AI costs extend far beyond tokens. As organisations scale AI, architecture, orchestration, human oversight and operational design increasingly determine the real economics. The question for leaders is no longer simply what AI costs, but what business value it creates.
As organisations move from AI experimentation into production, a new challenge is emerging – AI is proving harder to cost, forecast and control than many business cases assumed.
The economics of enterprise AI extend far beyond the model itself. This becomes particularly important as organisations adopt agentic AI, where systems reason, retrieve information, use tools, make decisions and execute multiple steps before producing an outcome.
The question therefore should not be: “What does the AI model cost?”
It should be: “What does it cost to produce a successful business outcome using AI?”
The four layers of AI cost
AI expenditure can broadly be considered across four layers:
Some of the fastest-growing costs may therefore sit outside the underlying model—in orchestration, security, monitoring and the repeated processing of information through the AI architecture.
This is why token pricing alone provides an incomplete view of AI economics.
Agentic AI changes the equation
Agentic AI makes cost management more complex because the cost of an agent is not the cost of one response. It is the cost of the complete decision and execution loop.
A single transaction may require an agent to interpret a request, develop a plan, retrieve documents, call tools, evaluate results, retry failed steps, obtain verification and produce an outcome. Every stage may consume tokens, computing capacity and external services.
More importantly, this path is not always predictable. Two apparently identical transactions could retrieve different amounts of information, require different numbers of reasoning steps, make different tool calls or escalate to different models.
AI cost should therefore increasingly be understood as a range or distribution rather than a single fixed number.
This has an important implication: AI consumption can rise even while model prices fall.
Cheaper models do not automatically produce cheaper AI. Architecture, workflow design, context management and agent behaviour all influence the final unit cost.
Start with the economics of the process
One of the biggest weaknesses in many AI business cases occurs before the AI solution is even designed.
Organisations estimate how much AI might save without properly understanding what the existing process costs.
A meaningful baseline should consider transaction volumes, handling time, labour, rework, outsourcing, technology costs, losses from poor decisions, revenue leakage and operational or compliance risk.
Only then can alternative operating models be properly compared. And the answer does not always have to be AI. The appropriate operating model could be:
* Human-led;
* Rules-based automation;
* AI-assisted human; or
* AI-led with human exception management.
The objective should be to select the lowest-cost reliable operating model capable of achieving the required business outcome, rather than automatically choosing the most sophisticated AI solution.
Use the minimum sufficient intelligence
This leads to one of the most important principles of AI economics:
Use the minimum sufficient intelligence required to achieve the outcome.
Simple extraction may need only rules, OCR or a small model. Classification may suit a specialised model. Standard summarisation may use a lower-cost general model. Complex analysis may justify a more advanced model, while frontier-level reasoning should be reserved for tasks that genuinely require it.
This is where model routing becomes important. Rather than sending every task to the most capable—and often most expensive—model, organisations can route work according to complexity, risk and confidence.
At scale, these architecture decisions can have a substantial impact on AI economics.
A Proof of Value must prove the economics
The same principle should change the way organisations approach AI pilots.
A Proof of Value should not simply answer: “Does the technology work?”
It should also answer: “Do the economics work?”
That means measuring business performance, quality, operational behaviour and economics simultaneously.
Metrics such as tokens per transaction, model calls, retries, human review, rework and total cost per successful outcome should sit alongside traditional measures such as accuracy, productivity and turnaround time.
A pilot that demonstrates impressive technology but cannot demonstrate viable unit economics has not yet proven that it can scale.
Put economic guardrails in place before scale
As AI systems become increasingly autonomous, they should operate within clearly defined economic boundaries.
These could include maximum cost per transaction, token limits, maximum reasoning steps, tool calls, retries, confidence thresholds and escalation rules.
The principle is simple: No autonomous system should operate without a defined mandate, budget and stopping rule.
This moves AI cost management beyond retrospective reporting. Economics become part of the design of the AI system itself.
From AI cost to Return on Intelligence
Ultimately, the management conversation needs to move away from the price of tokens and towards unit economics and value.
* What does it cost to resolve a claim?
* What does it cost to assess a loan?
* What does it cost to onboard a customer?
* What does it cost to identify a fraudulent transaction?
And, most importantly: How much incremental business value is created for every unit of AI operating cost?
This is a more useful way of thinking about Return on Intelligence.
That value may come from increased employee capacity, higher revenue, reduced fraud and losses, faster turnaround, improved customer retention, lower outsourcing costs or reduced risk.
The first phase of enterprise AI was dominated by the question: Can AI do this?
The next phase will increasingly be defined by another: Should AI do this—and at what cost?
The organisations that succeed will not necessarily be those deploying the most sophisticated AI. They will be those that become best at converting machine intelligence into measurable business value.
