PARTNER CONTENT
By Andrey Darenberg, founder, RateYourCyber.com
Generative AI has moved through financial services faster than most internal governance frameworks were designed to absorb. What began as experimentation inside research and technology teams is now embedded across investment analysis, due diligence, compliance, onboarding, client communications and internal reporting. In a number of firms, it has become difficult to remove without disrupting day-to-day operations.
Of course, there is nothing wrong with this trend – productivity gains are real, analysts complete research faster, information moves more freely through organisation. In some cases, decision cycles have shortened in ways that were not previously possible.
The issue is not adoption. It is interpretation: most firms still treat AI as a software procurement problem. A new tool, a new vendor, a new layer of security, a new policy framework. Regulators, on the other hand, increasingly appear to treat it differently – as part of the operational fabric of the firm itself.
That gap in interpretation is becoming more important than the technology itself.
Financial regulation was built around friction. Approval chains, review layers, escalation paths and documentation requirements all assumed that meaningful decisions moved at human speed, through human bottlenecks, with human accountability clearly traceable at each stage. AI reduces much of that friction, and in doing so, it changes the shape of operational risk.
For decades, financial institutions treated technology risk, conduct risk, governance risk and operational risk as separate domains. AI does not respect those boundaries.
A single output can move across them almost instantly: a research summary generated with the assistance of a model may later appear in investor materials. An internal note drafted with AI may influence a credit decision. A junior employee may paste confidential material into a public system without malicious intent, only for questions about oversight and disclosure to emerge later. A model-assisted recommendation may become embedded in a process that no longer has a clear human point of validation.
Individually, none of these issues is novel. In combination, they begin to resemble something different.
Over the past year we spent time testing AI security, monitoring and governance tooling used across financial services. Some tools perform well within narrow constraints. Prompt protection systems can reduce obvious data leakage risks and flag attempts to manipulate models. Monitoring platforms can provide visibility into usage patterns. Governance tools can help structure inventories, policies and risk registers. However, the broader market reality is more fragmented than it appears from vendor materials.
Firms often end up combining multiple products across different layers of the stack, each producing its own dashboards, alerts and logs (we have once seen more than 50 tools used by a regulated firm with 25 employees). What is less common is a unified governance structure that connects those signals into a coherent control framework.
This becomes visible most clearly after an incident, and these are creeping across regulated firms around the world.
Technical systems can often show what happened. They can log prompts, flag anomalies, record usage and trace interactions with models. What they are less able to do is answer the questions that tend to matter most in supervisory or board-level discussions.
- Who approved the workflow in the first place.
- What risk assessment justified its use.
- Whether the firm could explain why the system was appropriate for that context.
- Who was accountable for reviewing outputs before they influenced decisions.
- And whether that accountability is consistent across similar use cases.
These questions are not new to financial services. What is new is the speed at which gaps in those answers can now emerge, and the scale at which they can propagate once they do.
In practice, many governance frameworks were designed for environments where operational processes evolved slowly. Policies were reviewed annually (but mostly designed out of templates and never really adhered to). Controls were validated periodically, if at all. Human review acted as a natural brake on system behaviour and regulated firms very rarely want anyone from outside to review their systems, services, processes and procedures.
AI changes that rhythm: it introduces systems that are adaptive, fast-moving and capable of distributing outputs across multiple workflows before traditional oversight mechanisms are triggered. Model behaviour itself is not static. Vendor terms evolve. Implementation patterns shift inside firms without formal approval. Informal usage often precedes formal governance. This type of operational efficiency is very hard to spot, even in an audit.
The result is a growing separation between how AI is actually used inside firms and how it is described in policy documentation or even understood by the management. That separation is where risk begins to accumulate.
It also explains why current conversations about AI security often feel incomplete inside regulated environments. Much of the focus remains on model-level issues such as hallucinations, prompt injection or vendor selection. These are relevant, but they sit below a more structural concern.
The concern is whether governance systems can still produce reliable accountability in environments where operational behaviour is no longer tightly coupled to formal process.
This is not purely a technology issue. It is increasingly an institutional one.
Financial institutions are introducing machine-speed systems into governance structures still largely operating at human speed. That mismatch is not immediately visible in normal conditions. It becomes visible when something goes wrong and the organisation is required to reconstruct not only what happened, but why it was considered acceptable in the first place.
This is also where external scrutiny is changing. Investors are beginning to ask more direct questions about AI governance during due diligence. Insurers are increasingly attentive to operational exposure linked to AI-assisted workflows. Regulators are applying existing frameworks around operational resilience, accountability, outsourcing and model risk management directly to AI systems, even in the absence of dedicated AI regulation.
The implication is not that firms should slow adoption. That is neither realistic nor, in many cases, desirable. It is that AI governance is becoming a constraint on how safely and credibly adoption can scale.
Firms with stronger governance structures are likely to be able to deploy AI more aggressively because they can evidence control. Firms without them may find that certain uses become harder to justify, not because the technology is unavailable, but because the organisation cannot adequately explain or defend its use under scrutiny.
That shift is already beginning to separate institutions in subtle ways.
The firms navigating this most effectively tend to treat AI as both an operational and governance issue at the same time. Operational controls matter, including approved environments, vendor oversight, access management, monitoring and data classification. Governance structures, and especially, GRC software tools matter equally, including ownership models, escalation procedures, auditability, board reporting and clear accountability around how AI influences decisions.
Neither layer is sufficient on its own.
What emerges from this is a more uncomfortable conclusion than many firms currently acknowledge. The challenge is not simply that AI introduces new risks. It is that it exposes inconsistencies in governance systems that already existed but were less visible in slower-moving operational environments.
Financial services already understand how to manage operational risk. The discipline is not new.
What is changing is the environment in which that discipline is being tested.
Increasingly, it is not the models that determine outcomes. It is whether the institution can still govern them once they are embedded deeply enough into how decisions are made.
Andrey Darenberg, founder, RateYourCyber.com – Andrey has spent the last 12 years in cybersecurity, with 10 years pervious experience in governance consulting. His background is in corporate strategy, investments and venture capital, and finance by training. Andrey has a PhD in Finance, MBA (London Business School), IT Systems Analysis and Design (Oxford), C-DORA-P, ISO 27001 Lead Auditor, C-DPO, CE auditor, IASME auditor.
