AI is an expensive game—and sometimes the most mature technology decision is not to use it
A recent Financial Times report described how large companies that rushed to introduce artificial intelligence tools into everyday operations are now beginning to restrict their use. They are imposing spending limits, directing employees toward cheaper models, and trying to connect token consumption to measurable business outcomes.
The phrase “we created a monster” makes a powerful headline. But it may describe the wrong problem.
The monster is not artificial intelligence.
The monster is AI adoption without economic architecture, clear objectives, usage controls, or the discipline to ask the most basic question first:
Do we actually need artificial intelligence to solve this problem?
AI has always been—and remains—an expensive game. The cost per token may fall for some models, and cheaper alternatives continue to emerge. But that does not mean we know when, or even whether, the total cost of operating AI at enterprise scale will become genuinely low and predictable.
Because the real cost is not just the model price.
It is integration.
It is data preparation and data quality.
It is security.
It is evaluation.
It is observability.
It is human oversight.
It is compliance.
It is error management.
It is the continuous work required when models, pricing structures, and capabilities change.
Most importantly, it is the cost of complexity introduced when AI is added to a system that did not genuinely need it.
From “Where can we add AI?” to “What are we trying to achieve?”
The current AI hype has created a dangerous reversal of priorities.
Many organizations begin with the question:
Where can we introduce artificial intelligence?
That is usually the wrong question.
The correct question is:
What are we trying to achieve, and what is the simplest, safest, and most economical solution capable of achieving it?
The difference may appear small, but it is fundamental.
The first question starts with a technology and searches for a problem that can justify it. The second starts with the desired business outcome and objectively evaluates which technology is required.
That should be the role of a serious digital solutions architect. The stakeholder should not be asked:
What do you want us to build?
The better question is:
What do you want to achieve with what we build?
The first question easily leads to features, platforms, and impressive demonstrations. The second leads to outcomes, capabilities, measurable value, and strategic direction.
Artificial intelligence is a tool. It is not an objective. It is not a strategy by itself. And it is certainly not proof that an organization has modernized.
A real case where the right solution was not AI
I was recently asked to examine the case of a company that wanted to introduce artificial intelligence to improve the workflow of a core business process.
The intention was completely understandable. Management wanted a smoother workflow, fewer delays, and greater efficiency. AI appeared to be the obvious modern solution.
But before recommending a technology, we studied the process.
We examined what actually happened at each stage, what outcome the organization wanted, and which functions genuinely needed to be automated. The conclusion was simple: everything the company needed could be achieved more effectively with technologies that already existed, at significantly lower cost and with less operational complexity.
It did not need an artificial intelligence model.
It needed better workflow design, appropriate automation, clear business rules, and proper integration between existing systems.
That was not a failure to adopt AI.
It was a successful architectural decision.
Sometimes the greatest value a technology adviser can provide is not recommending a new technology. It is preventing an organization from paying for technology it does not need.
Automation and AI are not the same thing
Much of the confusion comes from using “automation” and “artificial intelligence” as though they mean the same thing.
They do not.
Automation executes predefined rules. When the conditions are clear and the required outcome is predictable, a conventional system can perform the process quickly, reliably, and at low cost.
AI is particularly valuable when there is uncertainty, unstructured information, language, imagery, complex classification, changing context, or a need to generate and evaluate alternative responses.
When a task follows the pattern:
If A happens, do B,
it probably does not need artificial intelligence.
When the requirement is to move data from one system to another, notify a responsible person, apply a specific rule, route an approval according to known criteria, or produce a predefined report, mature and cost-effective technologies already exist.
Workflow engines, APIs, business rules, event-driven architectures, integrations, and conventional automation did not suddenly become obsolete when large language models appeared.
They are often still the best solution.
AI is justified when it adds a capability that conventional systems cannot provide easily or economically. It is not justified when it replaces a clear rule with a more expensive probabilistic process.
Predictability has business value
Technical capability is not the only consideration in business systems.
Predictability, reliability, cost, speed, security, and auditability matter as well.
A conventional rules-based system may not look as impressive in a presentation. But it can deliver the same result every time, at a known cost, with full traceability and transparent logic.
