time. **</p><p>Together, these two objectives capture both the structure and dynamics of financial activity, producing reusable representations without requiring any manually labelled data during pre-training.</p><h3><strong>Training at banking scale</strong></h3><p>A foundation model needs more than a good architecture; it needs exposure to large, diverse data, and the infrastructure to learn from it in a practical amount of time. Atum 100M was trained on a broad corpus of customer histories (<em>roughly 18 billion transactions)</em> spanning many spending patterns, income behaviours and life events. Moving from 10 million to 100 million parameters increases the model’s capacity to learn richer representations while keeping the same hierarchical design. Getting there meant solving two distributed problems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/568/1*msZG9wBtqYYIf7RyGI7dMg.png" /><figcaption>Diagram showing Atum’s distributed training pipeline, where billions of banking transactions are tokenised in parallel and processed across multiple GPUs to train the foundation model efficiently.</figcaption></figure><p>This is not autoregressive forecasting. The encoder is bidirectional and can observe the full sequence throughout training. The objective instead encourages representations that capture event order and elapsed time, providing a structural learning signal rather than a forecasting one. It is also assigned a lower weight than the MEP objective, so the overall training loss remains dominated by MEP.</p><p><strong>Distributed tokenisation</strong></p><p>Before the model sees anything, billions of transactions must become windowed, tokenised sequences. At this volume that is a distributed data problem. The raw corpus is divided into many shards held in cloud storage, and workers tokenise them in parallel, each turning its slice of transactions into the sequence format the model reads. The output is a corpus of ready-to-train windows, produced once and reused across every training run rather than being rebuilt each time.</p><p><strong>Distributed training</strong></p><p>Training itself follows a chief-and-worker architecture coordinated by <strong>Ray[7]</strong>. A lightweight CPU-based chief performs the coordination and does not use a GPU. The main compute workload sits on a fleet of GPU workers utilising CUDA-X for processing, and <strong>Ray Data</strong> streams the tokenised sequences to them continuously, so the GPUs are kept busy rather than waiting on input, which can otherwise become a bottleneck at this scale. Each worker computes gradients on its share of customers, and the workers synchronise every step, so they stay a single model rather than drifting apart. The GPU workers comprise four g4-standard-384 machines with 32 NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs in total.</p><p><strong>Fitting a bigger model in memory</strong></p><p>The harder constraint at 100 million parameters is memory, not compute. We use <strong>DeepSpeed[8]</strong> to shard the optimiser state across the whole fleet, so no single GPU has to hold the full training footprint, a technique from the ZeRO family of memory optimizations<strong>[8]</strong>. Two more choices keep us comfortably within budget: mixed-precision training, which reduces the memory required by most tensors, and gradient checkpointing, which recomputes some intermediate values on the backward pass instead of storing them, trading a little extra compute for a large memory saving.</p><h3><strong>What this unlocks</strong></h3><p>For the bank, Atum 100M’s greater capacity enables deeper, richer representations, with the potential to improve performance across our use cases. The overlapping-window objective also makes these representations more stable, helping downstream models distinguish genuine behavioural change from arbitrary window boundaries.</p><h3><strong>What we’ll share next</strong></h3><p><strong>Atum 100M</strong> is another milestone on the road to foundation models for financial services. By combining large-scale transaction data, modern transformer architectures and self-supervised learning, it shows how reusable behavioural representations can cut our dependence on handcrafted features and open up a more scalable approach to machine learning in banking.</p><p>In future articles we will show how Atum embeddings are consumed across banking applications, walk through use cases and evaluation, and set out the Responsible AI principles guiding how we build and deploy foundation models in a regulated environment. We will also open the next stage of the roadmap: our work towards Atum 1B.</p><h3><strong>References</strong></h3><ol><li>Bommasani et al. (2021). <em>On the Opportunities and Risks of Foundation Models.</em> arXiv:2108.07258.</li><li>Devlin et al. (2019). <em>BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.</em> NAACL. arXiv:1810.04805.</li><li>Padhi et al. (2021). <em>Tabular Transformers for Modeling Multivariate Time Series (TabBERT).</em> ICASSP. arXiv:2011.01843.</li><li>Zhang et al. (2023). <em>FATA-Trans: Field And Time-Aware Transformer for Sequential Tabular Data.</em> CIKM. arXiv:2310.13818.</li><li>Revolut Research &amp; NVIDIA (2026). <em>PRAGMA: Revolut Foundation Model.</em> arXiv:2604.08649.</li><li>Vaswani et al. (2017). <em>Attention Is All You Need.</em> NeurIPS. arXiv:1706.03762.</li><li>Moritz et al. (2018). <em>Ray: A Distributed Framework for Emerging AI Applications.</em> OSDI. arXiv:1712.05889.</li><li>Rajbhandari et al. (2020). <em>ZeRO: Memory Optimizations Toward Training Trillion Parameter Models (DeepSpeed).</em> SC. arXiv:1910.02054.</li><li>Sennrich et al. (2016). <em>Neural Machine Translation of Rare Words with Subword Units (BPE).</em> ACL. arXiv:1508.07909.</li><li>Babaev et al. (2022). CoLES: Contrastive Learning for Event Sequences with Self-Supervision. KDD. arXiv:2002.08232.