By Arun “Rak” Ramchandran, CEO, QBurst
Walking into a boardroom full of AI agents isn’t some far-off possibility anymore, it’s close to becoming the default. The real question isn’t whether that day arrives. It’s how we get there: having exhausted token budgets chasing consumption with nothing concrete to show for it, or having built something with a return anyone can actually point to. The future doesn’t belong to whoever deploys the most AI. It belongs to whoever makes it purposeful.
Sometime around April 2026, it came out that Meta had been running an internal leaderboard called Claudeonomics, ranking roughly 85,000 employees by how many tokens they burned. Top users were clocking hundreds of billions of tokens a month and picking up titles like “Token Legend, and within weeks, “tokenmaxxing” had a name, and enterprises everywhere started asking the same anxious question: are we spending too much on AI, and how do we make it stop.
I understand the instinct. What I don’t think enterprises have got right yet is the diagnosis.
Tokenmaxxing isn’t the disease. It’s a symptom, and it’s pointing at the wrong number.
The Wrong Number
A token is a unit of billing. It tells you how much you consumed, not what you got for it. Someone burning tokens on a trivial, low-value task looks identical on a dashboard to someone resolving a genuine customer problem, and neither looks any different from an agent stuck in a loop, quietly consuming budget on nothing at all. If your only lever is a flat per-employee cap, you’re treating all three the same, and that’s precisely the problem. A customer-facing task that creates real value should carry a meaningfully higher allowance than routine internal work. It shouldn’t be the same number, whatever that number is, for both.
Uber’s experience earlier this year is a preview of where this goes if you don’t fix it. The company burned through an entire year’s AI coding budget in about four months. That wasn’t a token pricing problem. It was a task visibility problem: nobody had a clean answer for what all that spend was actually buying. The instinct to cap spend after something like that isn’t wrong. But a cap without task-level understanding just moves the friction somewhere else. You’ll keep hitting the ceiling and never answer the question that actually matters: what is this spend buying us.
Where the Real Cost Lives
Here’s what I think most finance and technology leaders are still missing: token cost is the cheapest and most visible layer of a much bigger bill. I think about agentic AI’s total cost of ownership across four layers, and only the first one shows up prominently on any invoice.
The inference base, the tokens and API calls everyone budgets for, is the layer that’s actually gotten dramatically cheaper. Two years ago we were paying roughly $60 per million tokens. Today it’s 30 to 50 cents, close to a 98 to 99 percent drop. If tokens were the whole story, AI economics should be getting easier every quarter.
They aren’t, because the other three layers are growing faster than the first one is shrinking. Orchestration overhead is real: a single customer interaction can now trigger half a dozen specialized agents working together, and every hand-off between them adds compute and latency. Salesforce’s own Connectivity Benchmark found enterprises already run an average of 12 AI agents, with that number projected to grow 67 percent over the next two years. Integration and data access, connecting agents to commerce platforms, customer databases, inventory and supply chain systems, and keeping those connections accurate and secure, ends up being 40 to 60 percent of true cost of ownership in many deployments, and it almost never shows up in the original budget proposal. And governance, observability, audit logging, human review, compliance, can add another 40 to 80 percent on top in regulated environments.
That governance number isn’t theoretical. When Air Canada tried to argue that its chatbot was a separate legal entity and the company wasn’t liable for the incorrect information it gave a customer, the tribunal rejected that outright. As agent interactions multiply, and IDC expects agent API calls to grow a thousandfold by 2027, the requirements for monitoring, compliance, and accountability are going to compound right alongside them, whether or not anyone budgeted for it.
The Economy You Can’t See
There’s a second layer of hidden cost that I think gets even less attention, and it isn’t only happening on the employee side. Yes, people running work through personal AI accounts is real, and it needs a governance policy, not just a cap. But SaaS platforms your enterprise already pays for are increasingly shipping agentic features baked into the subscription. That consumption doesn’t show up as a new AI line item.
It’s absorbed into a bill you already thought you understood. Add in the fact that agents now call other agents, some from vendors, some built in-house, and you have a cost layer most companies aren’t measuring at all, let alone tying back to output. A fair amount of what gets marketed as agentic commerce today is honestly sophisticated automation wearing new vocabulary, and that blurring makes the shadow economy harder to see, not easier.
The fix isn’t a policy memo. It’s what I’d call Glass Box AI: every agent, wherever it comes from, should have clear provenance and be able to show its work, what data it used, what rule it applied, how it arrived at an output. You cannot govern a cost, or a decision, that you cannot see, and right now most enterprises can’t see most of this.
From Tokenmaxxing to Valuemaxxing
None of this means enterprises should stop watching token spend. It means token spend is the easiest number to watch and the least useful one to optimize in isolation. The shift I’d encourage every enterprise to make is from tokenmaxxing to what people are starting to call valuemaxxing: stop asking how to cut tokens.
Start asking what it actually costs to complete a task, and whether there’s a cheaper way to get the same outcome. Build for optionality. an open architecture that lets you route work to the right model instead of defaulting to the most expensive one out of habit. And don’t ignore the organisational cost nobody puts in the model: reskilling people, cleaning up data, redesigning workflows.
You cannot automate chaos. Agents don’t fix messy data; they amplify it, faster than a person ever could.
The enterprises getting this right have stopped being fascinated by AI and started being accountable for it. That’s the conversation the industry needs to have before more companies find themselves where Uber did.