<?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 Caroline Collins on Medium]]></title>
        <description><![CDATA[Stories by Caroline Collins on Medium]]></description>
        <link>https://medium.com/@caroline.a.collins?source=rss-20e7fec29931------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*fF6fFR6rtcpjZhUND54p2Q.jpeg</url>
            <title>Stories by Caroline Collins on Medium</title>
            <link>https://medium.com/@caroline.a.collins?source=rss-20e7fec29931------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 22:07:59 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@caroline.a.collins/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[From Interactions to Understanding: Why Responsible Memory Matters for the Next Generation of AI in…]]></title>
            <link>https://medium.com/the-ai-perspective-lloyds-banking-group/from-interactions-to-understanding-why-responsible-memory-matters-for-the-next-generation-of-ai-in-08e997102cf1?source=rss-20e7fec29931------2</link>
            <guid isPermaLink="false">https://medium.com/p/08e997102cf1</guid>
            <category><![CDATA[ai-in-banking]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[conversational-ai]]></category>
            <category><![CDATA[customer-experience]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <dc:creator><![CDATA[Caroline Collins]]></dc:creator>
            <pubDate>Fri, 09 Oct 2026 15:40:02 GMT</pubDate>
            <atom:updated>2026-10-09T15:40:02.037Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>From Interactions to Understanding: Why Responsible Memory Matters for the Next Generation of AI in Banking.</strong></h3><p><em>The future of banking AI isn’t just intelligence. It’s continuity.</em></p><p><strong>For decades, we have taught machines to recognise language, infer intent and generate increasingly fluent responses. But fluency is no longer the hardest problem. An important test for the next generation of AI in banking is whether it can preserve the right context across time, channels and hand-offs, without remembering more than it should.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*79T3SPt4NHBxMs05kYDw7A.png" /><figcaption>Illustration of a customer interacting with an AI assistant that uses conversation history and contextual information to provide personalised, continuous support across banking touchpoints.</figcaption></figure><h3><strong>As a linguist, I have always sat at the intersection of science and art</strong></h3><p>There is the science of language: how speech is recognised, how meaning is constructed, how intent is inferred and how context changes what words mean. Then there is the art of conversation: understanding what somebody is really trying to achieve and responding in a way that is useful, natural and human.</p><p>I’ve spent most of my career exploring that intersection.</p><p>I began in research and development, working on natural-language technologies long before AI became a boardroom priority. Since then, I have worked across successive generations of conversational technology, from DTMF, speech recognition and voice biometrics to chatbots, deterministic automation, retrieval-augmented generation and generative AI.</p><p>Each generation improved what machines could recognise or produce; none removed the need to understand what a person is trying to achieve.</p><h3><strong>Conversation is choreography. Right now, today’s conversational AI is strictly out of step</strong></h3><p>The psychologist Herbert Clark describes conversation as a waltz. Not an exchange of words, but a jointly accomplished sequence with each turn shaped by the last, each move anticipating the next. Both partners hold the structure in memory. They read cues. They adjust. They remember where the dance has been.</p><p>It’s a precise description of something we all do without thinking. It is also a precise description of what much of today’s banking AI still struggles to do consistently: preserve the relevant history of a customer’s journey across conversations, channels and hand-offs.</p><p>The customer remembers the moves. They remember what happened last week, what they were promised, what was resolved and what wasn’t. To them, every interaction is part of one continuous relationship.</p><p>But conversational AI typically begins each conversation as though it’s the first. That gap between interaction and understanding is the defining design problem for the next generation of banking AI. Closing it will require more than larger models or longer context windows. It will require a governed memory layer that can decide what should be carried forward, what should be forgotten and what requires human judgement.</p><p>The problem isn’t that organisations don’t have enough data. The problem is that there is more to do to turn more of that data into understanding.