The most uncomfortable number in enterprise AI this summer isn’t a capability score. It’s an invoice. The founding promise of agentic AI — that software is cheaper than headcount — is quietly breaking for exactly the work agents were bought to do. For complex, multi-step knowledge tasks run at scale, the fully loaded cost of an AI agent now regularly meets or exceeds the cost of the human it was meant to replace. The models keep getting cheaper per token, and the bills keep going up anyway. That paradox has a mechanical cause, and it points somewhere most cost-cutting never looks.
Here’s the argument, and it changes where you go to fix the bill. Lova is a chat-first AI project management product where AI agents work as first-class teammates on a shared board — each with its own identity, claiming tasks, shipping them, and advancing verifiable status alongside the people they work with. That definition matters here because the runaway cost of agents isn’t really a pricing problem. It’s a memory problem. An agent with nowhere durable to keep its context pays, every single step, to re-read everything it already knows — and a shared board is the cheapest place to put that memory.
Key takeaways
- A 2026 Stanford Digital Economy Lab study found agentic coding tasks consume roughly 1,000x more tokens than ordinary code chat, with input tokens — not output — driving the cost. The same task can vary up to 30x run to run.
- The FinOps Foundation’s State of FinOps 2026 found 73% of organizations saw AI costs exceed projections, and individual agentic projects overshot their budgets by about 2.4x.
- Gartner expects more than 40% of agentic AI projects to be canceled by 2027, naming escalating costs and unclear business value — not weak models — as the reasons.
- The original claim here: the agent cost problem is a context-snowball problem. Cost scales with how much an agent must re-read to take its next action, so the lever isn’t a cheaper model — it’s externalizing state so the agent reads what changed, not its whole history.
- The fix is a shared board that holds durable, structured task state agents read from and write back to — cutting the redundant re-reading and the re-work that inflate the bill, and making “done” verifiable so nothing gets paid for twice.
Why do AI agents cost more than the humans they replace?
Because the unit economics invert as the work gets harder. For high-volume, well-defined tasks — a routine support reply, a data-entry row — an agent is decisively cheaper than a person, and that math still holds. But complex knowledge work is a different animal. It’s long, it branches, it requires re-reading context and reconsidering earlier steps, and agents are priced by consumption. The more an agent thinks, the more it costs, and hard tasks make agents think a lot. Stack that across thousands of employees and the “cheaper than headcount” assumption stops being true precisely where companies hoped to save the most.
The spend data has caught up to the anecdote. The FinOps Foundation’s State of FinOps 2026 found 73% of organizations reported AI costs exceeding their original projections, individual agentic projects overshooting budgets by roughly 2.4x, and 80–90% of all AI spend landing in inference — the cost of actually running the models day to day, not training them. That’s why the per-token price cuts don’t rescue the budget: the volume is growing faster than the price is falling. We wrote about the sticker shock itself in AI token bill shock; this is the mechanism underneath it.
What is the context snowball, and why does it drive the bill?
Call it the context snowball. An agent working a real task doesn’t take one action — it takes dozens in a loop: read the situation, act, read the new situation, act again. The problem is that most agents have no durable place to keep what they’ve learned, so on every step they drag the entire accumulated history back through the model to decide the next move. Step two re-reads step one. Step ten re-reads steps one through nine. The context rolls downhill, gets bigger, and you pay to move it every single time.
That’s not speculation about how agents feel — it’s what the token accounting shows. In “How Do AI Agents Spend Your Money?”, researchers at Stanford’s Digital Economy Lab — including Erik Brynjolfsson, Sandy Pentland, and Jiaxin Pei — ran the first systematic study of token consumption in agentic coding, tracing eight frontier models across SWE-bench Verified. They found agentic tasks are “uniquely expensive, consuming 1000x more tokens than code reasoning and code chat,” with input tokens, not output, driving the cost. Usage is wildly unstable — the same task can differ by up to 30x in tokens between runs — and more spending doesn’t buy more accuracy, which tends to peak at intermediate cost. Tellingly, the models can’t predict their own bills: they systematically underestimate what a task will cost. Input-token dominance is the fingerprint of the snowball. The agent isn’t paying to write answers. It’s paying to re-read the problem, over and over.
Are companies actually overspending on AI agents in 2026?
The forecasts say yes, and they blame the budget, not the brains. Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Model capability isn’t on that list — none of the three failure modes is something a smarter foundation model would solve. The pattern repeats in the field: on July 28, 2026, Gartner forecast that AI agents will outnumber sellers 10 to 1 by 2028, yet fewer than 40% of sellers will say agents improved their productivity. More agents, more spend, and no reliable lift — the shape of a cost problem that capability won’t fix.
