AI Consumption Pricing in Banking: The Cost Trap of an Uncapped Workforce

For years, automation in banking has followed a simple promise:

• Fewer people,
• lower cost.

That promise was credible when technology was a fixed cost tool. A bank bought licences, staff used the software and the additional cost of serving another employee was limited. But this logic begins to break when a system does not merely support work, but performs it.

A bank can reduce a hundred operational roles and still increase its operating cost.

Not because automation failed. Because it worked.

As automated workflows take on more cases, they consume model capacity, retrieve more data, process longer documents, invoke more tools, produce evidence and create escalating demands for quality control. The bank is no longer buying software for its workforce. It is paying for a digital workforce whose cost rises with the work it performs.

The question for banks is therefore no longer simply how many roles can be removed. It is whether they understand the economics, dependencies and recovery capacity that replace them.

The old Business Case for AI Consumption Pricing

The standard automation case compares a human salary with a software licence.

A customer service officer costs a known amount per year. A KYC reviewer, investigator or credit officer has a salary, a manager, a defined mandate and a finite workload. Their cost is predictable even when individual cases become difficult.

Automated work behaves differently. Its cost is not defined by the number of employees who access a system. It is shaped by the number and complexity of the cases delegated to it.

A more realistic equation is:

Cost per quality assured outcome = model use + data and integration + human review + governance + failure remediation

This is the distinction banks need to make. A cheap automated interaction is not necessarily a cheap customer outcome. A cheap automated document review is not necessarily a cheap and defensible KYC decision.

The management error is to compare the salary of a person with the price of a licence. The relevant comparison is the fully loaded cost of an automated outcome against the fully loaded cost of a human outcome.

Success drives Consumption

The cost issue does not begin when a bank’s system fails. It begins when staff learn to use it well.

The adoption path is usually gradual. Employees initially use a system to summarise documents, draft responses or search internal knowledge. A smaller group becomes more capable and begins to delegate more complex tasks. The bank then connects the workflow to customer data, policies, transaction records and operational systems.

At that point, the system stops being an assistant and becomes part of the operating model.

Enterprise data points in this direction. OpenAI reported significant growth in structured workflows and reasoning consumption as companies moved from experimentation to more embedded operational use. It also found highly uneven usage, with a small group of heavy users consuming far more capacity than the typical employee.¹

The implications are uncomfortable. Management may celebrate rising active users, more automated cases and faster cycle times. Under consumption pricing, each of these success metrics can also drive supplier expenditure.

The more valuable the automated work becomes, the less meaningful a price per user becomes.

Three early Warnings: Consumption shock and Capacity shock

The first warning comes from Uber. The company reportedly exhausted its annual AI tooling budget within four months as engineers adopted Claude Code at scale. It later introduced spending limits on individual tools.²³

This is not a story of useless technology. It is a story of successful adoption under an outdated financial model. The tool was sufficiently valuable that people used it more intensively than the budget anticipated.

The second warning comes from Klarna. The company initially presented automated customer service as a major efficiency success. Later, its chief executive acknowledged that cost had been given too much weight and that service quality had suffered, leading Klarna to recruit human support capacity again.⁴⁵

The third warning is closer to banking. Commonwealth Bank of Australia planned to eliminate 45 customer service roles after rolling out an AI voice bot. It later reversed the decision, accepting that the roles remained necessary.⁶⁷

These examples describe different aspects of the same problem.

Uber illustrates the consumption shock. Usage grows more rapidly than conventional software budgets can accommodate.

Klarna and Commonwealth Bank illustrate the capacity shock. Human capacity is reduced before the organisation has proven that automation can manage the complete operating demand.

Automation exposes the real Work

Banks are especially exposed because their most attractive automation targets are operationally repetitive but rarely simple.

Customer service, onboarding, KYC, periodic review, transaction monitoring, fraud operations and credit administration all contain routine tasks that can be standardised. But each also contains cases where the cost of being wrong is disproportionately high.

The difficult cases do not emerge only after automation. They already exist. Today, they often compete for attention with routine administration, repeat checks, incomplete documents and low value alerts. Officers may know precisely which cases deserve deeper investigation but lack enough time to give them that attention.

Automation can change that.

It can remove administrative work, retrieve evidence, classify documents and prepare a structured case file. Used well, it gives officers more time for judgment, investigation and customer interaction. Research and bank executives increasingly frame the effect as productivity growth and task redesign, even while expecting workforce reductions in some functions.⁸⁹

Process: Customer service
Work automation can reduce:
status questions, standard requests and navigation
Work that needs more attention:
fraud, disputes, complaints, vulnerable customers and trust repair

Process: KYC and periodic review
Work automation can reduce:
document extraction, evidence retrieval and missing document reminders
Work that needs more attention:
complex ownership, conflicting sources, high risk customers and accountable sign off

Process: Transaction monitoring
Work automation can reduce
routine triage and repetitive false positives
Work that needs more attention
connected parties, unusual patterns and suspicious activity decisions

Process: Credit
Work automation can reduce:
data collection and standard scoring
Work that needs more attention:
fairness assessments, exceptions, adverse action reasons and appeals

The opportunity is not automatically to reduce the reviewer population. It is to give reviewers sufficient time to address complex ownership structures, conflicting evidence and higher risk relationships properly.

The risk begins when the bank treats released capacity only as a reason to remove people. Complex cases remain. The bank may then recreate the old problem at a higher level, fewer people, too little time and a greater concentration of decisions that require genuine judgment.

Automation should not be judged by how many people it replaces, but by whether it gives the remaining people enough time to handle the work that always required human judgment.

