Ask the head of operations at most banks or insurers where the money goes, and the answer is rarely the happy path. Payments that clear first time, policies that bind without a query, trades that match on the first pass — these are cheap and largely automated already. The cost sits in everything that falls out: the payment that needs repairing, the KYC file missing a document, the claim that arrives as forty pages of scanned attachments, the trade break that three teams argue about over email.
That is why the conversation about AI in financial services operations has shifted. The question now is whether AI can handle the judgement-heavy, unstructured, exception-driven work that rules-based automation never reached — and whether it can do so inside the control framework a regulated institution has to run.
It can, but not by being sprinkled over existing processes. The firms getting real operational efficiency from AI are the ones that start with where the work actually is, pick techniques that fit the problem, redesign the process around them and measure the result with the same discipline they would apply to any other investment.
Where operational cost actually sits
Before choosing a model, map the work. In most institutions, a small number of process families account for the bulk of manual effort, and they share three traits: high volumes of unstructured inputs, frequent exceptions, and hand-offs between teams that each hold part of the picture.
- KYC refresh and onboarding. Periodic reviews generate requests for documents, re-screening, and analysts reconciling what a client has provided against what the file already holds.
- Loan origination. Commercial and mortgage lending still depends on financial statements, tax returns, valuations and legal documents that arrive in every format imaginable.
- Claims. A regional P&C carrier may receive photos, repair estimates, police reports and handwritten notes for a single motor claim, each needing to be read, classified and checked against policy terms.
- Trade breaks and reconciliations. Mismatches between internal books, custodians and counterparties are investigated largely by people comparing records and chasing confirmations.
- Payment repairs. Incomplete or malformed instructions drop out of straight-through processing and land in a repair queue, where operators fix fields, look up beneficiary details and resubmit.
The common pattern is that the core system works; the cost is in the gaps around it. That is precisely where AI is strongest.
Matching AI techniques to the problem
“AI” covers a wide range of tools, and much disappointment comes from applying the wrong one. A useful discipline is to name the specific task before naming the technology.
Reading and classifying
Intelligent document processing combines layout-aware extraction, classification and validation to turn documents into structured data. It is the workhorse for KYC packs, loan files and claims submissions. Classification and triage models do the same for unstructured inbound traffic — routing emails, service requests and exceptions to the right queue with the right priority, so that skilled staff spend less time sorting and more time resolving.
Predicting exceptions before they happen
Predictive models trained on historical exceptions can flag payments likely to fail validation, trades likely to break or applications likely to stall. The value is not the prediction itself but the earlier intervention it allows: fixing a standing settlement instruction before the trade date, or requesting a missing document at onboarding rather than at funding.
Drafting and summarising
Generative AI is well suited to the writing work that surrounds operations: summarising a long claim file for an adjuster, drafting a client request for outstanding KYC documents, producing a first-pass investigation narrative for a trade break.
Agentic workflows with human approval
AI agents in operations go a step further, chaining several actions — looking up a beneficiary in reference data, proposing a corrected payment field, assembling the evidence and queuing the repair for sign-off. Designed well, an agent does the legwork and stops at a defined approval point. Designed badly, it becomes an opaque actor with system access. The difference is entirely in the controls built around it.
Redesign the process before you automate it
The fastest way to waste an AI budget is to automate a process nobody would design that way today. Many operational workflows are the sediment of past system migrations, mergers and one-off regulatory fixes. Adding a model to the end of a twelve-step hand-off chain makes a slow process slightly faster; removing six of the steps changes the economics.
Adding a model to a broken process does not fix it — it makes the breakage harder to see.
Consider a mid-sized retail bank whose payment repair queue is staffed across two locations. Before introducing intelligent automation, the useful questions are structural. Why do these payments fail in the first place? How many failures trace back to a handful of corporate clients or a single channel? Could validation at the point of capture prevent the error entirely? Often the best outcome is a combination: upstream fixes that remove a share of exceptions, and AI that handles the residual variety far better than static rules.
