A deceptively simple arithmetic problem is quietly distorting how the global banking industry evaluates one of the most consequential technology shifts of the modern era. When a bank replaces a workforce of 100 people — costing £10 million annually — with artificial intelligence systems that carry a price tag of £12 million, the traditional cost-reduction spreadsheet delivers a clear verdict: the project has failed. The numbers are worse, not better, and the investment case collapses before it reaches the board. The problem, as analyst and author Chris Skinner argues in the latest instalment of his ongoing series on The Finanser, is that this verdict is profoundly, structurally wrong.
The £10 million-versus-£12 million comparison is not an apples-to-apples calculation. It is a comparison between two entirely different categories of productive capacity. A workforce of 100 humans operates within well-understood constraints: working hours, cognitive bandwidth, error rates, sick leave, regulatory training cycles, and the compounding inefficiencies of organizational hierarchy. An AI system operating across the same functional scope faces none of those constraints in the same way. If those AI systems are capable of performing ten times the volume and quality of work that the human team could deliver, then the relevant comparison is not £10 million against £12 million. It is £10 million against the cost of hiring 1,000 people to match the AI's output — a figure that would run into nine figures for any institution of meaningful scale.
This is not a minor methodological quibble. It represents a fundamental category error that is shaping capital allocation decisions across the financial services industry right now. Banks and financial institutions, conditioned by decades of cost-efficiency programs — from offshoring in the 1990s to robotic process automation in the 2010s — have inherited a mental model in which technology's value is denominated almost exclusively in headcount reduction. Every previous wave of operational technology did, broadly speaking, map onto that model. Fewer clerks per transaction processed. Fewer tellers per branch. Fewer compliance officers per regulatory filing. Artificial intelligence, particularly the generative and agentic systems now entering production environments, operates on a different logic entirely.
The distinction matters enormously for how financial institutions structure their investment cases, measure success, and ultimately decide whether to accelerate or retreat from AI deployment. If the prevailing benchmark remains the cost-reduction spreadsheet — if chief financial officers continue to ask only "what headcount does this eliminate?" — then the industry will systematically underinvest in the technology most likely to determine competitive positioning over the next decade. Worse, institutions that do invest but frame their return on investment incorrectly will declare early-stage AI programs failures precisely at the moment those programs are beginning to deliver the most value.
The Multiplier That Traditional Models Cannot See
Skinner's argument centers on what might be called the productivity multiplier: the capacity of a well-deployed AI system not merely to replicate human output at lower cost, but to expand the total envelope of what an institution can do. A fraud detection model that processes millions of transactions per second does not replace a team of fraud analysts on a one-to-one basis — it enables a risk posture that no human team, regardless of size or salary, could physically sustain. A customer-facing language model that handles complex mortgage queries at three in the morning is not a cheaper version of a call-centre agent; it is a service capability that did not previously exist. The value created by these systems is additive, not substitutive, and conventional return on investment frameworks have no register for additive value that does not first pass through a cost-saving line item.
This has significant implications for how banks approach vendor negotiations, technology architecture decisions, and talent strategy simultaneously. If AI investment is evaluated solely on its ability to reduce the salary bill, procurement teams will optimize for the cheapest models with the narrowest deployment scope — exactly the wrong incentive structure for technology whose value compounds with breadth and depth of integration. The institutions most likely to extract durable competitive advantage from AI are those willing to accept short-term cost increases in exchange for capability gains that only become legible in revenue and risk metrics over a multi-year horizon.
Rethinking the Scorecard Before It Is Too Late
The urgency here should not be understated. Major financial institutions globally are at varying stages of committing substantial capital to AI infrastructure, and the governance frameworks they establish now — including the metrics by which AI programs are judged — will be difficult to revise once embedded in annual planning cycles and board reporting templates. If the scorecard is wrong from the outset, the strategic decisions built on top of it will compound that error across years of deployment.
What Skinner's analysis ultimately demands is not that banks abandon financial discipline in evaluating AI — prudent stewardship of capital remains non-negotiable — but that they build a more sophisticated and honest accounting of what AI actually produces. That means measuring throughput, not just headcount. It means capturing risk-adjusted revenue improvements alongside operational cost lines. It means acknowledging, in formal investment cases, that a system performing ten times the work of its human predecessors at 120% of their salary cost is not a failed investment. It is, by any rational measure, one of the most efficient deployments of institutional capital available in the current technology landscape.
The banks that internalize this shift in framing earliest will not merely outperform their peers on efficiency ratios. They will define what banking capability means for the generation of customers and competitors that follows. The spreadsheet was never the enemy of good decision-making — but mistaking it for the whole picture very much is.
Written by the editorial team — independent journalism powered by Codego Press.