Evidence and accountability ran through this year’s RegTech Summit London. As artificial intelligence takes on more compliance work, firms need to explain decisions, trace reported data to its source and assign responsibility for errors. Across this year’s sessions, speakers returned to fragmented systems, continuous controls and the expertise needed to challenge automated results.
The opening keynote warned that AI could compress the time firms have to identify vulnerabilities, contain attacks and restore services. Simultaneously, dependence on common cloud and technology providers could have systemic effects spreading disruption across several institutions at once. Firms therefore need to test how they would recover when their providers and counterparties face the same incident. As the speaker put it: “frontier AI is changing what severe but plausible actually means.”AI Under Scrutiny
The first panel examined what separates a successful AI pilot from a dependable compliance process. Speakers called for agreed measures of success, clear ownership and decision records that humans can understand months later. One warned: “Agents must earn the right to have less oversight.”
Human review needs substance. Approving recommendations without challenge adds little protection. An audience question raised a longer-term concern: if agents perform junior analysts’ tasks, how will future decision-makers acquire the experience to judge their work?
The reporting keynote warned that regulators can arrive with evidence of errors before firms’ own controls detect them. Supervisors compare submissions across institutions, trading venues and reporting periods, then ask firms to explain discrepancies. Passing a repository’s validation checks cannot answer those questions: firms must show that they reported the right transactions and that each value reflects the underlying trade.“Schema validation, which firms do regularly, is a very, very low bar,” one speaker said. Firms need to reconcile reports to source data, investigate unusual patterns and track problems through to resolution. AI can help find root causes, provided firms understand their data and its transformations.
Data and Architecture
The data consistency panel challenged the pursuit of a single physical repository. Common definitions and consistent treatment across systems matter more than where firms store information. Speakers discussed AI-assisted mapping of legacy data and the detection of errors that pass syntax checks, alongside established rules and controls.
Peer benchmarking has limits. Being better than other firms does not establish that a report is right. Reporting simplification could raise expectations for the quality of the data supervisors retain.
Cloud migration can increase costs before firms retire legacy reporting systems, the architecture panel warned. Teams must keep existing platforms running while developing the skills to operate their replacements. Cloud services can absorb reporting peaks and reduce infrastructure maintenance, but usage-based charges introduce the risk of unexpected bills. Firms must also cost the controls that let operations teams trace data, investigate exceptions and reconstruct past submissions.Vendor exit plans need to preserve the data and audit history required to reconstruct reports. Coding agents create another maintenance responsibility. Operations teams can build prototypes and dashboards, but someone must test and maintain the software. As one panellist cautioned, “just because you can write some code doesn’t mean you should”.
Converging Demands
The pre-lunch keynote connected AI accountability, quantum threats to retained records and shorter settlement cycles. A surveillance archive may need to preserve an agent’s reasoning for years and protect that evidence against changing cryptographic threats. Faster settlement leaves less time for controls to detect problems and for people to intervene.
The speaker urged firms to assess these demands together when setting budgets. Agent inventories should record permissions, named owners and the means to stop an agent. One test captured the retention challenge: “Could you prove in 2035 what it was authorized to do in 2026?”
After lunch, the unstructured-data panel examined how firms turn regulatory updates into work. Finding and summarising a rule is one step. Determining which entities, products and controls it affects requires internal business context. Speakers stressed source provenance, documented interpretation and human responsibility for the resulting decisions.
Holistic Surveillance Remains Elusive
The surveillance panel described uneven progress across communications and trade surveillance. Connecting transactions with communications and market events (holistic surveillance) remains a challenge for firms to get a complete view of suspicious behaviour. Back-testing and audit trails help teams assess AI outputs, while investigators contribute business context and judgement.
Prediction markets and extended trading hours expose gaps in surveillance built around established products and next-day reviews. One panellist asked, “who is going to review alerts at 3:00 AM?” Another challenged the need for continuous staffing, arguing that surveillance remains a detective control. The discussion left firms with a practical decision: which risks require immediate escalation, and which can await investigation?
Be a Part of the Discussion
Continue these discussions at RegTech Summit New York on 19 November 2026, where sessions will examine production AI, prediction-market surveillance and technology risk. View the agenda and book your place HERE.
Next year’s London summit will be on October 1st, so mark your calendars and keep an eye on the event page for further event details.
Subscribe to our newsletter
