In Case You Missed It
Is AI Blunting Your Strategy? – As strategy development relies increasingly on LLMs, true differentiation from competitors is becoming more elusive.
Executive Summary
AI spending is rising quickly, and model charges are only part of the cost.
Budgeting for AI is challenging as unit prices are falling, usage is rising, and AI agents are changing consumption patterns.
Productivity gains are often hard to convert into realised financial benefit.
For budgeting purposes, organise AI spend into five funding categories, each with different owners and controls.
Approve a funding envelope, maintain a rolling forecast, and release funds in stages as the evidence strengthens.
Unpredictable Spend, Untraceable Returns
Budget season is upon us, and the battle lines are already being drawn. On one side, the advocates – colleagues who are already seeing benefit from AI usage and would like to ramp up spend. On the other, the sceptics – those who are finding it hard to see where AI is hitting the bottom line and prefer a more disciplined approach.
That tension is a normal and healthy part of any technology budgeting exercise. While there will always be gamesmanship, the debate can be mediated using tools such as benchmarking, cost-driver analysis (e.g., headcount) and project ROI ranking.
Yet budgeting for AI is harder. AI spend is growing but not in a predictable way. Hidden costs, such as verification of AI output, need to be accounted for. Returns are both obvious and invisible.
Bain’s 2026 survey of 102 CFOs found that 83 per cent expect to raise enterprise‑wide AI spending by more than 15 per cent over the following two years and 42 per cent expect an increase of at least 30 per cent. Yet only 31 per cent were highly satisfied with current outcomes.
The harsh reality is that the agreed budget number will probably be wrong within months. As one analyst put it, a lot of CFOs are about to see their AI bill and “freak out.”
Cheaper and More Expensive
Token costs are collapsing. Stanford’s AI Index found that the price of querying a model at GPT-3.5 performance fell about 280-fold between November 2022 and October 2024.
But lower prices stimulate usage, as students of Jevons know. As more capable AI models are released, new use cases become viable. For example, early LLMs couldn’t handle complex customer interactions, but the latest models perform far better.
More significantly, use of AI is shifting from synchronous chatbot interactions to autonomous agents. An agent can call other models, invoke tools and repeatedly iterate a task. Agents also spin up other agents, making a mockery of headcount-based budgeting. Claude Code creator Boris Cherny says he sometimes runs hundreds of agents in parallel.
Similar dynamics were seen as other general-purpose technologies emerged. The cost of a unit of electricity in the US fell by more than 97 per cent between 1902 and 1950, yet total spending on electricity rose, because cheaper power led to people buying more electrified devices.
And the AI bill is climbing. In BCG’s 2026 survey, respondents expected their organisations to spend about 1.7 per cent of revenue on AI, more than double the 2025 number, although with significant differences by sector. Businesses on Ramp that spend on AI allocate nearly 15 per cent of their software budget to AI tools, on average.
AI models themselves are only part of the total bill. Workflow redesign, data, systems integration, evaluation, security, human review and upskilling also need to be factored in. BCG’s 10-20-70 rule of thumb for AI transformation splits the spend 10 per cent to algorithms, 20 per cent to data and technology, and 70 per cent to people and processes.
Some of the increase in AI costs is being driven by the shift to agentic workflows. EY estimates that the cost of a single customer service interaction has risen from about four cents in 2023 to around $1.20 in 2026, a thirtyfold jump. The token price per unit fell over that period, but the workflow around the model drove up costs due to reasoning steps and repeated attempts by agents to complete tasks.
Understanding Your AI Spend
The first task in an AI budgeting exercise is to find the current spend. AI spend is buried in SaaS renewals, cloud accounts, consulting projects, data teams and shadow AI. Sometimes AI spend arrives unseen, when a software upgrade switches on AI features. KPMG found that only 26 per cent of large US organisations had full visibility into the cost of running AI at scale.
SaaS spend shouldn’t simply be classified as AI spend simply because AI is bundled into the product. Include only clearly identifiable additional AI costs. Once the scope is agreed, organise the spend into five practical funding categories, each with different owners, drivers and controls:
Committed operating costs are contractually committed, or broadly fixed, within the planning period. They include licences, subscriptions and embedded AI features bought at a flat price or agreed contract tier. These costs should be funded and managed by the function or business unit consuming them, using centralised procurement for leverage.
