The first phase of workplace AI adoption had a fairly simple message: use it. Give people access. Experiment. Put copilots into workflows and look for opportunities to automate work or improve productivity.
For a while, high usage was almost treated as evidence of progress. Silicon Valley even gave it a name: tokenmaxxing.
Now the bill is starting to arrive.
Businesses that have become accustomed to relatively predictable software licences are discovering something different with AI. Usage can increase quickly, costs are often consumption-driven and a seemingly straightforward task can involve far more processing than the user sees.
Gartner warned in June that rising token consumption and the move towards consumption-based licensing are putting pressure on enterprise AI budgets. It predicts that, by 2028, the cost of AI coding could exceed the average developer salary as token consumption increases.
Recent reporting suggests the issue is already being felt. TechCrunch reported that Uber exhausted its 2026 AI coding budget by April, while a Priceline employee said a routine renewal for its Cursor coding platform came back four to five times more expensive. Microsoft has also introduced greater controls around internal AI consumption.
This does not make AI less valuable.
It does mean businesses need to understand what they are consuming.
The next phase of adoption requires organisations to move beyond asking how much AI they are using and start asking a more important question: what value are we getting from the intelligence we are paying for?
In this article we will cover:
- what tokens are and why businesses should care about them
- why AI costs can rise rapidly as adoption increases
- how agentic AI changes the economics again
- why AI dependency without cost visibility creates business risk
- why value per token is becoming a more useful measure of AI maturity
What is a token and why does it matter?
The terminology can make this sound more complicated than it needs to be.
A token is essentially a small unit of information processed by an AI model.
When you ask an AI system a question, the information it reads consumes input tokens. The response consumes output tokens. Depending on the system, additional processing may be required for reasoning, conversation history, retrieved documents, searches, tool use and other activity required to complete the task.
Microsoft describes tokens as similar to mobile data or call minutes. More tokens generally mean more computation, which in turn means greater cost.
For somebody occasionally using ChatGPT to help write an email, this is largely invisible.
At organisational scale, it starts to matter.
Most businesses understand the economics of conventional SaaS. Fifty people need fifty licences. The annual cost is reasonably easy to forecast.
AI increasingly introduces another variable: consumption.
Cost can change according to how many people are using the technology, which models they select, how large the information being processed is, how long conversations become and how many steps an AI system undertakes before completing its work.
The user may not even see those tokens directly. They can sit behind subscriptions, credit systems, usage allowances or software platforms that incorporate AI into a wider service.
But the underlying resource is still being consumed.
As AI moves further into everyday business operations, understanding that consumption becomes increasingly important.
Why AI costs can rise even when the technology becomes more efficient
There is an important distinction in the economics of AI.
The underlying cost of performing AI inference is falling rapidly. Gartner forecasts that the cost to AI providers of running inference on a very large model could fall by more than 90% between 2025 and 2030 as hardware and infrastructure improve. Goldman Sachs also expects AI chipmakers’ unit costs to continue falling.
But that does not mean the practical cost of AI to businesses will necessarily fall.
Consumption is expanding at an extraordinary rate.
Goldman Sachs forecasts that growing consumer and enterprise adoption of AI agents could drive a 24-fold increase in token consumption by 2030 compared with 2026.
This creates a fairly simple economic effect.
The cost of processing an individual unit of intelligence can fall while the amount of intelligence being consumed rises much faster.
AI is also being asked to do considerably more.
It is moving from occasional prompts into coding, document analysis, customer service, research, marketing, finance, legal work, procurement and operational systems. Existing software providers are adding AI functionality while organisations are building their own applications and agents on top of foundation models.
Each new use case creates additional consumption.
So while the technology itself is becoming more efficient, the experience for many businesses is entirely consistent with rising AI expenditure.
This is one reason token economics are starting to resemble the earlier development of cloud computing. Cloud made infrastructure dramatically easier to access and scale, but businesses eventually discovered that scalable infrastructure also required financial discipline.
AI appears to be reaching that point much more quickly.
Agents multiply the problem
Agentic AI changes the economics again.
A conventional chatbot exchange is relatively easy to understand. A person sends a prompt and receives a response.
An agent can operate very differently.
Give it an objective and it may search for information, retrieve company data, read documents, use software tools, analyse results, revise its approach and repeat parts of that process before completing the task.
The employee still sees one instruction and one eventual outcome.
Behind it can sit many individual interactions with an AI model.
Microsoft Research published a study in April examining token consumption across agentic coding tasks. It found that agentic tasks could consume around 1,000 times more tokens than conventional code reasoning or code chat. Token consumption for the same task could also vary by as much as 30 times between different runs, while higher consumption did not consistently produce greater accuracy.
That introduces a completely different challenge for business forecasting.
The amount of AI resource required to complete a piece of work is not necessarily obvious from the complexity of the request.
Nor does spending more automatically produce a better outcome.
As organisations begin deploying agents across operational processes, understanding how those systems behave becomes just as important as understanding what they can do.
The bigger risk is dependency without understanding
Cost is only part of the issue.
AI adoption inside many organisations has happened from the bottom up.