</li></ol><p>** This is not autoregressive forecasting. The encoder is bidirectional and can observe the full sequence throughout training. The objective instead encourages representations that capture event order and elapsed time, providing a structural learning signal rather than a forecasting one. It is also assigned a lower weight than the MEP objective, so the overall training loss remains dominated by MEP.</p><h3><strong>Authors</strong></h3><p><a href="https://www.linkedin.com/in/temiafeye/">Temi Afeye | LinkedIn</a>, <em>Senior AI Scientist, Commercial Analytics at Lloyds Banking Group</em></p><p><a href="https://www.linkedin.com/in/siyulincicely/">Siyu Lin | LinkedIn</a>, <em>ML Engineer, Commercial Analytics at Lloyds Banking Group</em></p><p><a href="https://www.linkedin.com/in/usman-choudahary/">Usman Choudahary | LinkedIn</a>, <em>ML Engineer, Commercial Analytics at Lloyds Banking Group</em></p><p><a href="https://www.linkedin.com/in/datanick/">Nick W | LinkedIn</a>, <em>Head of AI &amp; ML Engineering, Commercial Analytics at Lloyds Banking Group</em></p><p><a href="https://www.linkedin.com/in/vivek-shah-b8033374/">Vivek Shah | LinkedIn</a>, <em>Lead Data &amp; AI Scientist, Commercial Analytics at Lloyds Banking Group</em></p><p><a href="https://www.linkedin.com/in/alexandercave/">Alexander Cave | LinkedIn</a>, <em>Commercial Analytics Director at Lloyds Banking Group</em></p><h3><strong>AI at Lloyds Banking Group</strong></h3><p>For more detail on our AI transformation, visit <a href="https://www.lloydsbankinggroup.com/who-we-are/group-overview/artificial-intelligence.html">Artificial Intelligence — Lloyds Banking Group</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b10f0abb70bd" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-at-lloyds-banking-group/inside-atum-our-foundation-model-for-banking-b10f0abb70bd">Inside Atum: Our Foundation Model for Banking</a> was originally published in <a href="https://medium.com/ai-at-lloyds-banking-group">AI at Lloyds Banking Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why we stopped worrying about which AI model is best]]></title>
            <link>https://medium.com/ai-at-lloyds-banking-group/why-we-stopped-worrying-about-which-ai-model-is-best-a694f1356bcf?source=rss-4c2e7038b290------2</link>
            <guid isPermaLink="false">https://medium.com/p/a694f1356bcf</guid>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[ai-platform]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[AI at Lloyds Banking Group]]></dc:creator>
            <pubDate>Wed, 02 Sep 2026 16:01:06 GMT</pubDate>
            <atom:updated>2026-09-02T16:01:06.548Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>In banking, the real challenge isn’t model choice. It’s uncontrolled repetition.</strong></p><p>Generative AI’s rapid adoption in banking revealed that the core challenge is not model selection but managing uncontrolled repetition and operational risks at scale. A central platform, CorteX, was developed to provide safe, consistent, and scalable AI access across teams.</p><ul><li>CorteX enabled Lloyds Banking Group to deliver around £50m in value with 50 live GenAI use cases by 2025 through a gateway approach embedding safety and governance.</li><li>Key lessons include the need to manage model changes, complex cost structures, and frictionless controls to avoid duplication and risk.</li><li>The platform is evolving from access to intelligent orchestration, using evaluation frameworks like Pegasus to optimize model routing by quality, cost, and safety, aiming to make AI adoption boring, safe, and scalable.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hqYbndXkoF-Ab_Z13mu4rQ.png" /><figcaption>A central AI platform connecting teams, data, governance and automation across a secure banking network.</figcaption></figure><h3><strong>Generative AI arrived with the subtlety of a brick through a window.</strong></h3><p>Overnight, every product roadmap had an AI strategy, every vendor had a copilot, and every conference had a panel on responsible innovation. The hype was exhausting but the opportunity was not.</p><p>For banks, the choice was uncomfortable but simple: ignore AI and risk missing opportunities to improve colleague and customer outcomes, or adopt it without the right controls and create a different problem. Neither was acceptable. The job was to move quickly without being reckless; to give teams access without letting access become fragmentation; to turn a fast-moving, probabilistic technology into something usable, governable and scalable.</p><p>That was the mission for CorteX: make the safe path the easy path, then make the easy path the best path.</p><h3><strong>The real problem isn’t model choice</strong></h3><p>Most AI conversations start in the wrong place.</p><p><em>“Which model should I use?”</em></p><p>That matters, but it is not the hard problem. Models change. Prices change. Benchmarks change. The thing that looks best today can be average tomorrow, expensive next week, and unavailable the week after.</p><h3><strong>At enterprise scale, the real problem is uncontrolled repetition.</strong></h3><p>It is every team solving integration, safety, cost, data handling, monitoring, auditability and resilience for itself. Many routes to many models through many patterns with many different controls.</p><p>That creates duplication, inconsistency and avoidable operational risk.</p><p>We knew the uncertainty would continue, so instead of ignoring it we designed for it. If models, costs, risks and usage patterns were all going to change, then the first job was to make sure those changes could be handled once, centrally, instead of rediscovered across the organisation hundreds of times.</p><h3><strong>Building the Gateway: Why glamour is overrated</strong></h3><p>CorteX started as the obvious engineering answer to an emergent technology problem: a gateway.