The problem with ‘tokenmaxxing’: More AI usage doesn’t add up to more value
As token consumption rises, Elastic’s Ravi Rajendran says companies need to focus on sharper context, trusted data and stronger governance to ensure AI delivers measurable returns
Add BT as a preferred source
AS BUSINESSES scale their use of AI, rising token consumption is prompting a tougher question: Are the gains in productivity and efficiency keeping pace with the cost?
That concern is especially acute where companies treat token use as a proxy for AI adoption – a practice sometimes called “tokenmaxxing”. While it can nudge employees to experiment with AI, it can also reward activity over outcomes.
The stakes are set to rise as AI agents become more widely used. Goldman Sachs Research projects global token consumption will increase 24-fold between 2026 and 2030 as agentic AI adoption accelerates.
As AI spending shifts from innovation budgets to core operating expenditure, the way businesses design their AI systems today will have a growing impact on both costs and returns.
A more focused approach, says Ravi Rajendran, Elastic’s area vice-president for ASEAN, Hong Kong, Taiwan and Korea, is to define the business outcome first, then work backwards to identify where AI can improve the speed, accuracy or quality of a process, rather than encouraging employees simply to maximise AI usage.
“Local IT leaders recognise that ‘tokenmaxxing’ often fails to deliver real value while unnecessarily driving up expenses – a problem that will only compound as enterprise token volumes grow,” he notes.
“Rather than deploying AI for its own sake, Singaporean firms are prioritising high-quality applications that produce tangible business results.”
Better context can reduce AI costs
Building effective AI applications starts with giving models access to the right enterprise data. Moving beyond tokenmaxxing means thinking less about how often AI is used and more about how it reaches the right answer.
One way, says Rajendran, is to move AI inference closer to where enterprise data already resides instead of routing sensitive records, operational telemetry and compliance signals through external model pipelines.
“It is far more efficient, and far less costly, to ground an AI system in the specific data it needs than to feed it everything and hope it finds the signal,” he says.
Done well, every token is being used more purposefully because the model reasons over trusted enterprise data within the organisation’s own governance perimeter.
For Singapore-regulated organisations, this is not just about efficiency. Keeping AI processing within a defined data environment can also help meet technology risk management and data governance requirements.
This is the principle behind context engineering: supplying AI with the right information, from the right sources and at the right point in a task, so it can reason from relevant data instead of sifting through unnecessary noise.
The ability to retrieve the most relevant information efficiently is equally important. Technologies such as Elastic’s Better Binary Quantization can reduce storage requirements for retrieval by up to 32 times while maintaining accuracy, helping businesses manage costs as their data volumes grow.
Managing context, however, is only half the equation. As organisations move from AI assistants that generate answers to AI agents that execute tasks across workflows, they also need to define what each agent is allowed to do.
One approach is to give each AI agent access only to the capabilities it needs for a specific task, rather than every available tool. Coordinating these specialised agents through a central orchestration layer also gives organisations greater visibility over how work is delegated, decisions are escalated and actions remain within defined boundaries.
“Well-grounded agents draw on specific skills and workflows only when a task requires them, reducing both token overhead and the risk of an agent taking unintended action by reasoning across capabilities it should not have engaged,” Rajendran says.
Governance becomes central to AI strategy
As AI agents begin making decisions and carrying out tasks across workflows, governance moves from policy to operational necessity. Organisations increasingly need to know what data an agent consulted, which tools it used and why it reached a particular decision.
Many early enterprise AI programmes focused on driving adoption before governance frameworks were fully established. Today, boards and regulators are increasingly asking whether AI outputs can be verified, explained and overridden when necessary.
Without that visibility, organisations may struggle to explain AI-generated decisions to regulators, customers or even their own boards.
Elastic’s agent observability capability helps address this by providing a real-time view of agent decisions, tool calls and data access events, allowing organisations to monitor AI activity without requiring separate monitoring infrastructure or manual audit reconstruction.
“This is not a sign that the technology has failed. It is a sign that the governance infrastructure needed to run it responsibly was not yet in place when the deployment began,” says Rajendran.
As AI agents take on more responsibility, people remain essential to define guardrails, oversee decisions and ensure AI continues serving business objectives rather than simply automating tasks.
The next competitive edge is not bigger AI models
Moving beyond tokenmaxxing requires businesses to rethink what creates competitive advantage in AI.
The companies that pull ahead may not be those that consume the most tokens or deploy the largest models, but those that can give AI the most relevant, trusted and well-governed data to reason from.
Rajendran says the next phase of enterprise AI will depend less on model selection and more on the “intelligence layer” around it – the retrieval, ranking, memory and workflow infrastructure that determines what a model can actually know and do.
A well-grounded smaller model, he notes, can outperform a larger model working from fragmented or imprecise context.
“The next wave of AI value will not be won by those with the biggest models. It will be won by those with the most precisely organised, most trusted data, and the infrastructure to put it in front of the right reasoning layer at exactly the right moment,” he adds.
Share with us your feedback on BT's products and services
