Companies Are Discovering The Cost Of Letting AI Agents Work Longer

OpenAI researchers using coding agents have pushed daily token consumption into territory that looks less like a software subscription and more like computing infrastructure.

Business Insider reported that some heavy internal users were spending more than $7,000 a day in tokens, while median daily use among researchers climbed sharply over the summer.

The numbers come from an AI laboratory rather than an ordinary corporate IT department, but they expose a cost model that enterprises will meet as agents move from answering questions to performing hours of work.

A chatbot has an obvious interaction pattern. An employee sends a prompt. The model answers. Usage ends. An agent reads files, plans a task, calls tools, revises its plan, writes code, runs tests, interprets errors and starts another iteration. Each step consumes model inference. Parallel agents multiply the work again.

Companies therefore need to measure AI productivity in units larger than the price of one prompt. A coding agent that costs €40 to complete a task once performed by an engineer for three hours looks inexpensive. An agent spending €40 on a task that an engineer reviews for two hours and then rewrites has produced a different economic result.

Wipro offers evidence from the other side. The company says AI has freed capacity equivalent to the output of around 20,000 workers and has reassigned that capacity elsewhere. The Indian IT services group has trained more than 100,000 employees in advanced AI skills as it reorganises work around human and AI collaboration.

Enterprises now need to connect those two ledgers: compute consumed and work removed.

Many early AI programmes measured licences or adoption. Finance teams knew how many employees had access to Copilot or another assistant. Those metrics worked while the software behaved like a seat-based productivity tool.

Agentic systems consume resources according to workload.

One employee might run a small agent twice a week. Another engineering team might send millions of tokens through several models every day. Procurement cannot understand the second workload by counting seats.

Routing becomes part of cost control.

A company does not need its most expensive model to classify every document or perform every intermediate step. Smaller models handle routine work while stronger models receive tasks that require deeper reasoning. Local models serve some predictable workloads without metered API calls, although the company still pays for the infrastructure underneath them.

Engineering teams already optimise databases, cloud instances and storage because usage creates cost. AI inference belongs in the same discipline.

Budgets also need limits that understand the task. A hard token ceiling prevents runaway spending but might stop useful work halfway through. An agent that has repeated the same failed test twenty times deserves a different intervention from one processing a legitimate archive containing thousands of files.

Observability therefore becomes financial as well as technical. Teams need to see which agent consumed the resources, which workflow triggered the run, how many retries occurred and whether a person accepted the output.

A company that cannot connect inference spend to completed work cannot calculate the return from agentic AI. The first generation of enterprise AI sold access. The next one sends a compute bill.