The hidden cost of AI: What businesses need to model before the bill arrives

Artificial intelligence is becoming increasingly accessible and, at first glance, appears relatively inexpensive to use. It’s starting to resemble the space race, with the demand for more features and increased usage ramping up each day.

A monthly subscription gives employees access to tools that draft reports, analyse documents, generate presentations, and assist with research. Of course, the quality of each output can (and does) vary significantly. Many businesses are also exploring more advanced applications, from customer-service automation to coding assistants and AI agents that complete multi-step tasks with limited human intervention.

Businesses are investing in AI with the expectation that it will improve productivity, streamline processes, and unlock new ways of working.

But stepping back from the top-down view of efficiency / savings, the cost question is quickly becoming more nuanced; demanding attention.  

The amount shown on an invoice may not capture the full economic impact of AI adoption. Subscription fees are only the most visible part of the picture and, in many cases, have been deliberately priced at a loss to drive uptake and usage. As the investment mania settles, and the access to bottomless capital dissipates, AI providers are coming under increasing pressure to demonstrate a path to profitability.  

As AI usage increases, businesses need to understand token consumption, model selection, data-centre capacity, electricity requirements, cooling and water usage, governance, and what can often be overlooked (and at times, the highest cost), the time required to validate outputs. For finance teams, this creates a challenge. Cost drivers are evolving quickly and becoming opaque. In simple terms, the ongoing costs of a tool that initially looked inexpensive during rollout, can look very different after wider adoption and being embedded into automated workflows.

Businesses need a more disciplined way to understand what they’re consuming, what value they are receiving, and how the economics may change as their usage matures.

AI pricing models: subscriptions per user vs token consumption

Many businesses first encounter generative AI through a relatively simple pricing structure of monthly charges per user. This resembles the software-as-a-service model that businesses have used for years. Finance teams can multiply the number of licences by the monthly fee and build a reasonably predictable budget.

However, AI pricing is beginning to evolve.  

For example, GitHub announced that its Copilot plans would move to usage-based billing from 1 June 2026. Monthly subscriptions remain, but each plan includes an allocation of AI credits. Additional usage is calculated based on token consumption, including input tokens, output tokens, and cached tokens – using the published application programming interface rates for each model.

GitHub Copilot has evolved from an in-editor assistant into an agentic platform capable of running longer, multi-step coding sessions, and iterating across entire repositories. A quick question and a multi-hour autonomous task can create very different demands on computing infrastructure. A pricing model that treats them as equivalent becomes difficult to sustain.

As AI tools become more capable, software providers will need to decide how much usage they can reasonably include within fixed subscription fees. Businesses may gradually find that the familiar per-user licence is only one component of the bill, particularly when employees use advanced models or agents undertake complex tasks in the background. Future cost models will thus need to reflect how the tech is actually being used. 

What is an AI token? Understanding the new AI cost structure

A token is a small unit of data processed by an AI model. Every prompt, document, spreadsheet, response, and follow-up instruction adds to the volume of tokens consumed. The model may also need to retrieve background information, interpret a long conversation history, or review its previous output before it completes a task. Each of these steps means more tokens consumed.

At a small scale, this can feel immaterial. But at an enterprise scale, the numbers increase quickly.

Tokens are an emerging unit of cost for AI, with implications for operating expenses, margins, forecasting and capital planning. A basic chatbot may generate millions of tokens per subscriber each year, while more advanced ‘super agent’ applications can generate substantially more.

The cost per token may fall as models become more efficient and competition increases. However, total expenditure can still rise if businesses consume tokens faster than prices decline.

Prompts are not tokens

Concentric AI’s 2026 Data Risk Report found that an average of 73% of employees used public AI applications. Users submitted an average of 6.5 prompts per day, rising to 20 prompts per user on peak usage days.

Those figures show strong engagement, which is encouraging. They also reinforce the need for better measurement. A prompt count gives us a useful starting point, but it does not tell us how many tokens were consumed, which models were used, whether employees generated short answers or detailed reports, and how much value each workflow created.