A smaller Queue is not enough

A lower backlog can be misleading.

A periodic review operation may clear low risk files quickly. That is valuable. Yet the remaining files can still contain complex legal structures, incomplete evidence, higher risk relationships and difficult judgement calls. The important question is not whether the queue has become smaller. It is whether the bank has created more capacity for the cases that most require scrutiny.

The same applies to customer service. A virtual assistant may answer routine questions efficiently, while human employees focus on customers who have already tried and failed to get help. Such customers may be frustrated, vulnerable, in dispute or at risk of leaving.

The mistake is not automation. The mistake is counting only the cases that leave the queue, while ignoring whether the remaining cases are receiving better decisions, faster resolution and more accountable treatment.

This is particularly relevant in KYC and financial crime. Routine handling can consume a disproportionate amount of officer time, but a bank should not assume that releasing that time automatically justifies a proportional reduction in experienced capacity. Better use of the capacity could mean stronger investigations, better evidence quality, earlier escalation or more useful customer engagement.

The Risk is larger than the Model Bill

It would be easy to overstate the evidence and claim that banks have already replaced staff only to discover that their AI costs exceed previous salaries. That specific chain is not publicly established at scale in banking.

The more defensible argument is different.

Banks risk reducing human capability before they understand the full variable cost and operational dependency of automated work.

Model usage is only one cost. A bank must also pay for data preparation, integration, evaluation, controls, human escalation, audit evidence and remediation when something goes wrong. In regulated processes, these costs can grow as a workflow becomes more autonomous.

But cost is only half the issue. The deeper problem is resilience.

Experienced KYC reviewers, investigators and complaint handlers do more than complete tasks. They know where source data is weak, which cases do not fit the standard workflow, how customers behave under stress and when an apparently complete file is not defensible.

This operational knowledge is difficult to document and slow to rebuild.

If a bank reduces those people too early, it may be unable to return to manual processing at scale when a supplier changes its commercial terms, a model becomes unavailable, data quality deteriorates or regulation changes.

The strategic exposure is not merely an expensive monthly invoice. It is a bank that can no longer operate a critical process without its automated workforce.

Revolut’s more mature Lesson

Revolut offers a more sophisticated example of what banking AI at scale requires. It has described deployments across fraud prevention, financial crime operations, customer support and transaction intelligence, including systems handling approximately two million financial crime tasks and up to 1.2 million support tickets per month.¹⁰

What matters is not merely the scale. Revolut has highlighted model routing, observability, cost control, fallback options and retained human accountability for customer affecting decisions.¹¹¹⁰

That is the right lesson.

A bank should not assume that one high capability model is the answer to every task. Stable work such as data validation, document classification and basic status communication should use deterministic workflows where possible. Routine interpretation can use lower cost models. The most expensive reasoning should be reserved for ambiguity, conflict or high risk.

A bank must also know when to stop. An uncertain workflow should not retry indefinitely, consume ever more context and generate an expensive explanation for a case that an experienced specialist could resolve more quickly and reliably.

Govern automated Work

Banks already know how to govern human employees. They set mandates, approval authorities, escalation routes, budgets and accountability.

Automated work needs equivalent controls.

First, every material workflow needs a cost ceiling. The system should have a maximum level of expenditure for a case type. If the cost or uncertainty exceeds that threshold, the case should be routed to a person.

Second, every workflow needs a decision boundary. The bank must define what the system can complete autonomously, what it can prepare for review and what it must escalate immediately.

Third, every workflow needs a named owner. That person should be accountable not only for accuracy, but also for total cost, customer outcome, control evidence and fallback readiness.

The right performance measure is not automation rate, prompt volume or number of cases processed. It is:

Fully loaded cost per quality assured outcome.

For periodic review, that means a review completed with current evidence, appropriate escalation and no material quality finding. For customer service, it means an issue resolved without repeat contact, avoidable transfer or complaint.

This shifts the incentive from maximising automation to maximising safe and sustainable value.

Price it, stop it, run without it

A board does not need a lengthy AI strategy to test whether an automation programme is sound. It needs three questions.

Can we price it?

Can the bank calculate the fully loaded cost of an automated outcome, including human review, controls, remediation and repeat work?
If it cannot, the project has not yet established its business case.

Can we stop it?

Can the bank impose spend limits, reduce autonomy, reroute work to another approved model or pause the workflow when costs rise or quality deteriorates?
If it cannot, the bank has given an external provider significant control over a critical operating cost.

Can we run without it?

Does the bank retain enough experienced people, documented procedures and manual fallback capacity to operate when automation is unavailable or no longer economic?
If it cannot, the bank has not automated a process. It has transferred operational sovereignty.

European supervisory attention is increasingly focused on governance, accountability and risk management, particularly for high risk use cases such as creditworthiness and credit scoring. Banks should apply the same seriousness to the economics and continuity of automated operations.¹²¹³

The real Board Question

The wrong question is: “How many people can this replace?” It encourages banks to treat people as a cost line and automation as a fixed investment. The better question is:

How much discretionary spending authority, customer impact and operational dependency are we prepared to delegate to an automated workflow for one customer case?

A KYC reviewer has a salary, a mandate, an escalation path and a manager. An automated workflow needs the equivalent:

• a cost ceiling,
• a defined authority and
• a human who can take over.

The next banking cost crisis may not come from a failed automation project. It may come from a successful one: a bank automates routine work, reduces its operational workforce and then discovers that its digital workforce is more expensive, less controllable and harder to replace than anticipated.

That is not simply a technology problem. It is a question of who still controls the bank’s capacity to operate.