For the last few years, many organisations have been experimenting with artificial intelligence in unusually forgiving conditions.
Powerful tools were cheap. Free tiers were generous. Subscriptions looked simple. Internal pilots could be launched quickly. Teams could test copilots, chatbots, AI agents, document tools, sales assistants, coding assistants and workflow automation without thinking too deeply about the full economic model underneath them.
That period is now changing.
A recent article in The Verge by Hayden Field, “You’re about to feel the AI money squeeze”, captured the shift clearly. The article describes an AI market moving from subsidised adoption into commercial reality, with rate limits, feature restrictions, higher prices, token-based enterprise plans, advertising experiments and more pressure on users to pay for the compute they consume.
This is not simply a technology pricing story. It is a warning sign for every business currently experimenting with AI.
The era of cheap, forgiving AI experimentation is giving way to a more financially disciplined phase. Businesses that do not understand the value of their AI use cases may soon find themselves paying for pilots, subscriptions and agents without being able to explain whether those initiatives are delivering a measurable return.
AI Is Not a Normal Software Cost
For many business leaders, AI is still being mentally categorised as software. That is understandable. Most AI tools are bought through familiar commercial structures: monthly subscriptions, enterprise plans, user licences, platform fees or API access.
But generative AI is not economically identical to traditional SaaS.
With conventional software, the marginal cost of one more user action is often relatively low. With generative AI, especially reasoning models and agentic workflows, usage can directly drive material compute cost. Every prompt, uploaded document, generated response, reasoning step, tool call, retrieval event and agent loop consumes tokens.
To the business user, an AI task may feel like one simple request.
To the economics, it may be a long chain of computational activity.
A one-line prompt to an AI agent can trigger thousands of tokens of hidden reasoning. A coding assistant may inspect files, generate alternatives, review its own work, run checks and revise outputs. A document review workflow may retrieve context, summarise evidence, compare clauses, classify risks and draft recommendations. A sales research assistant may search, interpret, score, summarise and generate account plans.
What looks like a single productivity gain at the user interface level may be a much larger consumption event underneath.
That distinction matters because, as AI providers face the realities of infrastructure investment, energy costs, GPU demand, data centre build-outs, model training and ongoing research, more of that cost will inevitably be passed through the value chain.
For businesses, that means AI cost control can no longer be treated as an afterthought.
The First Wave Was About Possibility
The first wave of AI adoption was largely driven by possibility.
Boards wanted to know what their organisation was doing with AI. Innovation teams wanted to demonstrate progress. Employees discovered tools that helped them work faster. Vendors sold compelling stories about productivity, automation and transformation.
That was not necessarily wrong. In the early stage of any major technology shift, experimentation matters. Businesses need to test, learn and build internal literacy. They need to understand where the technology helps, where it fails and where it introduces risk.
But experimentation is not the same as value creation.
A pilot can look impressive when the underlying cost is low, subsidised or hidden inside a flat-rate bundle. A chatbot that saves a few minutes feels useful. A document assistant feels efficient. A sales research agent feels clever. A coding tool feels transformational.
The commercial test arrives later, when usage scales.
At that point, the relevant question changes. It is no longer simply:
“Can we use AI here?”
It becomes:
“Is this AI use case financially worth doing?”
That is a much harder question.
AI Activity Is Not the Same as AI Value
Many organisations now have AI activity. They have pilots, subscriptions, proofs of concept, employee experiments and vendor demonstrations. Some have internal AI champions. Some have working groups. Some have informal usage across multiple departments.
But AI activity is not the same as AI value.
AI value exists when a business process improves in a measurable way. That improvement might be financial, operational, commercial, risk-related or quality-related. It might show up as lower cost, faster cycle time, higher conversion, reduced error rates, improved compliance, better customer outcomes or increased capacity.
Without that measurable connection, an AI initiative risks becoming a cost centre with a good story attached.
This is where many businesses are vulnerable. They have pilots, but no value model. Usage, but no unit economics. Enthusiasm, but no financial governance. Subscriptions, but no clear owner for benefit realisation.
Then pricing changes. Rate limits tighten. Usage grows. Vendors introduce higher tiers. Agentic workflows consume more tokens than expected. The business starts paying more, but cannot clearly explain what it is getting in return.
That is the real danger.
Not that AI costs money. Of course it costs money.
The danger is paying for AI without knowing what value it is meant to create.
Start With the Business Outcome, Not the Tool
A mature AI value case starts with the business outcome, not the technology.
“We are using AI in customer service” is not a value case.
“We are reducing average handling time while maintaining customer satisfaction and first-contact resolution” is closer.
“We are using AI for legal review” is not a value case.
“We are reducing contract review cycle time while maintaining escalation quality and lowering external legal spend” is closer.
“We are using AI for sales” is not a value case.
“We are increasing qualified opportunity conversion by improving account research quality and reducing preparation time per prospect” is closer.
The distinction is important. AI does not create business value simply because a model generates an output. It creates value when that output changes a process, decision or outcome in a way the organisation can measure.
That means every serious AI use case needs a defined value hypothesis.
What is the process being improved? What is the current baseline? What metric should move? What is the expected benefit? What level of confidence is required before scaling? What evidence will prove the change is working?
Without those answers, the organisation is not managing AI investment. It is funding experimentation and hoping value emerges.
The Baseline Problem
One of the biggest weaknesses in AI adoption is the absence of baselines.
If a business does not know how much a process costs today, it cannot prove AI has reduced cost. If it does not know how long a process currently takes, it cannot prove cycle time has improved. If it does not track error rates, escalation rates, conversion rates or customer outcomes, it cannot prove quality has improved.
