By Oliver Fleetwood, Head of Data Science & AI at Scila
The consensus among industry leaders at the most recent XLoD Global in London is clear: “Artificial Intelligence (AI) is a tool, not an alibi”. While financial institutions are eager to move from experimentation toward strategic deployment, regulators like the Financial Conduct Authority (FCA) have issued a firm reminder that responsibility for model outcomes cannot be outsourced.
As the post-event report1, “The End of the Beginning, Not the Beginning of the End: The Future of Non-Financial Risk & Controls across the 3 Lines”, highlights:
“If you can’t explain your models, your data, or why alerts collapsed from hundreds to a handful, you have a supervision problem.”
At Scila, our strategy is built on Vertical AI integration with a strong emphasis on Explainable AI. To us, this means moving beyond general-purpose tools to provide specialized intelligence that is purpose-built for the surveillance environment.
Having consulted extensively with customers and the industry, we have added AI specifically where it adds the most value to our users. We avoid ”AI for AI’s sake” by embedding intelligence directly into the surveillance workflow. This is a modular, optional, and human-centric evolution of the surveillance operating model.
1. Breaking the ”black box”: Insights into model behavior
A primary barrier to AI adoption is the ”opaque” nature of deep learning. Regulators and internal audit functions demand to know why a model flagged a specific sequence of trades. Scila addresses this by using established techniques to demystify our deep learning models so that every anomaly comes with an immediate insight into model behavior.
Ultimately, every anomaly is caused by a deviation from the training data—the model has identified a pattern that doesn’t fit the ”normal” behavior it has learned. To make this understandable for a human analyst, Scila Surveillance provides clear context:
- Feature contribution:
Every prediction is accompanied by scores showing which specific market attributes (such as price, volume, or timing) led the model to its conclusion. - Statistical context:
We complement these insights with distribution charts and statistical metrics per feature. By showing how a current event sits in relation to historical norms (using metrics like Z-scores), we turn a complex mathematical score into a relatable business insight. - Risk-free deployment:
For firms hesitant to deploy AI immediately, Scila allows you to run anomaly detection rules as ”near misses” (shadow alerts). This allows you to observe model behavior in parallel with existing rules, or utilize Scila’s backtesting functionality to validate the AI against historical data before it goes live.
2. Expanding coverage and refining precision
Traditional surveillance often struggles with complex, multi-variable patterns that don’t fit into rigid boxes. Scila’s AI framework closes these gaps through a dual-path strategy that enhances both detection and efficiency:
- Pattern-based anomaly detection:
We use machine learning to detect outlier deviations that traditional rules might miss. Crucially, our models take time into account, allowing the system to identify abnormal sequences and evolving trading flows. This ensures that surveillance is not just looking at a single point in time, but at the ”story” of the trade. - Alert classification:
To solve the industry-wide problem of ”alert fatigue,” Scila employs models that continuously learn from your case management data. By analyzing historical user decisions—which alerts were escalated and which were closed as false positives—the system helps analysts prioritize the ”true positives” that demand immediate attention.
3. Generative AI: Productivity with enterprise guardrails
The ”real prize” of Generative AI (GenAI) is removing the administrative burden of investigation. Scila’s GenAI integration targets key ”choke points” in the surveillance workflow:
- Context-aware case support:
When generating issue titles or summaries, Scila provides the model with relevant metadata and alert context. This ensures the AI isn’t ”guessing” but is grounded in the specific facts of the breach. To ensure the output aligns with internal standards, customers can provide their own templates and rate LLM responses, allowing the integration to be customized to work best for their specific needs. - Configurable AI Quality Assurance (QA): We offer an optional QA module for automated issue review. Users can configure the specific instructions given to the GenAI. The system then analyzes created issues against these guidelines, flagging inconsistencies for human review.
- Enterprise-grade security & hosting: Scila recognizes that many firms have already vetted their own secure GenAI solutions. We do not ask you to bring another system to the table; instead, we offer flexibility:
- Secure API Integration: We connect to your own vetted enterprise providers (e.g., your internal Azure OpenAI or Anthropic instances).
- Self-hosted LLMs: For firms requiring total data isolation, we can host open-source models locally within your own infrastructure.
To ensure accuracy, we utilize self-consistency and multi-step verification—prompting the model multiple times to verify logic before it reaches a human. Looking ahead, Scila will expand this with AI chatbots and agents for workflow automation scheduled for release later in 2026.
Conclusion: Human-in-the-loop by design
The future of surveillance is not about replacing humans with machines; it is about empowering humans to oversee a more complex market. At Scila, every AI feature is modular—it can be turned on or off based on your firm’s specific risk appetite.
By prioritizing Vertical AI Integration, we ensure that these advanced tools live exactly where the work happens. This pragmatic approach allows firms to meet the ”growth agenda” of their business while maintaining the rigorous, explainable oversight demanded by the modern regulatory landscape.
References
1. The End of the Beginning, Not the Beginning of the End: The Future of Non-Financial Risk & Controls across the 3 Lines. Insights from XLoD Global – London, 13 JAN 2026 (https://www.1lod.com/1st-line-risk-control/7949570/the-end-of-the-beginning-not-the-beginning-of-the-end-the-future-of-non-financial-risk-controls-across-the-3-lines)
About the author
Oliver Fleetwood is the Head of Data Science & AI at Scila, where he leads the firm’s strategic AI initiatives and roadmap. With an extensive background as a CTO and startup founder, Oliver has a proven track record of building complex AI-driven products and teams from the ground up. He holds a Ph.D. in Computational Biophysics from the Royal Institute of Technology (KTH), and has published multiple scientific papers on Explainable AI.
If you would like to discuss Scila’s AI roadmap or explore how these technologies can be integrated into your surveillance operations, please get in touch via LinkedIn or contact our team here.