 The hidden cost of AI governance and correcting AI hallucinations

Direct financial costs are only part of the equation. AI tools can produce work that looks polished and plausible while containing inaccuracies, missing important context or drawing unsupported conclusions. This is particularly relevant when employees use AI to assist with research, financial analysis, legal documents, client communications or strategic decision making.

An output that saves an hour of drafting time may still require careful review. In some cases, the human validation process can offset a meaningful share of the productivity gain.

There is also a risk that teams become less careful as the quality of AI-generated writing improves. A poorly written response is easy to question. A fluent and confident response can be more dangerous if it encourages the reader to accept an incorrect answer without checking the underlying assumptions.

Finance teams understand this problem intuitively. A model can be technically sophisticated and beautifully presented while still producing unreliable forecasts if the assumptions are weak. AI outputs deserve the same scrutiny.

The cost model should therefore include the time required for review, correction, and escalation. Higher-risk use cases need stronger controls. Businesses should also consider the financial impact of errors, especially where an AI-generated recommendation informs a decision involving customers, capital allocation, regulatory obligations, or reputation.

Good governance is part of the economics and should not be treated as a separate compliance exercise added after deployment.

The important role of the finance function 

Wolters Kluwer’s 2026 Future Ready CFO Survey shows that among senior finance leaders in APAC, 83% said AI adoption and implementation is a key force reshaping finance and 72% expect AI to have a significant impact on their finance function within the next three years. They identify financial planning and analysis, forecasting and scenario modelling, and risk management and compliance monitoring as the activities most likely to be transformed.

However, adoption without measurement creates a blind spot. Finance teams do not need to own every technical decision. They need enough visibility to understand how usage translates into cost, how the business is measuring productivity and when the economics justify a different approach.

That may require closer collaboration between finance, technology and operational teams. It may also require new reporting disciplines: usage by team, cost by workflow, model selection, productivity gains, review time and the value created from specific use cases.

The most useful financial models make uncertainty visible. AI adoption is a good example of where that discipline matters. The cost structure is still developing, usage patterns are changing, and some of the largest expenses may emerge only after the technology becomes embedded in everyday work.

The inputs that matter when modelling AI investment

Traditional software budgeting tends to focus on licence fees, implementation costs, support, and training. AI requires a broader approach because usage can become more variable as adoption expands. A practical financial model should consider several layers. 

• Adoption: How many employees will use the tool? How frequently will they use it? Will usage grow gradually or accelerate as teams discover new applications? 
• Workflow intensity: How many tokens are consumed by a typical task? Does the tool answer a single question or complete several steps? Does it retrieve documents, process images, generate presentations, or run continuously in the background? 
• Model selection: Which tasks genuinely require a premium model? Could a smaller or less expensive model handle routine work such as classification, extraction, or basic summaries? A business that sends every task to the most advanced available model may pay for capability it does not need. 
• Infrastructure exposure: Is pricing fixed, metered, or likely to become more consumption based? Does the business rely on a software provider, a cloud platform, or privately hosted infrastructure? How might energy, cooling, and capacity constraints affect costs over time? 
• Human oversight: How much review is required? Which outputs can employees use as a first draft? Which decisions require formal validation, audit trails, or approval workflows? 
• Value creation: What measurable benefit does the business receive? Does the tool reduce time spent on repetitive tasks, improve the speed of analysis, support better customer service, or help employees complete work that would otherwise require specialist support?

These inputs allow businesses to test a range of scenarios, identify the variables that have the greatest influence on cost, and understand how seemingly small changes in usage can have a material financial impact. 

Modelling AI costs for your business

Robust financial modelling can help organisations move beyond assumptions and develop a clearer understanding of how AI may affect costs, productivity and business performance over time.

RSM Australia's financial modelling specialists work with businesses to evaluate investment decisions, test scenarios and build models that support confident decision-making. Get in touch to learn more about how we can help you make confident decisions about AI in your business.