By Neil Vernon, chief product officer at Gresham.
The financial services industry has spent considerable energy debating whether AI is production-ready for regulated operations. That debate has been asking the wrong question. Readiness is not the constraint; targeting is. The real question is whether firms have correctly identified where AI can address costs that have never properly been measured, and have therefore never properly been managed.
In most financial operations environments, the majority of real effort is consumed before anyone reaches the visible task. Translating legacy logic before a migration can begin. Investigating why a data feed looks anomalous before anyone can act on it. Chasing context across disconnected systems before an exception can be resolved. Rerouting questions to senior colleagues because institutional knowledge is unevenly distributed. None of this registers on a P&L, but collectively it constitutes a significant overhead that firms have absorbed for years as an unavoidable cost of doing business, when in fact it is neither unavoidable nor necessary to absorb.
Earlier technology investment was largely directed at accelerating the visible work. AI is beginning to eliminate the friction that prevents teams from reaching it. That distinction is where the real return sits, and it is already demonstrable across four areas where operational overhead has historically been highest.
How AI Is Changing the Economics of Migration
The barrier to migrating from legacy reconciliation platforms has never been the licensing cost. It has been the professional services effort required to translate proprietary matching logic, remap data schemas, run parallel environments and validate outputs at scale. For many firms, that effort has made the commercial case impossible to close, and migration has been deferred indefinitely, not because the platform is good enough, but because the cost of leaving feels too high.
AI changes that calculus directly. The same capabilities that allow a model to read and interpret unstructured documentation can be applied to legacy rule sets, translating configurations that would previously have required months of specialist engagement. Each phase of the migration effort that has historically consumed the majority of project time becomes compressible. The result is not just a faster migration, but a migration that becomes viable for firms that have long concluded it was not worth attempting. But migration is only the entry point. Once firms are operating on modern infrastructure, the overhead does not disappear. It simply moves.
Building Pipelines That Adapt to Real-World Data
Real-world data does not arrive in the format it was configured to accept. Columns shift, delimiters change, field names drift between versions of the same counterparty feed. These are not exceptional events. They are the routine texture of operating at scale across a fragmented data landscape, and the traditional response to them is a support ticket, a developer investigation and a period of pipeline downtime while the cause is traced.
The cost is not the variant, it is the downstream consequence. Exceptions that accumulate before the root cause is identified, engineering attention diverted from higher-value work. Intelligent ingestion that recognises known variation patterns and adapts without intervention removes a category of operational noise that firms have long treated as structural rather than solvable. Cleaner pipelines reduce the volume of exceptions that reach the operations team. But they do not eliminate them, and what happens when an exception does arrive matters just as much as how many get through.
Automating the Investigation Before the Decision
The decision that ultimately closes a break typically takes seconds; the work required to reach it does not. Gathering context from multiple systems, classifying the break type, identifying the root cause and assembling the information needed to act or escalate can consume the majority of each resolution cycle, and it is almost entirely repeatable.
AI applied at this layer does not replace the human decision. It eliminates the assembly work so that the decision is what the operator actually does. For regulated firms, this distinction carries specific weight. Human judgement remains in the loop; it is simply applied at the point where it genuinely matters, with the audit trail intact and the accountability chain unbroken. This is a structural change in what operations teams are asked to do. It also changes the question operations leaders need to ask: not just how exceptions are resolved, but whether the team resolving them can sustain that capacity as volume grows and people move.
Making Institutional Knowledge Available to Everyone
Knowledge concentration is one of the least-discussed structural risks in financial operations. The practitioners who can interpret an anomalous break pattern, navigate a counterparty dispute or identify the likely source of a feed discrepancy are typically a small subset of any team. When they are unavailable, on leave or managing an elevated volume day, everyone else operates at reduced effectiveness. That reduced effectiveness rarely appears in any reporting. It simply accumulates as slower resolution, more escalations and more rework.
Context-aware guidance embedded directly in existing operational tools addresses this without requiring a training programme, a manual or a structural reorganisation. Guidance that adapts to what an operator is doing and delivers the institutional knowledge of a senior practitioner to everyone on the team from their first day. This is a change in how operational knowledge is distributed, one that makes teams more resilient to the staff movements, volume spikes and knowledge gaps that are a permanent feature of operating in financial services.
Where the Measurable Return Actually Sits
The firms seeing the most defensible return from AI investment are not those automating the visible work; they are those eliminating the overhead that precedes it. In regulated markets, that shift comes with one further requirement. The AI doing that work must be explainable, auditable and traceable, not as a compliance preference, but as a condition of deployment. Firms that apply that standard as their evaluation framework are the ones best placed to realise the return.
Subscribe to our newsletter