Controls, model risk and resilience
Operations is where AI touches customers’ money and regulatory obligations directly, so the control design has to be part of the build, not an audit finding afterwards.
Human-in-the-loop by design
Define, per decision type, what the system may do alone, what it may propose for approval, and what it must never touch. Confidence thresholds should route low-certainty cases to people, and approval steps should present the evidence the model used, not just its answer. Reviewers need to be able to disagree easily — and those disagreements should feed back into monitoring.
Model risk management
In the US, SR 11-7, the supervisory guidance on model risk management issued by the Federal Reserve and the OCC in 2011, remains the reference point for how banks inventory, validate and monitor models. Its principles — sound development, independent validation, ongoing performance monitoring, clear governance — apply to document extraction and triage models as much as to credit scorecards. Generative and agentic systems add new questions: how to validate outputs that are not a single number, how to test for drift when the underlying foundation model is updated by a vendor, and how to document prompts and tool permissions as part of the model. Treat these as model risk problems, not a parallel governance track.
Operational resilience
Once a model sits inside a critical process, its failure becomes an operational incident. Firms need fallbacks: manual procedures that staff still know how to run, the ability to switch a model off without halting the queue, and clarity on third-party dependencies. In the EU, the Digital Operational Resilience Act has applied since January 2025 and brings ICT third-party risk — including AI providers — firmly into scope.
Measuring value and escaping pilot purgatory
Many institutions have a portfolio of AI proofs of concept that impressed in a demo and never reached production. The usual cause is not the technology; it is that success was never defined in operational terms, and nobody owned the path from pilot to scaled service.
Agree the baseline before building. The metrics that matter are the ones operations leaders already manage:
- Straight-through processing rate — the share of cases completed with no manual touch.
- Cycle time — elapsed time from receipt to completion, including waiting time between hand-offs.
- Cost per case — fully loaded, including review effort and exception handling.
- Error and rework rate — measured downstream, not just at the model’s output.
Two disciplines prevent pilot purgatory. First, scope each pilot as the first release of a production service, with a named business owner, an integration path and a control design from day one. Second, set a decision point: after a defined period against the baseline, the use case scales, is reworked or is stopped.
People, change and sequencing
Operations teams know where the bodies are buried. They know which counterparties always send malformed files and which exceptions are really policy questions. Involving them in redesign is not a courtesy; it is how you get a process that works. Roles will shift — from keying and sorting towards investigation, quality review and handling the genuinely complex cases — and that shift needs training, revised performance measures and honest communication about what is changing.
On sequencing, a pragmatic order tends to work:
- Diagnose. Quantify where manual effort and exceptions concentrate, using your own operational data.
- Fix upstream. Remove avoidable exceptions through validation, reference data and process simplification.
- Read and route. Deploy document understanding and triage, which deliver value quickly and build data foundations.
- Predict and draft. Add exception prediction and generative assistance for investigators and case handlers.
- Orchestrate. Introduce agentic workflows with explicit approval points once controls, monitoring and confidence are established.
Each stage funds and de-risks the next, and each leaves the institution with reusable capability — cleaner data, a model inventory that covers AI, and teams that know how to work alongside it.
Key takeaways
AI can optimise financial services operations meaningfully, but the gains come from targeting exceptions and unstructured work, redesigning processes first and building controls in from the start.
- Operational cost concentrates in exceptions and document-heavy work — KYC refresh, origination, claims, trade breaks and payment repairs.
- Match the technique to the task: document processing, triage, prediction, generative drafting and agentic workflows each solve different problems.
- Redesign and simplify processes before automating them; remove avoidable exceptions upstream.
- Treat AI as models under existing model risk frameworks such as SR 11-7, with human approval points and tested fallbacks.
- Measure against an agreed baseline — STP rate, cycle time, cost per case, error rate — and set clear scale-or-stop decisions.
- Sequence the programme from diagnosis to orchestration, and bring operations staff into the redesign.
