This June, the biggest AI story was not a new model or a new capability. It was a bill. Companies that budgeted for a full year of AI usage burned through the entire amount in four months. Uber watched adoption of AI coding tools jump from 32% to 84% of its engineering org, and by April the annual AI budget was gone. One enterprise reportedly spent 500 million dollars in a single month because nobody set a usage cap.
Everyone spent 2025 asking whether AI agents work. In 2026 the question changed: can you afford to run them? I run OpenLoop, a small AI-first agency in Kashmir, and this is the conversation I now have with every client before we write a single line of code. Here is what the token bill actually looks like, and how a small business avoids the trap the enterprises just walked into.
Why Agents Cost 30 Times More Than Chatbots
A chatbot conversation is cheap. You send a question, the model sends an answer, and the whole exchange might cost a fraction of a cent. The mental model most business owners have for AI pricing was formed by this experience, and that mental model is now wrong by an order of magnitude.
An agent does not answer one question. It reasons through a task in a loop. It reads your instructions, calls a tool, reads the result, thinks about what to do next, calls another tool, checks its own work, and repeats. Every one of those steps consumes tokens, and every step carries the full context of everything that came before it. Industry data from this year puts the multiplier at 5 to 30 times more tokens per task than a standard chat. One analysis found that a typical orchestrated agent interaction now costs around $1.20, compared to about 4 cents for a simple chat interaction in 2023.
That is not a pricing scandal. It is just what the work costs. The scandal is that most companies discovered the multiplier after deployment, when the production bill arrived. The deployment decision came before the cost model, and that order is exactly backwards.
The ISP Lesson: Nobody Budgets for Usage They Cannot See
I recognized this failure pattern immediately, because I have lived it in a different industry. At ViberNet, the fiber ISP I run in Kashmir, bandwidth is a metered cost on our side and a flat fee on the customer's side. The entire business depends on knowing, precisely, how much usage the average customer generates. If we guessed wrong by 30x, we would be dead in a quarter.
Enterprise AI teams in 2026 essentially sold themselves unlimited plans without ever measuring usage. They assumed conversational volumes, deployed agentic workloads, and let every employee run loops all day with no caps, no per-task cost visibility, and no owner watching the meter. Microsoft started canceling internal AI coding licenses over runaway token bills. A routine tooling contract renewal at Priceline reportedly came back 4 to 5 times more expensive.
The fix is not mysterious. It is the same discipline any infrastructure business learns in year one: meter everything, cap by default, and know your cost per unit of work before you scale the units.
What a Task Should Actually Cost
Here is the exercise we now run at OpenLoop before building anything. Take the task the agent will do. Estimate the tokens per run, honestly, including the loops and retries. Multiply by the runs per month. Then compare that number against what the task costs today with a human doing it, or against what the task earns.
A billing triage agent that costs 8 rupees per ticket and saves a staff member ten minutes is an obvious yes. A research agent that burns 300 rupees per query to produce a summary someone skims once is an obvious no, even though the demo looks impressive. The math is not complicated. What is rare is doing it before deployment instead of after.
This is also where understanding AI agents in business pays off in a very literal way: architecture is the biggest cost lever you have. A narrow agent with a short prompt, minimal context, and three tools costs a fraction of a general agent that hauls your entire knowledge base into every step. The design choices that make agents reliable, small scope, few steps, human checkpoints, are the same choices that make them cheap. That is not a coincidence. Both reliability and cost degrade with every extra loop.
Get new posts by email, daily-ish writing on ISP economics, AI in production, and building in Kashmir.
The Small Business Advantage, Again
There is a strange comfort in watching companies with thousand-times-larger budgets make this mistake, because it confirms something I keep seeing from Srinagar: constraints are a feature. A Kashmiri shop owner spending 2,000 rupees a month on an AI workflow checks the bill. Every month. The moment cost drifts past value, the workflow gets fixed or killed. There is no budget slack to hide a bad decision in.
Enterprises had the opposite problem. The budgets were large enough that a 30x usage miss could hide for a full quarter. By the time finance noticed, the overspend was headline-sized. Small businesses adopting AI in 2026 should not envy that scale. Being forced to know your unit economics from day one is the advantage.
Practically, it looks like this. Set a hard monthly spend cap on every API key before anything ships. Log cost per task, not just cost per month, so you know which workflow is eating the budget. Prefer cheaper, smaller models for the boring steps and save the expensive model for the one step that needs judgment. And re-check the bill when you change anything, because a single added tool or a longer prompt quietly compounds across every run.
The Bill Is the Feedback
My family name, Shahmiri, traces back to the Shah Mir dynasty that governed Kashmir for over two centuries. You do not hold anything together that long by ignoring what things cost. Sustainable operations, whether a medieval administration, a fiber network, or an AI workflow, survive on boring accounting, not on impressive capability.
So if you are adding agents to your business this year, treat the token bill as information, not as an annoyance. It tells you which workflows are actually designed well. A lean bill usually means a lean agent, and a lean agent is the one that also behaves reliably in production. The companies that blew their budgets in four months did not have an AI problem. They had a measurement problem.
If you want help scoping an AI workflow where the cost math is done before the deployment, not after, that is exactly the kind of project OpenLoop takes on. Reach out through the contact page. The first thing we will build together is the spreadsheet, not the agent.