</p><p><strong>A gateway is not glamorous. </strong>That is deliberate.<strong> </strong>A gateway gives you a boundary. It breaks coupling between systems that change independently. It creates an inflection point where standards, controls and conventions can be applied once and consumed many times.</p><p>CorteX exposed multiple models from multiple providers through a familiar OpenAI-compatible API. Teams could build in a standard way. The organisation could manage safety, cost, performance and data handling centrally.</p><p>Responsible AI became embedded into product capability: guardrails, monitoring and usage controls built into the request path rather than stapled on at the end.</p><p><strong>That distinction matters.</strong> If the responsible route is harder than the workaround, people will find the workaround. If the responsible route is fast, familiar and useful, governance becomes enablement rather than drag.</p><p>The result was impact. In 2025, Lloyds Banking Group moved from a small number of experiments to approximately £50m of benefits delivered by 50 live GenAI use cases, all supported by a central gateway and reusable enterprise patterns.</p><h3><strong>Four hard lessons from operating at scale</strong></h3><p>With enterprise-wide access to models solved, the problem shifted. Teams could build, and the real operational challenges arrived.</p><p><strong>1. Models change; outcomes do not</strong></p><p>Model refreshes are routine, and with every version, cost, latency, bias, and safety can change.</p><p>Teams should not need to become benchmark analysts to do their jobs. They need confidence that their use case clears the bar for quality, latency, safety and cost.</p><p>To help teams deliver consistent outcomes in this changing landscape, we developed ‘Ask CorteX’ (our internal AI builder guide). It recommends the right models, guardrails, tools and services to accelerate delivery through practical design recommendations.</p><p><strong>2. Cost is real, and far harder than the pricing pages suggest</strong></p><p>Tokens are easy to price; end-to-end use cases are not. Retries, long-running context, tool calls, and non-deterministic behaviour turn a tidy unit cost into a moving target.</p><p>Finance needs visibility, teams need control, and leaders need a number they can plan against.</p><p>Our CorteX Console provides that visibility and control through a consolidated view of token usage, costs, guardrail calls and evaluation data.</p><p><strong>3. Friction creates risk</strong></p><p>It’s not news that AI can be dangerous. Controls are a necessity, but when poorly designed, they can discourage teams from taking the safe, approved routes and encourage duplication. The safe path must also be the easy path; every additional hour imposed by the safe path creates an incentive to avoid it.</p><p>To keep things simple, we placed our common AI guardrails within our gateway.</p><p><strong>4. Standing still is not an option</strong></p><p>As an industry, we’re still learning the right ways to make and use AI.</p><p>In months, the world moved from Model + Context to RAG to SFT SLM to Agent to Harness to Skills (and back again). It will move again before this post is a week old.</p><p>Churn is a fact of life in AI, so a product that can only handle today’s pattern is tomorrow’s migration project.</p><h3><strong>The Next Wave: From gateway to orchestration</strong></h3><p>CorteX is now moving from access to orchestration.</p><p>The first version answered: <em>how do we give teams safe access to models?</em> The next version answers: <em>how do we help teams adopt AI safely, consistently and efficiently?</em></p><p>The version after that answers the more interesting question:</p><p><strong>How do we match each workload to the right model, with the right guardrails, at the right cost and performance level, without asking every team to solve that puzzle themselves?</strong></p><p>That is where intelligent routing becomes important. Teams should describe what they need: quality, latency, cost, resilience, safety, data handling and evaluation requirements. CorteX should increasingly handle the mechanics: model selection, fallback, routing, capacity, observability and control.</p><p><strong>Pegasus, our evaluation framework,</strong> gives us the evidence layer to make that possible. Instead of routing by opinion, habit or vendor preference, we can route by evaluation: accuracy, cost, latency, capacity and risk.</p><p>In enterprise engineering, better evidence beats louder opinions every time.</p><p>CorteX is also becoming a multi-tenant AI inference platform. That means predictable behaviour, sensible limits, stronger resilience, clearer operating models and better visibility for teams running real workloads.</p><p>One use case should not be able to quietly ruin another. One model outage should be managed centrally without requiring every application team to respond separately. One new provider should not trigger a thousand integration projects.</p><h3><strong>The Goal: Make AI Boring</strong></h3><p>Scaling AI in banking is hard. It does not need to be chaotic.</p><p>The model market will keep moving. Regulation will keep evolving. Costs will keep changing. Users will keep asking for more. None of that is a reason to slow down. It is a reason to build the gateway layer properly and to make the whole thing boring enough to operate at scale.</p><p>CorteX started as a gateway. That was the right answer for access. It is now becoming the intelligent router for enterprise AI adoption: safe by design, useful by default, and adaptable enough to survive the next wave of change.</p><p>For teams across the Group, this means faster access to approved AI capabilities, less time solving repeat engineering problems, and more focus on delivering value safely and consistently.</p><p>The prize is not choosing the best model. The prize is making AI boring enough to operate, safe enough to trust, and flexible enough to keep moving.