In partnership with Elastic
On a quiet afternoon in a regional tech office, a finance team stares at an alarming cloud bill that has quadrupled in a single quarter, all thanks to a swarm of AI agents dutifully running their routine checks.
Across Southeast Asia, this scene has played out with sobering frequency. Tokenmaxxing, the habit of feeding oversized context windows, sprawling instruction sets, and heavy reasoning tasks into top-tier artificial intelligence models, has shifted from a symbol of AI ambition into a major boardroom concern.
High-growth regional firms have blown through annual AI budgets in a matter of months, while technical teams across major Asean markets have quietly dismantled internal token leaderboards to curb runaway infrastructure spend.
Despite these cautionary stories, enterprises across Southeast Asia are not backing off from their digital transformation goals.
With the regional digital economy expanding rapidly and private investments pouring into local AI tech stacks, business leaders in markets from Indonesia and Malaysia to Thailand and Vietnam are recalibrating how their AI systems are constructed.
A third of Singapore respondents have already moved 40 per cent or more of their AI pilots into production, and almost three-quarters plan to roll out agentic AI across multiple parts of the business within two years, according to consulting firm Deloitte.
The current transition is fundamentally about solving systemic architectural flaws rather than retreating from technology adoption, said Nicholas Lim, director of solutions architecture for Asean at Elastic.
An AI agent engineered to automatically load its full instruction set, complete toolkit, and entire conversation history into every interaction is inherently expensive to run, he explained.
In multi-turn sessions, repeatedly re-transmitting unneeded context creates financial deadweight that scales rapidly across high-volume enterprise operations, he added.
This realisation has driven a structural pivot across Asean tech teams towards context engineering over routine prompt engineering.
Rather than dumping broad organisational knowledge into a prompt window, system architects are building pipelines that deliver only the exact context required for a specific operational task.
Dynamic tool allocation is central to this shift. Platforms such as Elastic Agent Builder enable AI agents to selectively pull in domain-specific tools and instructions only when a user prompt explicitly requires them.
Combining dynamic loading with context summarisation, which condenses previous conversation turns instead of resending raw chat histories, lowers token consumption with minimal impact on output quality.
For organisations seeking maximum return on investment, managing AI usage requires a multi-layered approach to workload management, starting with model routing.
Automatically directing routine tasks like text classification, intent detection, and document summarisation to smaller, cost-effective models allows enterprises to reserve expensive, high-tier models strictly for complex, multi-step reasoning.
Operational data shows that companies deploying a tiered model configuration pay a fraction per million tokens compared to those defaulting all enterprise traffic to premium flagship models.
However, model choice alone hits a limit if the underlying retrieval pipeline remains unoptimised, said Lim, who added that precision retrieval before generation ensures models are not forced to process thousands of loosely relevant tokens.
Advanced compression techniques, such as Better Binary Quantisation, can reduce vector retrieval storage footprints by up to 32 times without sacrificing search accuracy, he noted.
When paired with hybrid search and re-ranking, far fewer tokens hit the model in the first place, yielding persistent savings regardless of which foundation model is plugged into the system, he added.
Another major source of budget leakage across regional deployments comes from wrapping basic operational tasks inside generative reasoning engines.
Routine processes like status checks, data enrichment, system logging, and ticketing dispatches do not require machine judgment.
By offloading these predictable steps to deterministic, scripted automation through tools such as Elastic Workflows, enterprises restrict expensive model calls purely to decisions that require genuine cognitive evaluation.
Consolidating these operational layers onto a single unified platform also prevents hidden technical tax, according to Elastic.
Stitching together separate vector databases, external retrieval pipelines, and standalone orchestration tools creates boundary friction where latency and token leaks quietly accumulate.
Treating search retrieval, tool access, and process execution as an integrated system eliminates these inefficiencies.
For enterprise leaders dealing with recent invoice surprises, the most damaging reaction is to abruptly choke off user access or suspend internal projects, said Lim.
Halting access merely delays organisational maturity while leaving fundamental architectural flaws unresolved, he noted.
A more effective roadmap begins with establishing granular visibility over workload spending.
Many organisations across the region still lack the tracking precision needed to identify which specific department, agent, or workflow accounts for their monthly invoice. Logging every tool call, retrieval query, and reasoning step creates the audit trail necessary for financial governance.
Next, technical teams should prioritise context cleanup before embarking on long re-platforming initiatives. Trimming search parameters, loading tools dynamically, and summarising long chat histories can be executed quickly to deliver immediate relief.
Finally, auditing active AI agents to separate routine automated steps from actual reasoning tasks eliminates unnecessary token usage at the root.
For Southeast Asian enterprises, aligning cost efficiency with strict data sovereignty mandates offers a distinct advantage, said Lim.
Changing regulatory framework requirements across the region are driving a shift towards hybrid AI architectures, keeping data processing closer to local infrastructure rather than sending sensitive payload data to offshore cloud endpoints, he noted.
Running platforms like Elastic Agent Builder or Elastic Workflows on-premises or within localised air-gapped environments addresses compliance and residency demands whilst eliminating external bandwidth and data transfer overheads, he added.
By treating data sovereignty and cost optimisation as a unified engineering goal, Southeast Asian enterprises can maintain aggressive digital momentum while keeping quarterly expenditures fully predictable, he advised.
Find out how Elastic’s AI tools can help transform your business with AI here