</p><h3><strong>We must show up consistently, wherever the customer moves</strong></h3><p>Customers don’t organise their lives around our products, channels or organisational structures. They are trying to buy a home, recover from fraud, manage their finances, save for the future or resolve a problem. The route they choose at any given moment is simply the one that feels most appropriate, convenient or reassuring at that point in their journey.</p><p>Yet customers arrive at every interaction carrying the context of everything that has come before. They remember previous conversations, promises made, frustrations experienced and issues that remain unresolved. To them, it is one relationship and one continuous story.</p><p>Organisations typically arrive with something different. We have systems of record. We know which products a customer holds, their transactions, balances and account history.</p><p>Those things matter. But they tell us what a customer <strong>has</strong>. They tell us far less about what a customer has <strong>experienced</strong>.</p><p>That is not a data problem. It is a memory problem.</p><h3><strong>We must listen for the rhythm</strong></h3><p>Every customer conversation contains value. It tells us something about what that customer is trying to achieve, what matters to them, what happened previously and what may need to happen next. More importantly, it contains signals.</p><p>· Signals that a customer is becoming frustrated before they formally complain.</p><p>· Signals that a customer has gained confidence after a major financial decision.</p><p>· Signals that an issue remains unresolved, even though the case was marked as closed.</p><p>· Signals that an earlier concern has been resolved and no longer needs attention.</p><p>· Signals that circumstances are changing.</p><p>· Signals that trust is growing, or being eroded.</p><p>· Signals that a customer prefers a particular way of engaging, communicating or making decisions.</p><p>Unlike the facts captured in systems of record, these signals rarely exist as structured fields. They emerge through language, behaviour and interactions over time. They tell us not just what a customer owns, but what a customer is experiencing.</p><p>Capturing these signals is only half the challenge. The other half is making sure that the knowledge travels — carries forward for action, as well as extending back for learning.</p><h3><strong>The next move matters</strong></h3><p>At Lloyds Banking Group, we are addressing this challenge through Conversation Memory: how customer interaction data might be transformed into trusted, relevant context that supports continuity while remaining selective, explainable and governed.</p><p>There are two potential dimensions.</p><p>The first is operational: helping a human or AI agent understand what has already happened for an individual customer and enabling them to continue the conversation from there.</p><p>The second is analytical: connecting interactions across customers and channels to help identify recurring points of friction, emerging needs or stages of a journey where customers frequently need more help.</p><p>These distinctions matter. Remembering the last conversation is useful. But the greater opportunity is combining immediate context with learning scaled across thousands of interactions. That means helping the next agent avoid asking a customer to start again, while also helping the organisation understand why customers need to contact us in the first place.</p><p>We are now working through what it takes to make that real: bringing together interaction data from different channels, turning individual conversations into concise memories, exploring how understanding can be consolidated over time, and making relevant context quickly retrievable by the agents that need it. Just as importantly, we are designing the governance alongside the capability, rather than treating it as something to add later.</p><h3><strong>Memory is a service, not an archive</strong></h3><p>The word ‘memory’ can be misleading because it suggests a place where information is stored.</p><p>The more useful way to think about it is as an active service: a capability that continuously extracts understanding from fresh interaction data, keeps that understanding current and serves the relevant part to the person or agent who needs it.</p><p>Sometimes that understanding will be:</p><p>· Episodic: what happened in the last interaction, what the customer wanted and whether it was resolved.</p><p>· Semantic: a pattern that becomes visible across several interactions, rather than a fact stated in one conversation.</p><p>· Procedural: a preference or adjustment that helps the next conversation work better.</p><p>But no colleague or AI agent needs everything. They need a concise, trusted view of what is relevant <strong>now</strong>, with a clear route back to the source where appropriate.</p><p>Freshness matters just as much as relevance. Old context can be unhelpful or misleading. Information must be updated as the customer’s situation changes, and it must arrive quickly enough to support the live interaction.