Stay connected with KayaToday, follow us on Instagram and Facebook for the latest news and reviews delivered straight to you.
When Uber handed its engineers access to Claude Code in December 2025 and set up internal leaderboards rewarding heavy usage, the company got exactly what it incentivised: consumption. By April, the entire AI coding budget for 2026 had been exhausted. What made the episode genuinely alarming was not the bill itself but what Uber President and COO Andrew Macdonald admitted afterward: there was no demonstrated link between that surge in usage and any improvement in the products reaching riders and drivers.
That admission crystallised a problem now spreading across the enterprise technology world. The term that has attached itself to the phenomenon is “tokenmaxxing,” meaning runaway token consumption that produces no measurable return. Gartner projects spending on AI agent software will reach roughly $207 billion this year, up more than 139 percent from $86.4 billion in 2025. Yet the executives and engineers trying to manage that spend are discovering that token costs behave nothing like the software licensing fees that finance teams have spent decades learning to model. The same engineer, using the same tool, on the same day, can generate wildly different invoices depending on whether they are auto-completing lines of code or running parallel agents through a major database migration.
The Default Settings Nobody Questioned
Uber responded by capping spend at $1,500 per employee per agentic coding tool each month. Microsoft questioned the cost of Claude Code licenses before cancelling them across its Experiences and Devices division. Duolingo reversed a plan to tie AI usage to performance reviews after employees pushed back against being told to use the tools for their own sake, regardless of outcome.
Dmytro Palaniichuk, head of engineering at language-learning platform Promova, traces much of the problem to the default configurations that AI vendors ship. “It turned out that on team and enterprise plans, the premium models are pre-selected; they run at high reasoning effort by default, and on our plan, Opus sessions upgrade to a 1M-token context window. Of course, nobody changes these in their own settings,” he told VentureBeat. Within months, Palaniichuk was watching a premium frontier model handle tasks as trivial as checking email.
Noe Ramos, VP of AI operations at enterprise contract lifecycle management platform Agiloft, frames the same problem differently. “Teams aren’t burning spend because they love waste. They’re burning it because the default infrastructure pushes them toward it,” he said. “Most enterprises still hand model selection to whoever is prompting, which means a frontier model is handling tasks that a cheap open-weight model could do just as well. That’s not a people problem. It’s a plumbing problem.”
The distinction matters because it points toward different remedies. If the issue is human behaviour, the answer is training and awareness. If the issue is infrastructure, the answer is architectural change at the routing layer, not coaching sessions.
Caps, Coaching, and the Limits of Both
The companies grappling with this have landed on strikingly different approaches, and none of them has clean proof that their approach is working.
Rick Spencer, GM of technology and product at SUSE, argues that the instinct to cut usage is itself the wrong starting point. “The first reaction should not be to ‘use less AI’. High token usage can mean a lot of different things. If someone is using $16,000 in tokens to save us $100,000, of course we want to encourage that,” he told VentureBeat. SUSE sorts AI usage into three categories that managers coach against: Daily Work, Autonomous Agents, and Curve Jumping for one-time strategic efforts. Notably, SUSE runs no automated routing layer at all, leaving model selection to individual developers, though Spencer says a proxy is on the roadmap.
Agiloft went in the opposite direction and scrapped its spending caps entirely. “74% of people were never hitting the old caps anyway. Caps were a false ceiling that created friction for heavy users without addressing the actual cost drivers,” Ramos said. Instead, the company made cheaper models the default and built escalation to frontier models into the infrastructure layer, routing by task complexity rather than leaving the choice to whoever is typing the prompt. “Don’t ration the tool. Fix the architecture underneath it. Scarcity governance is a patch. Intelligent routing is the fix,” Ramos said.