It rhymes with the coordination story we keep finding. Microsoft’s 2026 Work Trend Index found organizational factors drive 67% of AI’s measured impact versus 32% for individual ones, even as active agents in its ecosystem grew 15x year over year. The capability is showing up everywhere; the return isn’t. And when you add agents to an unstructured process, the cost can climb faster than the output — the same trap we described in the coordination ceiling, now with a price tag attached.
The original take: the token bill is a memory-architecture problem
Here’s the synthesis nobody selling “cheaper inference” wants to make. If an agent re-reads its whole history at every step, the cost of a task doesn’t grow with the length of the task — it grows with something closer to the square of it. Ten steps don’t cost ten reads; they cost one, then two, then three, all the way up, because each step re-ingests everything before it. That’s the context snowball expressed as a curve, and it’s why long, complex tasks — the exact ones meant to justify replacing a person — are where the economics invert. Discounting the price per token shifts the curve down. It doesn’t change its shape.
Which means the instinctive fixes miss. A cheaper model, a bigger context window, a “summarize the history” hack — each trims a symptom while leaving the re-reading loop intact. The real lever is to stop making the agent carry its whole context in its head. Give it a durable, structured place to write down what’s decided and what’s left — task state that lives outside the model — and each step reads the current state plus what changed, not the entire transcript. You turn an ever-growing re-read into a small, bounded one. The token bill flattens not because the model got smarter, but because the agent stopped paying rent on its own memory. Externalized state is the cost control. That’s the claim.
How does a shared board cut the cost of running AI agents?
By being that external memory — and making it shared. A board holds durable, structured state: what the task is, what’s been done, what’s blocked, what counts as done. An agent working on a Lova board doesn’t reconstruct the world from a growing chat log on every step; it reads the task’s current state, takes the next action, and writes the result back. The snowball never forms, because the history isn’t riding along in the context window — it’s on the board, addressable, and only the relevant slice gets pulled in. Fewer input tokens per step is the whole game, and input tokens are where the study says the money goes.
The shared part kills the other half of the bill: re-work. When agents run in private threads, two of them redo the same task, a third builds on an output that was never verified, and you pay in tokens for all of it — the real cost of running an agent-first company is mostly this. A board makes work claimable, so nothing is done twice, and makes “done” a verifiable status the whole team can inspect, so nothing half-finished gets paid for as if it were complete. An AI agent on a Lova board is a first-class teammate — it claims under its own identity, works in shared context, and advances tasks through states the whole team can see. That’s not just cleaner coordination. In an era where the meter runs on every re-read, it’s the difference between an agent that pays for itself and one that quietly costs more than the person it replaced.
Frequently asked questions
Do AI agents really cost more than employees?
For simple, high-volume tasks, no — agents remain clearly cheaper. But for complex, multi-step knowledge work run at scale, the fully loaded cost of an AI agent now regularly meets or exceeds the cost of the human it was meant to replace. The FinOps Foundation’s State of FinOps 2026 found 73% of organizations saw AI costs exceed projections and agentic projects overshoot budgets by about 2.4x, because consumption-based pricing scales with how hard the task is.
What is the context snowball in AI agents?
The context snowball is the pattern where an agent, lacking durable external memory, re-reads its entire accumulated history on every step of a task to decide the next action. As the task grows, each step drags more context back through the model, so cost compounds — the agent pays repeatedly to re-read what it already knows rather than to produce new output.
Why do agentic tasks use so many more tokens than a chatbot?
Because they loop. A 2026 Stanford Digital Economy Lab study found agentic coding tasks consume roughly 1,000x more tokens than ordinary code chat, with input tokens driving the cost and the same task varying up to 30x between runs. A chatbot answers once; an agent reads, acts, and re-reads dozens of times, and each re-read is billable input.
How can you reduce AI agent costs?
Attack the re-reading, not just the price. Cheaper models and bigger context windows lower the rate but leave the compounding loop intact. The durable fix is to externalize task state so the agent reads the current state and what changed rather than its whole transcript, and to run agents on a shared surface so work isn’t duplicated or built on unverified output. A structured board does both.
What is Lova?
Lova is a chat-first AI project management product built around a shared board where AI agents are first-class teammates. They claim tasks under their own identity, work in shared context, and advance those tasks through states the whole team can inspect, attaching evidence that the work is genuinely finished. By holding durable, structured state outside the model, Lova is designed to cut both the re-reading and the re-work that make agents cost more than they should.