</p><h3><strong>The goal is judgement, not maximum retention</strong></h3><p>The answer cannot be ‘remember everything’.</p><p>In banking, more information is not automatically safer or more useful. We need to know what is relevant, what is permitted, what is still current and what should no longer be used.</p><p>Was something explicitly stated by the customer, or inferred by a model? How confident are we? Who should have access? Could using it cause harm? Should a person review it before action is taken? When should it decay or be deleted?</p><p>These are not compliance questions to bolt on later. They are part of the product design.</p><p>This principle is not only good product design; it is consistent with the direction of UK regulation.</p><p>Responsible memory should therefore be selective, explainable, current and governed. Its purpose is not to know more about a customer. Its purpose is to help colleagues (human and AI) provide better support while keeping human accountability and customer interests central.</p><h3><strong>This is where I believe the next chapter of AI becomes interesting</strong></h3><p>The challenge is not simply building systems that can speak more naturally. A system can be fluent and still fail a customer. It can generate an excellent answer while missing the fact that this is the third contact about the same unresolved problem.</p><p>The opportunity is to transform interaction data into understanding. To continuously extract meaning from conversations, refine what we know about a customer’s goals, experiences, preferences and changing circumstances, and make that understanding available whenever and wherever it is needed.</p><h3><strong>Almost full circle</strong></h3><p>When I look back across DTMF, speech recognition, natural language, voice biometrics, chatbots, RAG and generative AI, it can look like a story of successive technologies replacing one another.</p><p>I see it differently.</p><p>It is a long-running attempt to close the gap between words and meaning, and between meaning and useful action.</p><p>We are closer than ever to closing the gap between words, meaning and useful action at scale. But models alone will not do it. Progress depends on connecting language intelligence, data, retrieval, governance and live customer journeys so that understanding is timely, traceable and useful.</p><p>The technology has changed. The problem has not.</p><p>We still need to understand what people mean and help them move forward.</p><p>The difference now is that we may finally have the ability to transform conversations into understanding, and make that understanding available wherever it is needed.</p><h3><strong>To finally become the conversational partner who remembers the dance</strong></h3><p>I<em>f you are building conversational agents, where should memory improve continuity, and where should deliberate forgetting be part of the design.</em></p><p><em>What’s working? What isn’t? Where do you think the next breakthrough will come from?</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><h3><strong>About the author</strong></h3><p><a href="https://www.linkedin.com/in/caroline-collins-48b88a/">Caroline Collins</a> is a Lead Product Owner at <a href="https://www.lloydsbankinggroup.com/who-we-are/group-overview/artificial-intelligence.html">Lloyds Banking Group</a>. A linguist by background, she has worked across the evolution of conversational technology, from natural language R&amp;D and speech systems to deterministic automation, retrieval-augmented generation and generative AI. Her current focus is how trusted, timely understanding from customer interactions can support more continuous experiences across human and AI channels.</p><h3><strong>Sources and further reading</strong></h3><ul><li>Clark, “Using Language” (1996)</li><li>Deloitte, “How banks can turn AI-assisted customer service into a business advantage” (2026)</li><li>McKinsey, “The AI-powered bank: Rewiring for excellence in customer care” (2026)</li><li>Maharana et al., “Evaluating Very Long-Term Conversational Memory of LLM Agents” (2024)</li><li>Information Commissioner’s Office, “Data minimisation”</li><li>Financial Conduct Authority, “Consumer support outcome: good practices and areas for improvement” (updated 2026)</li><li>Financial Conduct Authority Handbook, PRIN 2A.6: Consumer support</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=08e997102cf1" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-ai-perspective-lloyds-banking-group/from-interactions-to-understanding-why-responsible-memory-matters-for-the-next-generation-of-ai-in-08e997102cf1">From Interactions to Understanding: Why Responsible Memory Matters for the Next Generation of AI in…</a> was originally published in <a href="https://medium.com/the-ai-perspective-lloyds-banking-group">The AI Perspective | 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>