</p><h3>Meet the Authors</h3><p><a href="https://www.linkedin.com/in/alexchittock/">Alex Chittock | LinkedIn</a>, Head of Engineering, Generative AI at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/angharad-williams-65902331/">Angharad Williams | LinkedIn</a>, GenAI Product Lead at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/jonathan-cannon-53104a45/">Jonathan Cannon | LinkedIn</a>, Product Owner, Generative AI at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/ollie-dell-5577a2a2/">Ollie Dell | LinkedIn</a>, Lead GenAI Software Engineer at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/amit-kudtarkar/">Amit Kudtarkar | LinkedIn</a>, Principal Engineer, GenAI Platform &amp; Cloud Architecture at Lloyds Banking Group</p><h3><strong>AI at Lloyds Banking Group</strong></h3><p>For more detail on our AI transformation, visit <a href="https://www.lloydsbankinggroup.com/who-we-are/group-overview/artificial-intelligence.html">Artificial Intelligence — Lloyds Banking Group</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a694f1356bcf" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-at-lloyds-banking-group/why-we-stopped-worrying-about-which-ai-model-is-best-a694f1356bcf">Why we stopped worrying about which AI model is best</a> was originally published in <a href="https://medium.com/ai-at-lloyds-banking-group">AI at Lloyds Banking Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Introducing Atum: A Foundation Model for Banking]]></title>
            <link>https://medium.com/ai-at-lloyds-banking-group/introducing-atum-a-foundation-model-for-banking-4d5291577d50?source=rss-4c2e7038b290------2</link>
            <guid isPermaLink="false">https://medium.com/p/4d5291577d50</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <category><![CDATA[foundation-models]]></category>
            <category><![CDATA[banking-innovation]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[AI at Lloyds Banking Group]]></dc:creator>
            <pubDate>Wed, 26 Aug 2026 16:34:52 GMT</pubDate>
            <atom:updated>2026-08-26T16:38:27.695Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The first chapter in turning over 18 billion annual transactions into proprietary intelligence that delivers better, faster and more personalised outcomes for our customers.</em></p><h3>The Summary</h3><p>A Transaction Foundation Model (TFM) is transforming billions of banking transactions into a reusable AI asset that improves efficiency and consistency across hundreds of use cases. This model provides a single, governed customer embedding that accelerates delivery and enhances fraud detection and other banking decisions.</p><ul><li>The TFM reduces duplicated effort by eliminating repeated feature engineering and supports a scalable AI platform for the entire organization.</li><li>It leverages a 10 million parameter encoder-only transformer architecture trained on ordered transaction sequences, with plans to scale to over 1 billion parameters.</li><li>Built on trusted, governed data and powered by GPU-accelerated distributed training, the model aligns with Lloyds Banking Group’s 2030 Accelerate Strategy to enable safer, faster, and more consistent AI-driven outcomes.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sBvY8XKV9Mxv54pDrjan7g.png" /><figcaption>Illustration of a banking AI foundation model, showing streams of transaction data flowing into a central neural network that powers fraud detection, customer insights, governance, and scalable AI decision-making across financial services.</figcaption></figure><h3><strong>The problem</strong></h3><p>Across banking, many data science use cases are built from the same underlying transaction data. Whether the goal is identifying customers who may benefit from a product or service, detecting fraud, assessing affordability, or predicting attrition, teams often start by engineering features from the same transaction history. The result is duplication of effort: multiple teams repeatedly transforming the same data into slightly different representations for individual use cases.</p><p>This approach creates several challenges. It slows delivery, increases development costs, makes it harder to apply data consistently across the organisation, and means improvements made for one use case rarely benefit others. Instead of building on a common foundation, teams continuously rebuild similar capabilities from scratch.</p><h3><strong>A different approach: Large Transaction Foundation Models</strong></h3><p>A new paradigm is emerging in AI and machine learning for financial services.</p><p>Traditional machine learning models, engineered over weeks or months to solve one business problem,<strong> are being challenged by a single, powerful, pre-trained model that can serve a whole host of business problems.</strong> Beyond potential performance improvements, this approach helps teams across the bank use data more consistently, responsibly and effectively, ultimately delivering better experiences for customers.</p><p>The idea will feel familiar. ChatGPT carries out many tasks, from text generation to translation to summarising documents, because it learned general structure from sequences of words. If GPT learns from sequences of words, a Transaction Foundation Model learns from sequences of transactions, encoding the semantics of payments.</p><p>That is exactly what we are building. By learning from a customer’s transaction history, a Large Transaction Foundation Model (LTFM) creates a single customer embedding that can be reused across a wide range of use cases. It gives teams a richer, governed representation of our customers, one that evolves with the customer’s transaction behaviour and is built with Responsible AI in mind. For customers, this means the bank can recognise needs earlier and respond more consistently. This could mean timely support, more relevant products or faster protection when something looks wrong.