Usage-variable operating costs are the costs of keeping AI running in production that vary with usage. They include model calls, agent runs, retrieval of information and associated human review activities. These scale with volume, which in turn is driven by the number of users, intensity of use and AI models used. A brand designer using an image model day in, day out may generate far higher usage costs than an executive who seeks occasional help to draft emails. The added complexity in this category is the proliferation of AI agents. As users deploy more agents, and agents spawn further agents, usage can grow non-linearly. These costs should also be funded and managed by the function or business unit consuming them but should be subject to central controls (e.g., metering, quotas) to avoid surprises. And surprises do happen. Uber reportedly exhausted its 2026 budget for AI coding tools by April.
Build and change costs are the costs of getting new AI into the business. They include project management, workflow redesign, systems integration, data preparation (substantial if existing data is of poor quality), user testing, and upskilling people. Any organisation whose systems cannot yet support agentic workflows must close that capability gap first, and for some that foundational cost will be substantial. For firms building or tuning their own AI models, model training costs will also be a factor. This expenditure is discretionary, at least until projects are green-lit. These costs should be funded like any other project, from a central change budget or the sponsoring unit, with funding released against milestones. Two things make AI projects harder to budget than other IT projects. First, AI systems need regular checks after launch because changes in data, workflows or the underlying model can cause their performance to ‘drift’ over time. That can make ongoing maintenance more costly and less predictable than other IT systems. Second, for firms that train or fine-tune their own models, build costs are challenging to estimate upfront, as there are many variables, such as the quality of training data, that cannot easily be pinned down until work starts.
Shared platform costs cover the common foundation that every function and business unit draws on. They include infrastructure, tooling, the model-access layer and shared data. This cost is step-fixed, growing as capacity and teams are added, and should be funded centrally and recharged based on consumption once reliable measurement is in place. Until then, it can be held centrally as an investment.
Control and assurance costs are the costs of keeping the entire system safe and defensible. They include evaluation, monitoring, security, legal, governance and audit support. These costs are semi-fixed, increasing with the number of deployments and regulatory scrutiny rather than volume. They should be funded centrally.
R&D costs and business unit experiments are not a sixth category. Rather, they are combinations of these five, with limited funding and an explicit decision to stop or scale up.
These categories combine purpose, cost behaviour and ownership to support budgeting and management. They are not accounting classifications. Individual build, platform and control costs may still be fixed, step-fixed or variable with usage. Some investments may be eligible for capitalization.
One risk is that costs assumed to be fixed can unexpectedly become variable. This can happen if agents drive usage above a contractual usage threshold and into metered pricing without appropriate controls.
Where the Value Goes Missing
AI business cases often project productivity gains that don’t easily translate into benefits. Employees report saving a few hours a week, which translates, in theory, to a significant benefit. But time saved is only crystallised when the business converts it into lower costs or higher revenue. Otherwise, it may still create value through faster service, better quality or lower risk. Time saved in minutes across hundreds of people must be aggregated before it hits the bottom line.
This is the same value leakage that I explored in an earlier article, which illustrated how the share of an AI investment’s potential value that reaches the P&L, its Deployment Yield, can be strikingly low.
The difficulty of turning time saved into money saved is not new. But the nature of AI exacerbates the problem. Past IT system deployments often landed on one team or function, where the time savings could be pooled and released as headcount savings. AI hands a few minutes back to almost everyone, so the time saved is more dispersed. And some of the saved time is spent checking and correcting AI’s output itself. Workday’s January 2026 survey of 1,600 leaders and 1,600 employees found that nearly 40 per cent of AI’s reported value was lost to rework and misalignment. Usage costs also rise as adoption expands, which further eats into net savings.
Benefits are even harder to project confidently than costs. In a 2026 survey of more than 900 companies, Bain found that 20 per cent of firms had realised savings of 10 per cent or less, yet 90 per cent of firms were raising their AI budgets again.
There is also a timing trap. In the near term, AI may increase the total cost base. For a while, companies will pay for the AI, its implementation, and running old and new workflows in parallel. Then-Walmart CEO Doug McMillon said AI would change “literally every job.” Separately, chief people officer Donna Morris said the company expected its global headcount to remain broadly flat for three years.
Fund Uncertainty in Stages
Addressing these issues requires a novel approach to AI budgeting, based on an annual funding envelope, a rolling forecast and a mechanism for releasing funds as evidence emerges. Funding should follow deployment in stages and be adjusted as costs, usage and value become clearer, rather than being fixed in advance. The system has five moves borrowed from other funding models but adapted to AI’s particular dynamics:
Fund at the edge, control at the centre. Consumption-versus-value trade-offs are best made by those close to the use case, so the associated costs should be owned and managed by operating teams. Agents can run up a huge bill unexpectedly, so devolved consumption relies on the centre to apply controls such as usage limits.
