For most of the past two years, the AI conversation in Singapore boardrooms has followed a familiar arc. First came experimentation: pilots, proofs of concept and the question “Can AI do this?” Then came adoption: licences, training programmes and dashboards tracking whether people were actually using the tools. Both phases mattered. Neither is where the hard work now sits.
The organisations I speak with are entering a third phase, where AI moves out of the chat window and into workflows, and agents begin to act on the business’s behalf. Here the executive questions get harder. Is it economically worthwhile? Should it be doing this, with this data and this level of autonomy? Can we see, constrain, audit and stop it? Those three questions map to three disciplines: tokenomics, governance and control.
Singapore is better prepared than most for the second and third. IMDA’s Model AI Governance Framework for Agentic AI, launched in January and updated in May, gives boards a practical reference: bound the risks upfront, make humans meaningfully accountable, implement technical controls, and equip end users to use agents well. The economic question, by contrast, is still too often answered by instinct.
A token is a meter, not a business outcome
A token is the unit AI platforms use to meter consumption, roughly a fragment of a word. AI tokenomics is the discipline of understanding and optimising the economics of AI consumption: from models and tokens, through architecture and capacity, to the cost and value of the outcome the business actually cares about.
That distinction matters because the token price is only the tip of the iceberg. Most AI business cases budget for the advertised price per million tokens. Below the waterline sit the costs that decide the bill: model choice and growing context windows, retrieval of enterprise data, agent loops, tool calls and retries, the choice between pay-as-you-go and provisioned capacity, evaluation and observability, integration, and the human time spent reworking outputs that weren’t good enough. Agents magnify all of this. They plan, retrieve, call tools, validate and re-plan before they answer, so consumption becomes harder to predict.
“Price per token is one line item. Architecture, autonomy and operations decide the bill.”
Measure up the ladder
Most organisations already measure the bottom of the ladder well: tokens, API calls and GPU hours, then the cost per token, inference or user. These are classic FinOps measures, and finance teams in Singapore have grown practised at applying them to cloud. AI economics lives in the rungs above. Quality: did it work, and was the output accepted? Outcome: what did it cost per case resolved, proposal drafted or incident avoided? Value: what revenue, avoided cost, cycle time or released capacity did it create? Few leadership teams could answer that top rung for their AI spend today. But it is the gap to close before AI spend scales further.
One equation, three factors
The way I frame it with leadership teams is a simple equation: Frontier AI = Business Value × Responsible Scale × Operational Control. It multiplies on purpose. If any factor is zero, the result is zero. Great economics with no control is a liability. Perfect governance with no measurable value is a cost centre. And an agent that is safe but economically irrational should not scale.
This is why tokenomics cannot be a finance side project. It belongs inside the same operating model as governance and control, with finance, security and IT looking at the same agents through different lenses. A runaway agent can cause a security incident. It can just as easily cause a cost incident.
What this means for Singapore
Every organisation in Singapore can reach the same AI models. The differentiator is the ability to scale AI economically and responsibly without losing control, which is exactly the behaviour our national frameworks are designed to encourage. Support such as the Enterprise Compute Initiative can help eligible organisations fund that journey. What remains is for each enterprise to bring the same rigour to AI economics that it is already learning to bring to AI governance.
Scale what creates value. Govern what creates risk. Control what takes action.
Author
