ON FUTURE OF WORK

Tokenomics Is the Budget. Humanomics Is the Bound.

About the Author

Rajeev Ronanki

Chief Executive Officer

Show bio ↓

Tokenomics is what you can compute, and it scales on its own. Humanomics is the harder discipline: deciding what is worth computing, and where the human line falls.

Barely four months into 2026, Uber had already blown through its entire AI coding budget for the year. Adoption tore through the engineering org, from a third of engineers in February to 84% a month later, driven in part by internal leaderboards that ranked people on how much AI they used. Then April came and the money was simply gone. By June the company was capping spend at $1,500 a month per employee, per tool, just to slow the burn. The people who owned the budget just watched it arrive.

In June, the Linux Foundation gave it a name. It launched the Tokenomics Foundation, a standards body meant to bring the same cost discipline to AI tokens that FinOps once brought to cloud spend. The reason was plain enough. One of its organizers, J.R. Storment, told TechCrunch that companies had begun calling in April to say they were already three times over their entire 2026 token budget, with eight months still on the calendar.

That is tokenomics in a sentence. It measures what you can compute, and it scales the moment you let it. Compute does not wait for a business case, and it does not ask permission. It just runs, and the bill arrives a quarter or two later, well after anyone remembers approving the work that caused it.

The figures point in two directions at once. Per-token prices keep sliding, which reads like relief on a pricing page. Underneath, the volume tells another story. EY estimates that a customer-service interaction costing about four cents in 2023 runs closer to $1.20 today, once you add tools, reasoning, and the loops an agent runs to check its own work. That is roughly thirty times more. Jellyfish's head of research, Nicholas Arcolano, told TechCrunch that spend per developer climbed about 18.6 times in nine months. So the price of thinking falls while the cost of thinking rises, and a finance team reading only the rate card never sees the turn coming.

This is where most of the conversation stops, and where the harder part starts. Tokenomics is an accounting question. It has an owner, a dashboard, and now a foundation with its name on the door. Humanomics has none of that yet. It asks what is worth computing at all, and where the line falls between the work a machine should carry and the judgment that stays with a person. A budget is a number you are cleared to spend. A bound is a stance on what your company is for, and which of its decisions should never go to a system tuned for the wrong objective. This is not a contest between people and machines over who costs less. The bound is the discipline of aiming both at the same outcome, so the compute you can afford goes to the results you actually want. You can hold a huge budget and no bound at all. That is more or less how a company spends $500 million dollars in a month on AI, without noticing.

The people selling the tokens have started saying this out loud. In June, Sam Altman told CNBC that whether AI spending ever pays for itself is the most fair criticism of AI going right now, and that customer worry about cost had moved from something he never used to hear to the second most common thing he does. When the vendor grants the ROI question, it deserves a real answer. PwC’s April 2026 AI Performance Study put numbers on the split: roughly three-quarters of the measured economic return from AI had pooled into about a fifth of companies, and everyone else was left dividing what remained. A great deal of activity; the number it was meant to move has barely twitched for most.

That leaderboard logic did not stay at Uber. Plenty of companies made the meter the goal, ranking teams by how many tokens they burned on the assumption that more usage meant more output. It is an easy thing to measure, which is most of its appeal, and it tells you almost nothing about whether the work was worth doing. Arcolano put the catch plainly: whether heavy spend pays off comes down to the business value of what actually ships, and most companies still can't measure that.

Jellyfish's own data shows why it matters. Its heaviest token users were roughly twice as productive as lighter ones, yet they spent ten times the tokens to get there. The leaderboard measured effort. Nobody was measuring whether the effort added up to anything.

The bound rests on the one input that does not get cheaper as compute does: human judgment, spent at the moment a decision gets made. And that judgment is under pressure from the very tools meant to support it. A World Economic Forum essay in June described “human in the loop” as a phrase we repeat for reassurance, without asking whether the person in that loop can actually overrule a confident, wrong machine. Its example is a bank manager in Kuala Lumpur named Diana, whose name sits on every lending call the model makes, and who interrogates its output rather than initialing it. Her employer can buy a sharper model next quarter. What Diana knows is not for sale on the same terms. The danger is measurable, too: studies of automation bias find that bad machine advice nudges people toward the wrong decision about a quarter of the time, and experience offers no immunity. In a survey of business leaders reported this spring, most now lean on AI for the majority of their decisions, and most also said their teams argue less than they once did. Debate is friction. Friction is frequently where judgment lives.

None of this is an argument for spending less. Spend heavily, wherever the compute earns its keep. The real work is drawing the bound, and that includes a third thing budgets ignore: the signals a leader can no longer afford to miss. In most organizations, the majority of the information that should drive a decision, by some counts as much as ninety-five percent, never reaches the person who could act while it still matters. It sits in a system three teams away, correctly logged, connected to nothing. Compute will not solve that. Add more agents and the room only gets louder. Put fifty to a hundred agents beside every employee, across thousands of them, and the scarce input stops being tokens. It becomes the attention to decide what the tokens were for.

That coordinating layer has a name: Systems of Attention, the fourth layer of enterprise software after Record, Engagement, and Insight. Signal Labs builds it. SignalOS™ allocates attention against the objective, lifting the few signals that matter out of the noise and putting them in front of the person who owns the decision, inside the window where the decision still counts. A dashboard reports what already broke; Systems of Attention convene the enterprise on what is breaking now, in time to act. Same engine, every industry, and only the vocabulary changes.

Which brings us back to the question underneath all of it. Tokenomics is the budget, and the budget will keep growing on its own, because that is what compute does. Humanomics is the bound, and the bound will only ever be a set of human decisions about what the institution is actually for. One of them has to answer to the other. The companies that last will be the ones that spent on compute the way you spend on anything serious, against an outcome someone owns, before the meter decided the outcome for them.