The State of AI Economics in Financial Services

Part one, what public research actually establishes

There is now a substantial body of public material on AI in financial services from supervisors and international bodies. Read together, it supports three conclusions with reasonable confidence: adoption is widespread and accelerating; governance frameworks are being stretched rather than rebuilt; and third-party dependency is the risk supervisors return to most often.

It supports almost nothing about unit economics. Surveys ask what firms use AI for and how they govern it. They do not publish cost per resolved case, cost per research note, or measured autonomous resolution rates, because firms largely do not measure them in a comparable way, and would not disclose them if they did.

Our reading of the gap: the industry has adoption evidence and governance guidance, and lacks a shared economic vocabulary. Firms cannot compare AI programmes to each other, or to the operations they replace, because there is no agreed unit. That is the gap this report exists to narrow, edition by edition.

Part two, four modelled workloads

Everything from here is a modelled scenario. The volumes, rates and cost lines are illustrative assumptions chosen to be plausible for a mid-to-large UK institution; they are not measurements, and they are not averages of anything. Every figure is computed by the same deterministic engine, and every slider is yours to move.

Workload 1, Retail client service

The most-funded workload, and the one where the economics are best understood. High volume, short interactions, a clear outcome definition, and an existing cost-to-serve baseline to compare against.

Workload 2, Onboarding and KYC review

Lower volume, longer handling, higher consequence. Automation here is document-heavy and control-heavy: the review step is mandatory for adverse findings, which caps the achievable autonomous rate by design rather than by capability.

Workload 3, Investment research support

Modest volume, very large context per output, and a verification step that cannot be removed. This is the workload where inference cost is most visible as a share, and still not the largest line.

Workload 4, Credit paper drafting

Low volume, high value per output, and a review step performed by expensive people. Here the case rarely rests on cost reduction at all; it rests on cycle time and consistency, with cost neutrality as the test.

What the four scenarios have in common

1. Model inference is never the largest line. Human escalation or verification is, in every one of the four.
2. Unit cost is a function of volume, because four of the seven cost lines are largely fixed.
3. The single most powerful lever is the rate at which work completes without a human, and it is a design and data-quality outcome, not a model-selection outcome.
4. Controls reduce that rate deliberately. Any case that looks excellent without them is describing an unrunnable system.
5. The benefit is credible only where a cost actually leaves a budget, or a loss is actually avoided.

What we would like to publish in the next edition

• Measured autonomous resolution rates by workload type, anonymised and ranged.
• Measured tokens per interaction in production agent loops, which would replace the weakest assumption in every model above.
• Observed escalation handling times compared with pre-AI handling times for the same queues.
• Realised versus modelled cost, twelve months after go-live.

Until then, treat every number in part two as what it is: a transparent model, built from stated assumptions, designed to show you which assumptions your own decision depends on.