This turns AI benefit into anecdote.
Anecdotes can be useful during discovery. They are not enough for investment decisions.
Before scaling an AI use case, organisations should understand the current state of the process. That does not always require a perfect measurement model, but it does require enough evidence to compare before and after.
For example:
How many hours are spent on the task today?
What roles are involved?
What is the cost of that labour?
How long does the workflow take from start to finish?
How often does work need to be corrected or escalated?
What does poor performance currently cost the business?
What quality threshold must be maintained?
What risk must not be increased?
Only then can the business evaluate whether AI improves the economics.
The Full Cost of AI Is Bigger Than the Invoice
Another common mistake is treating the AI vendor invoice as the total cost.
The true cost of an AI workflow may include subscription fees, API usage, token consumption, integration work, internal support, monitoring, governance, data protection, security review, human oversight, training, change management, vendor management and ongoing optimisation.
There may also be hidden costs in workflow redesign. If AI produces outputs that humans must heavily review, correct or rework, the apparent productivity gain may be overstated. If employees use AI inconsistently, quality may vary. If teams rely on high-cost models for low-value tasks, the economics may deteriorate quickly.
This is why AI investment needs unit economics.
What does it cost to complete one successful AI-assisted task? What does it cost per customer interaction, per document reviewed, per sales lead qualified, per report generated or per support ticket resolved? What happens when usage increases by 10 times, 100 times or 1,000 times?
The economics of a small pilot may not survive production scale.
Model Choice Is Now a Financial Decision
In the next phase of AI adoption, model selection will become a financial governance issue.
Not every task needs the most powerful model available. Some use cases require advanced reasoning. Some require speed. Some require consistency. Some require privacy. Some require retrieval from trusted business data. Some can be handled by smaller, cheaper models. Some are better suited to deterministic automation. Some should not use generative AI at all.
The question should no longer be:
“Which AI model is best?”
The better question is:
“Which model is economically appropriate for this task, risk profile and value target?”
High-value, high-risk workflows may justify high-cost models and deeper human oversight. Low-risk, repetitive workflows may not. Some AI agents may be worth the token consumption because they replace expensive manual work or create new revenue capacity. Others may simply generate expensive background activity without enough measurable benefit.
The businesses that understand this will make better investment decisions. They will know when to pay for premium AI capability and when not to.
AI Governance Must Include Value Governance
AI governance is often discussed through the lens of risk, compliance, data protection, security, regulation and ethics. Those issues remain critical.
But governance also needs to include value.
A business should be able to explain why an AI use case exists, what outcome it is expected to improve, what it costs, what benefit it creates, what evidence supports that benefit, who owns it and when the organisation should scale, stop or redesign it.
That is not bureaucracy. It is investment discipline.
AI governance should help organisations make better decisions about where to deploy AI, which use cases deserve funding, which risks are acceptable, which controls are required and which initiatives should not proceed.
Without value governance, businesses risk building AI estates that are technically active but commercially weak.
The CFO Lens Is Becoming Essential
The next stage of AI adoption will not be won by the organisations with the most pilots. It will be won by the organisations that can prove which pilots deserve to survive.
That requires a CFO lens.
A CFO lens asks whether the business case is real. It examines cost, value, risk, evidence and return. It challenges vague productivity claims. It separates genuine operating leverage from activity. It asks whether the expected benefit can be measured, whether the cost is understood and whether the economics remain attractive at scale.
This is particularly important for SMEs. Smaller and mid-sized businesses cannot afford uncontrolled AI sprawl. They need to know which AI investments will improve performance, which will create risk and which will quietly consume budget without delivering material value.
The good news is that this discipline does not prevent innovation. It strengthens it.
When a business understands AI value properly, it can invest with more confidence. It can scale the right use cases faster. It can stop weak initiatives earlier. It can negotiate better with vendors. It can choose models more intelligently. It can explain AI investment to directors, finance teams, shareholders and operational leaders in language they understand.
From AI Experimentation to AI Economics
The free ride was useful. It lowered the barrier to adoption. It helped businesses experiment. It created momentum. It allowed organisations to build familiarity with tools that will almost certainly become part of the future operating model.
But it was never the final state.
The next phase of AI will be less about access and more about economics. Less about what can be tried and more about what should be scaled. Less about visible demos and more about measurable value.
That does not mean businesses should slow down. It means they need to become more disciplined.
Every AI initiative should be assessed against a clear business outcome, a measurable baseline, a total cost view, a risk profile, an ownership model and a value case that can withstand scrutiny.
The businesses that do this well will not necessarily spend less on AI. In many cases, they may spend more. But they will spend deliberately.
They will know when expensive AI is worth it. They will know when cheaper models are good enough. They will know when an agent is creating genuine leverage and when it is simply burning tokens. They will know which pilots deserve production investment.
That is the commercial maturity businesses now need.
At Strategic AI Guidance Ltd, we are building this discipline directly into how organisations assess AI investment. Our CFO Evaluator tool is designed to understand and log AI value against cost, helping businesses determine whether an AI initiative has a genuine economic business case before it becomes another expensive pilot with no measurable success criteria.
For organisations experimenting with AI but struggling to explain the financial case for scaling it, this is the moment to bring value, cost, evidence and return into the centre of the conversation.
Citation: “You’re about to feel the AI money squeeze”, Hayden Field, The Verge. https://apple.news/Avnt3LORISWWHRrYwMmn5Cw