The discipline gap: why banks must manage AI the way they manage capital
Managing AI as an investment portfolio; connecting cost, consumption and outcomes for holistic value management
Banks are among the most disciplined allocators of capital in the economy. Yet their fastest-growing investment category – AI – is largely escaping that discipline. Closing the gap will become a defining test of leadership, and those organisations that address it soonest will set the pace for the future ahead.
Key Takeaways
A tale of two portfolios
Consider how a bank lends: before a pound leaves the balance sheet, the borrower is assessed, the exposure is priced and terms are agreed. The position is then monitored throughout its lifecycle. If performance declines, the bank responds. This discipline is so embedded it is almost invisible, becoming a fundamental part of how the bank works.
Now, consider how such an organisation invests in AI. Use cases may pass through rigorous technology, risk, and governance processes, yet costs, value, and performance may not be consistently managed with the same rigour, neither before nor post-production. As a result, the total cost of operating an AI solution at scale may not be fully understood, expected value may not be expressible in measurable terms, an accountable executive may not be assigned and performance may not be tracked consistently over time.
If banks manage their lending as a connected portfolio of investments with clear economics, lines of accountability and performance discipline, why then are the same disciplines not applied to AI?
UK financial services firms were already using AI by 20241
AI value is cited almost twice as often as cost as a barrier to adoption2
of businesses report no revenue impact from AI adoption4
From licenses to tokens: AI creates a new economic model
AI has moved rapidly from experimentation to core investment. By 2024, three-quarters of UK financial services firms were already using AI,1 with the average number of AI portfolios were expected to more than double by 2027.2 Yet as investment accelerates, many organisations are discovering that AI requires a fundamentally different set of economics from other, more traditional technology areas.
Historically, technology costs were relatively predictable. Organisations purchased licenses, deployed infrastructure and managed expenses that were largely fixed or linked to workforce growth. Generative and agentic AI change that equation. Every prompt, response, retrieval, reasoning step and agent action consumes tokens. And, unlike traditional software, the cost of AI does not settle down once a solution is deployed. It continues to accrue with every task performed, evolving with the changing burden of inference placed upon it by the organisation, making consumption a critical economic driver.
This shift makes ‘tokenomics’ the foundation of AI value management. Tokenomics describes the discipline of translating token consumption into cost, understanding how costs evolve as usage scales and how technical, usage and design decisions influence the economics of delivering outcomes. In practice, this means decisions that once sat largely within technology teams – such as model selection, context length, retrieval architecture and agent orchestration – now carry direct financial consequences. In consequence, two use cases delivering similar business outcomes may have materially different operating costs depending on factors such as model choice and prompt efficiency. And a solution that appears viable at pilot stage may become significantly more expensive when deployed across thousands of employees or millions of customers depending on adoption and use.
As AI portfolios grow, understanding token consumption is therefore no longer just a technical exercise. It becomes a business necessity. Organisations need to understand not only what AI is costing them today, but how those costs will evolve as adoption, complexity and autonomy increase. Yet, even understanding token consumption alone is not enough.
Grasping the economics of AI consumption will help organisations understand what AI costs and why. However, understanding cost is only part of the challenge. Organisations must also determine whether AI is generating sufficient value and whether that value is being realised over time.
This is where many AI programmes struggle – in articulating value, achieving scale and proving the value promised at inception.
Why AI value discipline breaks down
AI is not the first transformation investment to struggle with value realisation, of course, and the challenge facing most organisations is not access to AI itself, but the absence of a consistent discipline for evaluating, prioritising and managing AI investments.
Together, these weaknesses expose three fundamental questions that many organisations struggle to answer consistently:
1. What value should their AI investment create?
2. Is the value identified worth the cost and risk required to deliver it?
3. How do we know for sure that value is being realised after deployment?
Deloitte's AI Value Management Framework addresses these questions through three disciplines: Define, Decide and Deliver (see Figure 1).
Source: Deloitte
Define: What value are we creating?
Before organisations can begin to manage AI value, they need a clear answer to a deceptively simple question: what does value creation through AI really mean?
Many AI investments are justified by what the technology can do rather than the outcomes it is expected to deliver. Yet AI alone creates no value. Instead, value is realised when AI improves business outcomes, whether through increased revenue, lower costs, reduced losses, better decisions, stronger customer outcomes or greater organisational productivity.