Force a rethink at each stage gate. As with other projects, AI funding should be released in stages. But each gate should reconsider not only delivery progress, but whether the use case, business case and technical solution still make sense. A newly released AI model may have removed the need to build a bespoke solution, for example.
Hold a reserve, not only a contingency. Keep an explicit central reserve for opportunities that become viable during the year as models improve or unit costs fall. Contingency covers unexpected costs in projects already approved. The reserve funds new projects that become viable.
Reforecast based on underlying drivers. A normal quarterly reforecast updates volumes and prices within a stable model of the business. With AI, the unit economics can shift quickly. The price per unit may fall while consumption per task rises. So, the forecast needs to be repriced based on a deeper review of the underlying cost drivers.
Seek proof of value, not promises of it. All projects start with promises of benefits and timing, and AI is no exception. AI investments face additional risks such as dispersed savings and unexpected usage costs so the burden of proof should be higher. Proof of value should include measurable benefits, net of the costs to deploy, operate and verify.
In this way, the budget becomes a control system as well as a funding plan. Fund shared platforms and controls centrally, rethink at each stage, hold a reserve, reforecast based on drivers, and release funding only as evidence of value improves.
Whatever goes into the AI budget, it’s unlikely to please everyone. But with the right budgeting system in place, everyone should understand where the money goes, who’s spending it and what you’re getting for it.
Monday Morning Actions for Executives
· Build a spend snapshot. Ask finance and IT for a full picture of current spend.
· Put guardrails on variable spend. Set alerts and monthly limits.
· Challenge live pilots. Ask for proof of value or a timeline for seeing it.
Questions the Board Should Ask
· Can management show us the full cost of AI, including hidden costs?
· How much of the spend is committed, and how much could rise with usage?
· Which use cases have generated realised financial benefits, and why?
· Are we funding the full deployment chain, or only models and tokens?
Sources & Notes
· Bain & Company, 2026 CFO survey: In a survey of 102 CFOs, 83 per cent expected AI spending to rise by more than 15 per cent over two years, 42 per cent by at least 30 per cent, and only 31 per cent were highly satisfied with current outcomes.
· The Wall Street Journal, 5 June 2026: D.A. Davidson’s Gil Luria warned that many CFOs would “freak out” when usage-based AI bills arrived, while KPMG found that only 26 per cent of large US organisations had comprehensive visibility into AI costs.
· Stanford HAI, 2025 AI Index Report: The cost of querying a system at roughly GPT-3.5 performance fell more than 280-fold between November 2022 and October 2024.
· WIRED, 26 May 2026: Claude Code creator Boris Cherny said he regularly runs dozens, and sometimes hundreds, of agents for eight to twelve hours at a time.
· Vaclav Smil, via IEEE Spectrum: In real terms, US residential electricity prices fell more than 97 per cent between 1902 and 1950 while consumption rose.
· BCG, AI Radar 2026: In a survey of 2,360 executives, respondents expected AI spending to rise from 0.8 per cent of revenue in 2025 to about 1.7 per cent in 2026, with significant differences by sector.
· Ramp, The $1 Trillion AI Spend Blind Spot, 2026: Among businesses on Ramp that spend on AI, the median company allocates nearly 15 per cent of its software budget to AI, based on anonymised transaction data across Ramp’s customer base.
· Business Insider, May 2025: BCG X leader Sylvain Duranton recommended allocating roughly 10 per cent of AI-transformation effort to algorithms, 20 per cent to data and technology, and 70 per cent to changing how people work.
· EY, Agentic AI Enterprise Token Cost, 2026: EY illustrates how a four-cent customer-service interaction can become a $1.20 agentic orchestration when tools, reasoning and repeated attempts are added.
· The Financial Times and The Verge, May 2026: Uber reportedly exhausted its 2026 AI coding-tools budget by April, then capped each employee’s monthly token spending at $1,500 per tool as executives questioned the value of rising usage.
· Deployment Yield: The share of an AI investment’s potential value that reaches the P&L, developed in an earlier article in this series.
· Business Insider, 14 January 2026: A Workday-commissioned survey of 1,600 leaders and 1,600 employees found that nearly 40 per cent of AI’s reported value was lost to rework and misalignment.
· Bain & Company, 2026 survey of more than 900 companies: Two in five reported savings of 10 per cent or less, while about nine in ten planned to increase AI budgets.
· The Wall Street Journal, 27 September 2025: Then-CEO Doug McMillon said AI would change every job, while chief people officer Donna Morris said Walmart expected headcount to remain broadly flat for three years as roles changed.