Everlaw, which provides AI tools for litigation and investigations, runs per-person token caps but frames them as the opposite of rationing. When an engineer hits the cap, a one-line email goes to the tools team and the cap is typically doubled the same day. Everlaw is also the only company in this conversation that offered hard numbers. One project on core Java infrastructure consumed $3,500 in tokens and compressed implementation time from 9.5 engineer-months to 2.5. A larger, still-unlaunched product has consumed $27,000 in tokens, is likely to reach $40,000, and cut estimated engineering time from 90 to 100 months down to 19. “These are real numbers; the ROI of $3500 to save seven months of engineering time is simply a no-brainer,” CTO Max Christoff said.
Not every Everlaw experiment worked. The team spent thousands having agents port interface code from Dojo to React and discarded the output entirely, because the two frameworks rest on fundamentally different assumptions about how state and view relate. The lesson they drew was to have the agent document the old system’s behaviour first and then build from that documentation rather than translating code directly.
The Routing Layer as the Real Battleground
A growing number of vendors are now selling the infrastructure fix that Ramos describes. Merge, Databricks, AWS Bedrock, and Azure AI Foundry all offer auto-routers that attempt to match task complexity to the most cost-appropriate model automatically. Databricks introduced Smart Routing inside its Unity AI Gateway at the Data and AI Summit in June, bundling it with hard spend caps and cost attribution across hosted models, coding agents, and custom agents. The feature is currently in beta.
David Nasi, director of product management at Databricks, explained that the router evaluates each request using a combination of deterministic signals and model-based classification, examining factors including prompt intent and length, referenced files, stack traces, the scope of the change, reasoning depth, and execution complexity. Crucially, the router is not only selecting a model. “We’ve found that the same model performs differently when leveraged with a different harness, so we’ve built Smart Routing to include that flexibility,” Nasi told VentureBeat. When a request hits a budget ceiling, administrators choose between a hard block and a fallback that tries a cheaper compliant model first.
Agentic workloads complicate routing further because a single task can fan out into dozens of model calls with no human approving each one. Databricks addresses this by evaluating at execution boundaries rather than per individual call. Early signals are encouraging: “Teams that previously ran all traffic through frontier models are starting to shift routine tasks, such as boilerplate generation, simple bug fixes, or minor edits, to lower-cost models without measurable drops in resolution rates,” Nasi said.
Even so, validating whether quality has actually held up after a routing change remains the customer’s problem. Palaniichuk described this as genuinely hard, noting that “you’re imposing deterministic checks on non-deterministic output, and models ship on roughly a quarterly cadence, meaning an eval tuned to one model doesn’t cleanly transfer to the next.” Christoff’s answer at Everlaw is to give engineers a wide model menu and a dollar budget and let them choose, on the theory that the person reviewing the output builds the fastest instinct for which model produces work worth keeping. He also flags a failure mode that no routing layer catches: “The coding agent will propose twenty surface-level fixes instead of addressing an underlying pattern,” leaving code that passes every test but that nobody can maintain.
Why the Discipline Will Outlast the Tactics
Both Palaniichuk and Christoff expect the specific optimisations they are running today to become obsolete as models and pricing evolve. What they believe will survive is the underlying habit of matching the tool to the task consciously rather than defaulting to whatever the vendor pre-selected.
Christoff offered a prediction that carries real implications for how technology budgets get built. “My prediction is that over the next one to two years, we’ll start to see token expense treated more like annual headcount or production costs, and less like an IT or general software expense,” he said. Department heads will walk into annual planning with a position on headcount and a position on tokens, each with its own business case. One team might plan for millions in tokens and almost no new hires. Another might invert that entirely.
For businesses in Malaysia and Singapore that are accelerating AI adoption across software development, legal, and financial services, the Uber episode is a useful warning before the bill arrives rather than after. The companies that emerge from this period with genuine returns will not be the ones that used the most AI. They will be the ones that built the measurement discipline to know the difference between consumption and value, and the infrastructure to act on that difference automatically.
Read More: GitHub’s HydraFusion Cuts AI Coding Costs Sharply, But the Quality Story Is More Complicated