An artificial intelligence model may offer greater flexibility. It may also produce different answers to the same question, require continuous evaluation, create unexpected edge cases, and introduce operating costs that grow with usage.
That does not mean AI should not be used. It means organizations need to understand what they are buying.
They are not merely buying a new feature. They are adopting a different category of system: one that is more dynamic, more probabilistic, and usually more demanding from a governance perspective.
Predictability is not a limitation. It is a business advantage.
When two solutions achieve the same outcome, the simpler and more predictable one is usually the more mature choice.
The hidden cost of AI agents
Cost pressure becomes even greater with the rise of AI agents.
A simple chatbot waits for a user to submit a question. An agent can operate in the background, execute multiple steps, call tools, read large amounts of information, repeat actions, and activate other agents.
Each of those actions may carry a cost.
When an agent completes a task, the expense might not be limited to a single answer. It can include dozens or hundreds of model calls, searches, data retrievals, tool calls, and iterative evaluations.
This creates a new category of variable operating expenditure.
In traditional software, we can usually estimate with reasonable accuracy the infrastructure required for a given number of users or transactions. In agentic systems, cost may depend on how many steps an agent decides to execute, how much context it consumes, and how many times it repeats a process.
AI can automate work.
It can also automate resource consumption.
That is why agents deployed without budgets, limits, observability, and clear business KPIs can quickly become a source of uncontrolled spending.
Falling unit prices do not guarantee lower total costs
Models are likely to continue becoming more efficient. Smaller models, caching, batch processing, model routing, and other techniques can already reduce the cost of individual workloads.
That is positive.
But a lower price per token does not necessarily mean a lower total AI bill.
As AI becomes more accessible, more use cases appear. More employees gain access. More processes are automated. Larger contexts are processed. More agents operate continuously in the background.
Efficiency can reduce the cost of a single call while dramatically increasing the total volume of usage.
No one can predict with certainty what the true long-term cost of enterprise AI will be. The technology, prices, business models, and infrastructure requirements are changing too quickly.
An organization should therefore not base its strategy on the hope that “AI will soon be almost free.”
It should design around today’s real costs, measure business value, and preserve options.
AI FinOps cannot correct the wrong technology decision
Organizations are beginning to adopt AI FinOps practices: tracking consumption, defining budgets and alerts, selecting cheaper models, using caching and model routing, and matching each workload to the appropriate capability level.
All of this is necessary.
But it does not answer the first and most important question:
Should this workload have used artificial intelligence in the first place?
A flawed architecture can be optimized. It will still be flawed.
The price per token can be reduced. But if the process never needed tokens, the expense remains unnecessary.
True financial discipline starts before model selection. It starts with selecting the correct category of technology.
A simple decision sequence
Before adding AI to a business process, several basic steps should come first.
Begin by defining the desired outcome. Not the technology, not the feature, and not the demonstration.
Then study the existing process. Which steps are genuinely necessary? Which exist only because they have always existed? Where is the real delay or source of friction?
Next, determine whether the problem can be solved more simply: through better process design, system integration, clear rules, conventional automation, or functionality that is already available.
Only when a problem remains that genuinely requires the interpretation of uncertainty, unstructured data, natural language, or complex judgment should AI become a serious candidate.
And even then, the goal should not be to select the most powerful model. It should be to select the smallest and most economical model that reliably achieves the required outcome.
That is mature architecture.
The ability to say no to AI
Today, it is easy to approve a project because it includes artificial intelligence. AI sounds modern, attractive, and powerful in corporate communication.
It is harder to explain that the best solution does not require it.
But that is precisely the difference between technology consulting and technology sales.
A solutions architect is not there to validate the client’s initial assumption. The architect is there to test it.
The role is not to use every available tool. It is to choose the right one.
The goal is not to construct the most impressive system. It is to create the simplest solution that reliably achieves the desired outcome.
Sometimes that means using a powerful artificial intelligence model.
Sometimes it means using a smaller, specialized model.
And sometimes it means using no artificial intelligence at all.
AI where justified.
Automation where sufficient.
Simplicity wherever possible.
Artificial intelligence may be the most powerful technological capability of our time. That is precisely why it should be used with judgment rather than obligation.
The maturity of an organization is not measured by how much AI it uses.
It is measured by whether it knows when AI creates real value—and when it does not.