</p><h3><strong>Why one model, not many</strong></h3><p>Most data science use cases start from the same problem: how do we turn raw customer activity into useful signals? Historically, each team has had to solve that problem repeatedly, building features for one model, one outcome and one delivery route. A Transaction Foundation Model changes that pattern, with the potential to deliver significant benefits across the organisation:</p><ul><li><strong>Reduce duplicated effort: </strong>eliminate repeated feature engineering across teams and use cases.</li><li><strong>Accelerate delivery: </strong>enable data scientists to focus on solving business problems rather than rebuilding the same foundations.</li><li><strong>Improve model performance: </strong>provide richer behavioural representations that can enhance downstream models.</li><li><strong>Create consistency: </strong>establish a common view of customer behaviour across products, channels and analytical teams.</li><li><strong>Scale improvements across the bank: </strong>when the foundation model improves, every use case built upon it can benefit.</li><li><strong>Strengthen governance: </strong>ensure customer representations are built from trusted, controlled and standardised data assets.</li><li><strong>Increase reuse of AI capabilities: </strong>shift from developing isolated models to building a shared AI platform that functions as the engine of enterprise services.</li></ul><p>This is the vision behind Atum: transforming billions of transactions into a reusable AI asset that can power a wide range of banking decisions, creating a shared intelligence layer for the entire organisation.</p><h3><strong>What we have built</strong></h3><p>We are building Atum, Lloyds Banking Group’s family of Transaction Foundation Models, as a reusable, pre-trained AI backbone. Atum learns from billions of ordered transaction events across 28 million customers and turns their transaction histories into compact customer embeddings. <strong>We have already built a 10 million parameter version of the model, with the 100 million parameter model now in training and a short-term path towards 1+ billion parameters.</strong> Rather than building a separate feature set for every use case, teams can start from a shared representation of transaction behaviour that has already learned patterns across income, spending, transfers and changes over time.</p><p>We named the model Atum, after the primordial creator god in Egyptian mythology. The name also evokes atoms, the fundamental building blocks of matter, whose interactions give rise to complex and interconnected systems. Likewise, Atum creates a foundational understanding of customer behaviour, capturing how economic events ripple through households, businesses and communities.</p><p>Atum’s architecture is an encoder-only transformer designed for transaction data rather than text. Each transaction is first processed by an event encoder, which combines fields such as amount, merchant, category and channel into a single dense event representation. A history encoder then reads these summary tokens and learns the rhythms and changes that make each customer’s behaviour distinctive.<em> Both encoders are bidirectional transformers that are trained on masked-event modelling.</em> The output is a customer embedding that can be reused across downstream models and improved centrally over time.</p><p>Transaction data sits at the heart of many banking decisions. Whether helping detect fraud, personalise customer experiences, predict attrition or support affordability assessments, success often depends on uncovering complex behavioural patterns hidden within transaction history. Atum learns these patterns directly from transaction sequences, creating customer embeddings that capture signals that would be difficult to engineer manually. These embeddings can then be used across a wide range of downstream use cases, giving models a richer understanding of customer behaviour.</p><h3><strong>Responsible by design</strong></h3><p>A shared model only works if it is built on trusted foundations. Atum is powered by <a href="https://medium.com/ai-at-lloyds-banking-group/the-day-data-insights-stopped-being-hard-727adae84dc8">EDB Intelligence</a>: governed, standardised data products and a single trusted customer view, rather than local extracts stitched together differently by each team. The same privacy, security and access controls that apply to customer data also apply to the embeddings the model produces.</p><h3><strong>Our inspiration</strong></h3><p>We are not starting from a blank page. The wider industry is exploring how foundation-model techniques can be applied to financial transactions, including work such as <a href="https://arxiv.org/abs/2604.08649v1">Revolut’s PRAGMA</a> and research from organisations including <a href="https://arxiv.org/abs/2507.23267">Nubank</a> and <a href="https://arxiv.org/abs/2511.19693">Visa</a>.</p><p>Our challenge is different: extending proven industry architectures so they work at Lloyds Banking Group scale. That means processing 50 million daily transactions across tens of millions of customers, within governed data products, technical standards and consumption routes that fit how teams across the Group build models.</p><p>To achieve this goal, we are working with NVIDIA teams experienced in accelerated computing for workloads of this scale.</p><h3><strong>Training setup</strong></h3><p>Training this model is as much an engineering challenge as a modelling one. We have scaled in stages, validating a 10 million parameter model before moving to 100 million, with a path towards larger models. To make this feasible, we use GPU-accelerated preprocessing with NVIDIA CUDA-X, distributed training across 32 NVIDIA RTX PRO 6000 Blackwell Edition GPUs, <a href="https://www.ray.io/">Ray</a> for orchestration, and <a href="https://www.deepspeed.ai/">DeepSpeed</a> for training optimisation. We also use a range of Google Cloud services, from data storage in BigQuery and Google Cloud Storage to managed execution through Vertex AI custom training jobs.