For the last two years, organisations have been driving to adopt AI.
The initial conversation was relatively simple; Which model should we use? How many tokens does it consume? What does each API call cost? But as AI moves from experimentation into core business operations, those questions are no longer enough.
The real challenge is not simply controlling the cost of AI, it is understanding the total economic equation behind an AI outcome. Because an AI project can have incredibly cheap tokens and still be a terrible investment. Equally, an AI workflow may consume significant server and model costs while delivering value to the organisation.
The question leaders should increasingly be asking is:
What does this AI capability cost us to operate, and what measurable outcome does it produce compared with the people, software and processes it replaces, improves or enables?
That is where the next phase of AI governance needs to go.
The token trap
Token consumption has become one of the easiest AI metrics to measure. Every interaction with a large language model has a cost. Input tokens, output tokens, API calls and increasingly inference or reasoning costs can all be tracked., and because they are measurable, organisations naturally focus on reducing them.
Teams optimise prompts, they select cheaper models, they introduce smaller models for simpler tasks, they set usage limits. All of this is sensible, but token optimisation alone can create the wrong behaviour.
Imagine two AI systems, the first costs $0.03 per transaction, the second costs $0.30 per transaction. On a token-cost dashboard, the first system looks ten times better, but what if the $0.03 system produces an outcome that still requires a person to spend five minutes correcting it?
And what if the $0.30 system completes the task accurately, removes the manual intervention and reduces a process that previously took twenty minutes? Which one is actually more expensive? The answer becomes obvious when we stop measuring AI as a technology cost and start measuring it as an operating model.
Cheap AI that produces little value is expensive.
Expensive AI that produces significant business value can be incredibly cheap.
AI has changed the unit economics of work
Traditional software economics were relatively straightforward. A business purchased software licences and employed people to operate processes.
The costs were generally visible:
- Employee salaries
- Software subscriptions
- Infrastructure
- Outsourcing
- Transaction processing
- Operational overhead
AI introduces a new variable into that equation, we now have the ability to dynamically substitute, increase, or accelerate human work with intelligence that has a variable usage cost.
That changes how organisations need to think about ROI; the economic comparison should no longer simply be;
Cost of AI platform vs cost of AI platform.
It should be;
Total cost of delivering a business outcome before AI versus the total cost of delivering that same outcome with AI.
For example:
| Traditional Process | AI-Enabled Process |
|---|---|
| Employee time | AI model/token consumption |
| Multiple software tools | AI orchestration platform |
| Manual data movement | Automated agent actions |
| Human decision making | AI-assisted or autonomous decisions |
| Outsourced processing | AI-enabled workflow |
| Slow processing times | Near real-time execution |
| Limited scalability | Variable AI consumption |
The comparison is not about whether AI costs money, of course it does. The comparison is whether the organisation can achieve a better outcome at a better overall economic cost.
The three costs every AI project should measure
To understand the true ROI of AI, organisations need to look beyond tokens, there are at least three cost categories.
1. AI Consumption Cost
This includes:
- Tokens
- Model inference
- API usage
- Agent execution
- Usage
- Vector databases
- AI infrastructure
This is the cost most organisations are beginning to monitor, but it is only one part of the equation.
2. Software and Technology Cost
AI rarely operates in isolation. An AI workflow may require,
- SaaS platforms
- Integration tools
- Automation platforms
- Data infrastructure
- Security systems
- AI orchestration layers
- Governance platforms
One of the risks emerging from the AI boom is that organisations are adding AI tools on top of existing software stacks without removing anything. Instead of reducing technology costs, AI can accidentally create another layer of SaaS sprawl. A business might have ten employees using five different AI tools, each connected to different data sources and software platforms.
- The token cost may look insignificant.
- The overall technology cost may not.
3. Human Cost
This is perhaps the most important, and most overlooked, part of AI ROI. If an AI system generates an answer but requires a person to;
- Review it
- Correct it
- Re-enter it into another system
- Approve every decision
- Handle exceptions
- Reconcile errors
Then the organisation has not necessarily automated the work, it may simply have created another step in the process. The real opportunity comes when AI changes the economics of the workflow, that could mean reducing manual work. But it could also mean increasing capacity, improving accuracy, accelerating decisions or allowing employees to focus on higher-value work.
The objective should not be:
How many people can AI replace?
The better question is:
How can AI improve the economics and outcomes of the work our organisation needs to perform?
From token ROI to outcome ROI
This is where AI measurement needs to evolve, every significant AI capability should have an identifiable outcome.