Organisations should therefore define AI value holistically but measure it precisely across two dimensions:
1. Financial value: revenue growth, cost reduction, avoided losses, and capital efficiency.
2. Non-financial value: productivity, faster cycle times, better decisions, reduced risk, and improved customer and employee outcomes.
AI Capability
Business Outcome
Enterprise Value
These are not competing measures. In many cases, non-financial value is the mechanism through which financial value is realised. For example, faster onboarding can improve customer retention, while better fraud detection can reduce losses and improved productivity can increase capacity without increasing cost. The discipline lies in making those relationships explicit and measurable.
The strongest organisations define value before funding is approved. Every use case should have a clearly articulated business problem, measurable outcomes, an agreed baseline and an accountable owner. The objective is not to describe what the AI can do, but to establish the business result it is expected to deliver.
Value is not what AI does. It is the measurable business outcome AI enables.
Decide: Is the value identified worth the investment?
A compelling value case is only part of the equation. Organisations also need to understand the economics required to achieve it.
This requires a view of the Total Cost of Ownership (TCO) for AI across the full lifecycle of a solution, including technology, data, integration, governance, talent, adoption, and ongoing support. For generative and agentic AI, tokenomics becomes an important added component of this assessment. Unlike many traditional technology investments, these costs can vary significantly with usage, making consumption a key economic variable.
As a result, organisations need to model how costs evolve at scale. Adoption rates, interaction volumes, model selection and agent complexity can all have a material impact on the economics of a solution.
However, the objective is not to minimise token consumption. It is to maximise returns. Organisations should compare expected value against total cost, test how the business case performs under different usage scenarios and assess whether AI represents the most attractive option relative to alternative approaches.
Hence, the critical question is not how many tokens are consumed, but whether the value that consumption creates justifies the investment required to deliver it.
Deliver: Are we realising the value?
The final challenge begins after deployment.
Many organisations have robust processes for approving AI investments, yet far fewer apply the same rigour once solutions enter production. Attention can instead shift towards operational metrics, while the original value case receives less focus.
Yet, this is precisely where value management becomes most important, requiring expected outcomes to be compared with realised outcomes and consumption to be monitored alongside business impact, with clear accountability for whether the promised benefits are actually being delivered. The goal is not simply to confirm that a model is working, but to understand whether it is creating measurable business value.
Every significant AI use case should have a defined baseline, agreed success metrics, an accountable business owner and a regular review cycle. This creates a direct link between investment, consumption and realised outcomes, allowing organisations to assess whether value is being created as expected.
Importantly, value realisation is not a one-off exercise. Costs, usage patterns and business priorities change over time. Use cases that outperform expectations should be scaled, while those that fall short should be optimised, redesigned or retired. The goal is not simply to validate past investments, but to continuously reallocate resources towards the use cases generating the greatest value. This requires organisations to treat AI as an actively managed portfolio rather than a collection of technology projects – with clear ownership, regular review and a willingness to challenge investments based on evidence rather than enthusiasm.
Value is not created when a use case is approved. It is realised, or lost, through the decisions made after it goes live.
Where to start
Organisations do not need to wait for a major transformation programme to begin. Much of the required discipline can be built through existing investment, finance and governance processes. We recommend the following steps:
Together, these disciplines transform AI from a collection of projects into a managed investment portfolio. The earlier they are established, the easier they are to apply consistently as AI adoption and consumption grow.
The timing is important too. AI portfolios continue to expand rapidly, and investment is accelerating across the market. Disciplines that are relatively straightforward to establish today will become significantly harder to retrofit once hundreds of use cases are in production.
Bottom line
As AI capabilities become more widely available, access to models alone will offer limited and often short-lived differentiation. More durable advantage will come from deciding where AI should be used, understanding its full economics and continuously allocating and reallocating resources towards use cases that deliver measurable value.Banks already apply this discipline to capital. The same institutions must now apply that same rigor to AI.
The advantage for banks is not simply in better AI. It is in better AI economics.
References
1. Artificial intelligence in UK financial services - 2024 | Bank of England
2. Management practices and the adoption of technology and artificial intelligence in UK firms - Office for National Statistics. Note: ONS is currently undertaking research to produce more recent figures for 2026.