</p><h3><strong>Key wins so far</strong></h3><p>This work is core to delivering several positive shifts for how we build and reuse AI across the Group.</p><ul><li><strong>Training our first comprehensive Transaction Foundation Model:</strong> we are moving from individual feature engineering to a reusable model that learns from ordered transaction history across the customer base and produces embeddings that can support many use cases.</li><li><strong>Working with massive compute at scale:</strong> we have built the training approach around GPU-accelerated processing, distributed training and managed cloud infrastructure, giving us a practical route to train larger models on very large transaction datasets.</li><li><strong>Designing for downstream consumption:</strong> the model is not just a research asset. It is being set up so teams can consume governed embeddings through clear routes, test them against existing models and reuse them without rebuilding the same transaction features each time.</li><li><strong>Creating benefits for teams:</strong> consuming teams get a faster starting point, more consistent customer representations and a stronger route to model uplift. This lets data scientists spend less time on repeated feature work and more time solving business problems safely and effectively.</li></ul><h3><strong>How this fits in our central AI strategy</strong></h3><p>This work supports our AI strategy: using shared, governed capabilities to help teams build more quickly, safely and consistently. Instead of every team rebuilding the same foundations, a central reusable model allows teams to start from rich, pre-trained customer representations. Data scientists can spend less time engineering repeated features and more time solving business problems, testing use cases and delivering value safely through governed routes.</p><p>That is the practical promise of Atum: a shared AI backbone that improves over time, helping us move faster, act more consistently and unlock more value from the data we already govern carefully.</p><p><em>Colleagues can focus on solving business problems rather than repeatedly building features from scratch.</em></p><p>This is early work, but it marks an important step in how we build AI at Lloyds Banking Group: reusable by default, governed from the start, and designed to help teams deliver better outcomes for customers. Atum has the potential to accelerate data science significantly, improve consistency across use cases and unlock more value from our data.</p><p><strong>In subsequent posts we’ll go under the hood: </strong>how we train a model of this scale using GPUs, distributed training and managed cloud infrastructure.</p><h3>Meet the Authors</h3><p><a href="https://www.linkedin.com/in/temiafeye/">Temi Afeye | LinkedIn</a>, <em>Senior AI Scientist, Commercial Analytics</em></p><p><a href="https://www.linkedin.com/in/datanick/">Nick W | LinkedIn</a>, <em>Head of AI &amp; ML Engineering, Commercial Analytics</em></p><p><a href="https://www.linkedin.com/in/usman-choudahary/">Usman Choudahary | LinkedIn</a>, <em>ML Engineer, Commercial Analytics</em></p><p><a href="https://www.linkedin.com/in/vivek-shah-b8033374/">Vivek Shah | LinkedIn</a>, <em>Lead Data &amp; AI Scientist, Commercial Analytics</em></p><p><a href="https://www.linkedin.com/in/siyulincicely/">Siyu Lin | LinkedIn</a>, <em>ML Engineer, Commercial Analytics</em></p><p><a href="https://www.linkedin.com/in/alexandercave/">Alexander Cave | LinkedIn</a>,<em> Commercial Analytics Director</em></p><h3><strong>AI at Lloyds Banking Group</strong></h3><p>For more detail on our AI transformation, visit <a href="https://www.lloydsbankinggroup.com/who-we-are/group-overview/artificial-intelligence.html">Artificial Intelligence — Lloyds Banking Group</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4d5291577d50" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-at-lloyds-banking-group/introducing-atum-a-foundation-model-for-banking-4d5291577d50">Introducing Atum: A Foundation Model for Banking</a> was originally published in <a href="https://medium.com/ai-at-lloyds-banking-group">AI at Lloyds Banking Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI-powered Data Issue classification: scaling data quality investigation through intelligent…]]></title>
            <link>https://medium.com/ai-at-lloyds-banking-group/ai-powered-data-issue-classification-scaling-data-quality-investigation-through-intelligent-4bb082c930d5?source=rss-4c2e7038b290------2</link>
            <guid isPermaLink="false">https://medium.com/p/4bb082c930d5</guid>
            <category><![CDATA[explainable-ai]]></category>
            <category><![CDATA[data-quality]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-management]]></category>
            <dc:creator><![CDATA[AI at Lloyds Banking Group]]></dc:creator>
            <pubDate>Wed, 19 Aug 2026 09:25:48 GMT</pubDate>