For example:
Customer Service Agent
Instead of measuring:
- Tokens consumed
- Conversations handled
- Cost per API call
Also measure:
- Resolution rate
- Escalation rate
- Cost per resolved case
- Customer satisfaction
- Reduction in human handling time
AI Software Development
Instead of measuring:
- AI subscriptions
- Model cost
- Number of prompts
Measure:
- Development cycle time
- Features delivered
- Defects introduced
- Time to production
- Cost per software outcome
AI Finance Processing
Instead of measuring:
- Documents processed
- Token consumption
Measure:
- Cost per invoice processed
- Processing time, Exception rate
- Human intervention, Accuracy
AI Commerce Operations
Instead of measuring:
- Agent calls
- Model usage
Measure:
- Orders processed
- Manual interventions avoided
- Fulfilment speed, Inventory accuracy
- Revenue recovered
- Cost per operational outcome
This is the shift from AI activity metrics to AI business metrics, because businesses don't generate value from tokens, they generate value from outcomes.
The emerging problem: AI costs without accountability
As organisations deploy more AI agents, another challenge is emerging, who actually owns the economics? The CIO may see infrastructure costs, the CFO may see software subscriptions, Business leaders may see headcount costs, individual teams may see AI productivity gains, but nobody sees the complete picture.
An AI agent could be:
- Consuming thousands of dollars in model usage
- Using three SaaS platforms
- Requiring human review
- Creating downstream exceptions
- Delivering genuine revenue
Or it could be doing all of the above while producing almost no measurable value. Without visibility across the complete lifecycle, organisations cannot make informed decisions about where to scale AI. This is why AI governance cannot simply become a risk and compliance exercise.
It also needs to become an economic management discipline.
Every AI agent needs an economic profile
As AI agents become part of the workforce, organisations will increasingly need to understand them in a similar way to any other operational capability.
For every significant AI system or agent, leaders should be able to see:
- What does it do?
- What business outcome is it responsible for?
- What models does it use?
- What is its token and infrastructure cost?
- What software does it depend upon?
- How much human intervention does it require?
- What risks does it introduce?
- What measurable value does it create?
- And critically: Is the value increasing faster than the cost?
This creates what could be described as an AI economic profile, not simply a technology inventory or a risk register, but a live view of the economics of AI operating inside the business.
AI governance needs to include financial governance
This is an area where we believe the conversation around AI governance needs to mature.
Today, governance discussions often focus on:
- Security
- Privacy
- Model approval
- Responsible AI
- Regulatory compliance
- Risk
These are all essential, but there is another governance question:
Are we spending money on AI intelligently?
As organisations move from a handful of AI experiments to potentially hundreds of agents operating across their business, uncontrolled consumption can become a significant problem. AI agents can scale incredibly quickly, unlike an employee, an agent does not get tired.
It can run continuously, it can trigger other agents, it can call multiple models, it can process thousands of transactions, and this creates enormous opportunity. It also creates the possibility of rapidly increasing costs without corresponding business value.
The future AI operating model will therefore need controls around:
- AI budgets
- Consumption thresholds
- Cost per outcome
- Agent performance
- Value creation
- Human intervention
- Software dependency
- Return on investment
The goal should not be to restrict AI, the goal should be to understand where AI is creating value and give organisations the confidence to invest more in the areas that are working.
The future metric: Cost Per Outcome
We believe one of the most important AI metrics will eventually become Cost Per Outcome
Rather than asking:
How much did we spend on tokens this month?
Organisations will ask:
How much did it cost us to resolve a customer issue? How much did it cost us to process an order? How much did it cost us to develop and deploy a feature? How much did it cost us to process a financial transaction?
And increasingly:
How does that compare with the cost before AI?
This creates a much more meaningful conversation, because sometimes the right decision will be to use the most advanced and expensive model available. If it delivers a dramatically better outcome, the additional AI cost may be insignificant compared with the value created. Other times, a smaller and cheaper model may be perfectly adequate, the answer should not be driven by token price alone.
It should be driven by economics.
AI should be managed like a workforce and a technology estate
The organisations succeeding with AI will not simply be those using the cheapest models. They will be the organisations that understand how AI is changing their entire operating model.
AI is becoming:
- A technology cost
- A software capability
- A digital workforce
- An operational dependency
- A variable consumption model
That means it needs a new level of management.
What AI is actually delivering value?
That question was part of the thinking behind TraphicLights.ai. We believe organisations need visibility not just into what AI they have, but increasingly into how AI is performing as part of the business, that includes risk and governance.
But it should also include economics.
What AI is actually delivering value?
The first phase of AI adoption was experimentation, the second phase is deployment, the next phase will be optimisation, and optimisation requires a more sophisticated understanding of ROI.
Organisations will need to connect:
AI Cost → Software Cost → Human Cost → Business Outcome
Only when those four elements are visible together can leaders understand the true economics of AI. The future will not belong to the organisations that simply consume the fewest tokens, it will belong to the organisations that can answer a much more important question:
For every dollar we spend on AI, what measurable outcome are we creating, and are we improving the economics of the business?
That is the shift from measuring AI consumption, to managing AI value, and it may become one of the most important disciplines of the AI native organisation.