Today we’re speaking with Laurent Giraid, a technologist specializing in artificial intelligence and its real-world application within large-scale enterprises. His work offers a unique window into how financial giants are moving beyond experimental AI projects to fundamentally re-architect their operations. We’ll explore the tangible impacts of this shift, from augmenting customer service with generative AI and freeing up wealth managers for more client-focused work, to revolutionizing internal processes like software development and financial crime detection. This is a story about operationalizing AI at a massive scale and the profound productivity gains that follow.
Expanding a digital assistant’s capabilities from four customer journeys to 21 is a significant leap. Could you walk us through the operational steps involved in this scale-up and how you measure the impact on both customer resolution times and the need for human intervention?
That kind of seven-fold expansion isn’t just a matter of flipping a switch; it’s the result of a deliberate, foundational strategy. The first critical step was restructuring the entire data estate to create unified customer views and migrating those workloads to a scalable cloud platform like Amazon Web Services. Once you have clean, accessible data, you can begin training the models. We started with the initial four journeys, treating them as a proving ground to refine the technology and our processes. The scale-up to 21 involved identifying the most common and impactful customer queries, training the generative AI on those specific pathways, and then rigorously testing them. We measure success very directly: by tracking the end-to-end resolution time for a query within the digital assistant and, most importantly, monitoring the deflection rate—the percentage of queries that are fully resolved without ever needing to be escalated to a human agent. Seeing those resolution times drop and deflection rates climb is the clearest indicator of success.
In wealth management, AI-driven summaries reportedly freed up 30% of advisors’ time for direct client engagement. Can you provide an example of how this works in practice and describe the safeguards you have in place to ensure the accuracy and security of these AI-generated client summaries?
Absolutely. Imagine a relationship manager who has just concluded a series of meetings and has a mountain of notes, call transcripts, and correspondence to process. Previously, they would spend hours manually sifting through this information to update a client’s file and prepare for the next interaction. Now, the system can ingest all of that unstructured data and generate a concise, accurate summary highlighting key discussion points, life changes, and financial goals. This doesn’t just save time; it surfaces insights that might have been buried. The 30% time savings is a direct result of automating that administrative burden, allowing advisors to focus on what they do best: giving advice. To ensure security and accuracy, we operate under a strict AI and Data Ethics Code of Conduct. The models are trained on anonymized internal data, and every summary is auditable. Furthermore, our participation in programs like the Financial Conduct Authority’s Live AI Testing ensures we are constantly stress-testing these systems against the highest regulatory and ethical standards.
Automating call summaries and complaint drafting has saved over 70,000 hours in your retail division. Beyond time savings, how has this technology changed the daily responsibilities and skill requirements for your customer service staff, and what was the training process for these new tools?
The 70,000-hour figure is impressive, but the deeper story is about the evolution of the customer service role. Instead of spending their days bogged down in repetitive summary writing and drafting initial complaint responses, our staff are now functioning more like editors and strategists. The AI provides the first draft, a solid summary of the customer interaction. The employee’s role shifts to reviewing, refining, and adding the nuanced, empathetic human touch that a machine can’t replicate. Their skills are now geared more toward complex problem-solving and high-level communication rather than administrative transcription. The training was hands-on; we didn’t just give them a new tool, we ran workshops showing them how to effectively prompt the AI, how to critically evaluate its output, and how to integrate it seamlessly into their workflow to enhance, not just replace, their judgment.
With AI now producing over a third of your company’s code, how do you manage quality control and ensure a consistent coding standard? Please also elaborate on the tenfold productivity increase seen in agentic engineering trials and how you plan to replicate that success.
Having AI generate over a third of our codebase for a team of 12,000 engineers is a testament to its power, but it requires a new paradigm for quality control. The AI acts as a sophisticated pair programmer, drafting code, running initial tests, and even reviewing submissions. However, the human engineer is always the ultimate arbiter of quality. We enforce rigorous, automated code review processes and maintain strict style guides that the AI models are trained on, ensuring consistency. The human engineer’s role becomes more architectural—they set the direction, validate the logic, and integrate the AI-generated components. The tenfold productivity jump in our agentic trials comes from empowering these systems to take on more complex, multi-step tasks. Instead of just writing a function, an agent can be tasked with an entire objective, like “build and test a new transaction monitoring feature,” and it will autonomously write the code, create the tests, run them, and iterate until the objective is met. We plan to replicate this by identifying more high-value, self-contained engineering objectives and systematically deploying these agentic systems to tackle them across the organization.
Providing tools like Microsoft Copilot to all 60,000 employees is a major investment. What was the primary business case for this universal rollout, and what tangible outcomes have you observed from the more than 50% of staff who pursued advanced AI training?
The business case was rooted in a simple but powerful idedemocratizing productivity. We realized that the benefits of AI shouldn’t be confined to just our tech or data science divisions. Every single employee, from marketing to human resources, has tasks that can be streamlined—drafting emails, summarizing long documents, analyzing data in spreadsheets. Providing a tool like Copilot to all 60,000 staff was an investment in unlocking that latent potential across the entire organization. The fact that over half our employees voluntarily pursued advanced AI training beyond the basics is the most telling outcome. It shows a genuine enthusiasm and a proactive desire to integrate these tools. We’re seeing tangible results in the form of faster project turnarounds, more data-driven presentations, and a general sense of empowerment as employees offload tedious work and focus on higher-value, more creative tasks.
Your financial crime units saw a tenfold productivity increase from agentic engineering trials. Can you describe what an “agentic” system does in this context and explain the key factors that led to such a dramatic improvement in performance over previous methods?
In the context of financial crime, an “agentic” system is an AI that can operate with a degree of autonomy to achieve a complex goal. Instead of a human analyst manually pulling data from ten different systems, running queries, cross-referencing results, and then writing a report, they can give the agent a single directive, like “Investigate this unusual transaction network for signs of money laundering.” The agentic system then independently executes the necessary steps: it queries databases, analyzes transaction patterns, identifies linked entities, flags anomalies based on pre-defined rules, and even drafts the initial suspicious activity report. The tenfold productivity leap is a direct result of this parallel processing and automation. The system can perform in minutes what would take an analyst hours or even days, dramatically increasing the volume and speed of investigations and allowing our human experts to focus their time on the most complex and ambiguous cases that require human intuition.
What is your forecast for the evolution of agentic AI in consumer banking over the next five years?
Looking ahead, I believe agentic AI will become the central nervous system of personal finance. We’re moving beyond today’s digital assistants, which primarily answer questions, to true financial agents that can proactively take action on the customer’s behalf. Imagine telling your banking app, “My goal is to save for a down payment in three years; optimize my spending and investments to get there.” The agent would not only provide a plan but would actively monitor your accounts, suggest real-time spending adjustments, automatically move funds to maximize interest, and even negotiate better rates on recurring bills. It will become a personalized, always-on CFO for everyone, transforming banking from a reactive, transactional service into a proactive, deeply personalized advisory relationship that truly empowers customers to achieve their financial goals.