            <atom:updated>2026-08-19T09:25:48.561Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>AI-powered Data Issue classification: scaling data quality investigation through intelligent automation.</strong></h3><h4><em>How we used machine learning to bring greater consistency, transparency and speed to data issue triage.</em></h4><h3>The Summary</h3><p>Machine learning was applied to improve consistency and speed in classifying data issues at Lloyds Banking Group by providing explainable and confidence-scored recommendations. This approach supports better triage and review processes without replacing human judgment.</p><ul><li>The model uses TF-IDF vectorization and Support Vector Machine, achieving 88% accuracy and flags potential misclassifications with confidence above 80%.</li><li>Explain-ability is integrated via LIME, helping users understand classification reasons and maintain accountability.</li><li>The solution is embedded into existing workflows and tracked with a new metric for classification quality, enhancing ongoing data quality management.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pDUM5dikeDXvwzcBnv6Ghw.png" /><figcaption>Machine learning system classifying data quality issues with confidence scores, explainable recommendations, and human oversight.</figcaption></figure><h3>The Problem</h3><p>Data issues raised in Collibra are captured as free-text descriptions, , which can make classification inconsistent and time-consuming at scale. At Lloyds Banking Group, we built a machine learning model to support data issue classification by giving colleagues a more consistent starting point for triage, while keeping human judgement and accountability firmly in the loop.</p><p>The model supports two practical moments in the data issue lifecycle:</p><ol><li>Helps identify issues for review where the model strongly disagrees with an existing classification, surfacing only when the model signal is strong.</li><li><strong>Recommends categories for new candidate issues, </strong>issues that have been logged but not yet categorised, at the point of triage, helping speed up decision-making and improve consistency.</li></ol><p>The solution uses a pragmatic machine learning approach, combining text analysis, classification models, and explainability to deliver transparent, confidence-based recommendations. It also informs how we track classification quality through a data management metric:<strong> the percentage of open issues flagged for classification review.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/618/1*PUHtaw_Bo24ClVyUp38ynw.png" /><figcaption>Diagram showing a machine learning model automatically classifying incoming data issues into Data Quality Issues or General Data Issues based on issue characteristics.</figcaption></figure><h3><strong>The challenge: when free text becomes hard to scale</strong></h3><p>Every data issue starts with a simple question: <strong><em>what exactly is the problem?</em></strong></p><p>In practice, answering that question is rarely simple. At Lloyds Banking Group, colleagues raise data issues in Collibra, our metadata repository, using free-text descriptions. These issues may relate to specific data elements already affected in an application, or to broader process and control weaknesses that create risk to the data colleagues produce and consume. Capturing them in free text is useful because it allows colleagues to describe issues in their own words, but it also introduces variability.</p><p>For data management teams, this creates a practical challenge: consistent classification is often needed before an issue can be prioritised and routed efficiently.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/601/1*MN_DGDG0fV_mZPAp60xJpw.png" /><figcaption>Diagram showing multiple business documents and data sources being combined into a free-text description pipeline before processing and analysis.</figcaption></figure><p>Historically, classification often required multiple triage discussions between the colleague raising the issue, data stewards, and subject matter experts. Those conversations were valuable, but as volumes increased, they also became a bottleneck. Time was spent aligning on what the issue <em>was</em> before teams could focus on what needed to be done about it.</p><p>We saw an opportunity to make the process more scalable, not by replacing human expertise, but by giving colleagues a reliable, explainable starting point for classification.</p><h3><strong>What the model does: two use cases, one capability</strong></h3><p>We designed the solution around two practical, repeatable use cases.</p><p><strong>1) Improving the quality of existing classifications</strong></p><p>The first use case focuses on <strong>retrospective quality improvement</strong>. The model reviews already-categorised issues and flags only those where it strongly disagrees with the current classification and has a confidence score above 80%, helping colleagues focus on the cases most likely to need review.</p><p><strong>2) Supporting classification at the point of triage</strong></p><p>The second use case supports <strong>proactive decision-making</strong>. For candidate issues that have not yet been categorised, the model provides recommendations across all entries. This gives users a consistent baseline they can validate or override, helping them move faster from issue logging to issue understanding.</p><p>Together, these two use cases support both ends of the lifecycle: improving historical consistency while accelerating new triage.</p><h3><strong>Our approach: pragmatic machine learning, built for real workflows</strong></h3><p>A key principle behind the solution was simplicity. The pipeline standardises issue descriptions and transforms the text using TF-IDF (Term Frequency-Inverse Document Frequency), helping the model identify patterns in how different types of issues are described.</p><p>Combined with a <strong>Support Vector Machine (SVM), </strong>this approach delivered the strongest balance of performance and practicality, achieving an <strong>F1 score of approximately 0.78</strong> and an <strong>accuracy of around 88% </strong>on unseen data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/615/1*_vs5GgeIHtrlqmu7Euu7VQ.png" /><figcaption>Diagram showing an AI-powered text classification pipeline that processes free-text issue descriptions and categorises them as Data Quality Issues or General Data Issues.</figcaption></figure><p>However, performance alone was not enough: colleagues also needed to understand why a classification had been recommended, when to rely on it and when to challenge it. That requirement shaped the next design choice.