Someone started using ChatGPT. Developers introduced AI coding tools. Marketing adopted another platform. Finance gained access to a copilot. Existing software suppliers began adding AI features into products people were already using.
Each decision may be perfectly sensible.
Collectively, however, they can create an AI operating model that nobody actually designed.
A business can gradually become dependent on AI without having a complete picture of which providers it relies upon, which models are being used, what information is being processed or how consumption changes as adoption grows.
This matters because AI is starting to move from an employee productivity tool into infrastructure.
When a workflow becomes dependent on a particular model or platform, changes to pricing, usage limits, model availability or commercial terms become operational considerations rather than simply software purchasing decisions.
KPMG’s latest Global AI Pulse research illustrates the problem particularly well. One third of leaders identified a limited understanding of AI usage costs as a significant challenge when deploying agents. Organisations with strong visibility over their AI operating costs were five times more likely to report established ROI than those without it.
This is an important connection.
Cost visibility is increasingly part of AI capability.
A business cannot confidently scale something if it does not understand what happens economically when usage increases.
From tokenmaxxing to value per token
This is why the conversation around tokenmaxxing is changing.
AP reported in August that businesses which had initially encouraged heavy AI consumption are beginning to focus much more carefully on efficiency, model selection and the business value created by that usage.
Uber provides a useful example.
Its use of frontier AI tools has reportedly quadrupled since the beginning of 2026, but it is now focusing on controlling costs through better prompt caching, model selection, improved visibility and greater use of open-weight alternatives.
This feels like a much more mature phase of adoption.
Organisations are beginning to understand that model choice, workflow design, context management and information architecture have commercial consequences.
The largest and most sophisticated model may be entirely justified for complex reasoning. A simpler task might be completed perfectly well with a smaller model at a fraction of the consumption. Information that is continually reprocessed can sometimes be structured or retrieved differently. An agent that repeatedly retries a task may need redesigning rather than simply being given a larger budget.
The goal is not to minimise AI usage.
It is to understand whether the resources being consumed are proportionate to the outcome.
That creates a much better measure: value per token.
The value might come through increased revenue, lower operating costs, faster delivery, reduced risk, better customer outcomes or allowing people to spend more time on valuable work.
It will be different for every organisation and every use case.
But there needs to be an outcome.
Token consumption on its own tells us very little about productivity.
AI adoption needs to become a business discipline
The current situation should not be surprising.
AI moved extraordinarily quickly from experimentation to widespread business adoption.
Commercial understanding has not always kept pace.
KPMG’s Q1 Global AI Pulse found that 95% of organisations surveyed already had an AI strategy, yet only 8% reported established ROI. Its more recent Q2 research shows businesses increasingly focusing on cost visibility and accountability as AI moves further into everyday operations.
For many businesses, the next step is therefore not another AI tool.
It is understanding the AI estate they already have.
Where is AI being used? Which workflows now depend on it? Which models and suppliers sit underneath those workflows? What company information is being processed? How does consumption grow as usage increases? And where is the measurable commercial return?
These are no longer purely technical questions.
They belong alongside finance, strategy, procurement, risk and operations.
Where this fits with the Invent Group approach
At Invent Group, we start with the problem, not the technology.
The token conversation reinforces why that matters.
If the starting objective is simply to increase AI adoption, more consumption can easily become a proxy for progress.
Start with a clearly defined business problem and the conversation changes. The task is to design the most effective way of achieving an outcome, then decide where AI should play a role within it.
That might require a powerful frontier model. It might require a smaller specialist model, better data, improved information retrieval or a redesigned workflow. Often it will involve a combination.
Our Consult and Create model is built around that principle.
Consult starts with the business problem, the workflow, the information available and the outcome that needs to improve.
Create turns the right opportunities into practical AI-powered products and systems that can deliver that outcome.
As AI becomes more deeply embedded into organisations, the quality of the architecture surrounding the model will matter increasingly. Cost control, model selection, data, governance and measurement become part of building something that works sustainably.
The objective is not simply more AI.
It is better capability.
Take home
- Token-driven AI costs are becoming a meaningful operating expense for businesses as adoption scales.
- Falling underlying inference costs do not automatically translate into lower AI bills because consumption is increasing much faster.
- Agentic systems can multiply token usage because a single human request can trigger many model interactions.
- AI consumption and AI productivity are not the same thing.
- Businesses need visibility over their models, providers, workflows, data and costs before AI becomes deeply embedded across operations.
- Value per token is a more useful measure of AI maturity than the volume of AI being consumed.
A practical first step
Map where AI is already being used across your organisation.
You do not need to begin with a complicated AI audit. Start with the important workflows and establish what technology is being used, what information it processes, how consumption is charged and what outcome the business expects in return.
Then look at scale.
What happens to the cost if usage doubles, increases tenfold or becomes agentic? Could another model perform the same job more efficiently? Is the same information being repeatedly processed? How difficult would it be to change supplier? Where is human judgement still necessary?
Most importantly, identify the value being created.
That moves the conversation from “How do we use more AI?” to “Where does AI create enough value to justify becoming part of how our business works?”
That is a much stronger foundation for AI innovation.