</p><h3><strong>Designing for trust: confidence scoring, LIME, and human review</strong></h3><p>Automation works best when it is applied thoughtfully.</p><p>The model does not simply produce a category and ask users to accept it.</p><p>Instead, each recommendation is paired with a confidence score, helping users distinguish between stronger and weaker model signals.</p><p>To make the model transparent, we built explainability into the experience using <strong>LIME (Local Interpretable Model Agnostic Explanations)</strong>. LIME provides a local explanation of the words and phrases within a description that contributed most to a specific prediction. This allows users to understand <em>why</em> a classification was made, rather than treating the output as a black box.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/596/1*uhfD4XK59eo1ocJEHwhrfQ.png" /><figcaption>AI explainability dashboard showing a machine learning model classifying a free-text issue as a General Data Issue and highlighting the factors that influenced the decision.</figcaption></figure><p>This transparency plays a crucial role in adoption. It helps colleagues validate predictions, understand edge cases, and build confidence in the model over time. Most importantly, it reinforces the model’s role as a <strong>decision-support tool</strong>, not a decision-maker: classification accountability stays with people.</p><h3><strong>Embedding into everyday data management (and measuring impact)</strong></h3><p>The success of the solution is not just in the model itself, but in how it is embedded into day-to-day data management practices and how its impact is measured over time.</p><p>Rather than introducing a standalone tool, model outputs were integrated directly into existing workflows, surfaced through dashboards and reporting views that colleagues already use.</p><p>We introduced a new key data management metric: <strong>the percentage of open issues flagged for classification review</strong>, defined as the proportion of issues where the model identifies a strong signal of potential misclassification.</p><p>This metric acts as a proxy for classification quality and consistency, providing an early indicator of drift in how data issues are being interpreted across the Group, helping teams focus review effort more effectively.</p><p>Importantly, it shifts the solution from a point-in-time classification aid to a continuous improvement mechanism embedded within data management controls.</p><h3><strong>What we learned</strong></h3><p>Two lessons stood out.</p><p><strong>First, input quality matters.</strong> The quality of issue descriptions directly influences classification accuracy, for both humans and machines. Inconsistent or incomplete descriptions make issues harder to interpret and increase variance in how categories are applied.</p><p>In response, we’ve introduced guided prompts within the issue logging process. These prompts help colleagues structure their descriptions more effectively, making it clearer what information is required. Over time, this has created a stronger foundation for more consistent classification.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/585/1*0cjPsLR3rU23BU5zoRh3ag.png" /><figcaption>Diagram showing how guided prompts and feedback create a continuous improvement cycle that enhances AI classification accuracy and overall results.</figcaption></figure><p><strong>Second, explainability is not optional.</strong> Users were far more comfortable adopting the model when they could see the reasoning behind a recommendation. Explainability made it easier to validate outputs, challenge uncertainty, and build trust in day-to-day use.</p><p>Together, these changes created a reinforcing loop: <br><strong>better inputs → better classifications → better outcomes</strong></p><h3><strong>What’s next</strong></h3><p>This work provides a blueprint for approaching similar problems in data management: start with a clear use case, make pragmatic modelling choices, and integrate the solution into real workflows, rather than around them.</p><p>Looking ahead, there is potential to build on this foundation through more granular classification and closer integration with clustering and predictive capabilities. Over time, these enhancements could help teams move from reactive triage towards more proactive and intelligent data quality management.</p><h3><strong>Final reflections</strong></h3><p>The most useful AI solutions are often the ones that solve a specific problem clearly and responsibly. In this case, the value comes from helping colleagues to move faster with more consistency, focus review effort where it matters most, and do so with greater clarity and confidence,</p><h3>Meet the Makers</h3><p><a href="https://www.linkedin.com/in/anukansha-gagneja/">Anukansha Gagneja | LinkedIn</a>, AI Product Manager at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/mojisolaolawepo/">Mojisola Olawepo ITIL®. CAITL®. | LinkedIn</a>, Lead AI and Automation at Lloyds Banking Group</p><p><a href="https://www.linkedin.com/in/ozgur-cebir-91a264333/">Ozgur Cebir | LinkedIn</a>, Data Manager at Lloyds Banking Group</p><h3>AI at Lloyds Banking Group</h3><p>For more detail on our AI transformation, visit <a href="https://www.lloydsbankinggroup.com/who-we-are/group-overview/artificial-intelligence.html">Artificial Intelligence — Lloyds Banking Group</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4bb082c930d5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-at-lloyds-banking-group/ai-powered-data-issue-classification-scaling-data-quality-investigation-through-intelligent-4bb082c930d5">AI-powered Data Issue classification: scaling data quality investigation through intelligent…</a> was originally published in <a href="https://medium.com/ai-at-lloyds-banking-group">AI at Lloyds Banking Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>