---
title: "AI agents need a capacity budget"
description: "OpenAI and Google show that agents raise output and consumption together. Why organisations should manage capacity, cost, and approvals as one control loop."
date: 2026-09-07
lang: en
tags: [governance, cost, automation, news]
author: "thinkai.at"
canonical: https://thinkai.at/en/blog/ai-agent-capacity-budget/
---

# AI agents need a capacity budget

**AI agents** are rapidly becoming a capacity question, not merely a licensing decision. On September 6, OpenAI reported from its own research operation that researchers used 3.1 agent workdays for every human workday by mid-August. Google Cloud is also highlighting new flexible billing and cost controls for agent workloads.

## More execution requires more control

OpenAI’s figures are not a general benchmark for other organisations. They do, however, expose an effect that IT and finance teams should plan for: when teams run several agents concurrently, code, experiments, and throughput increase. Inference consumption, tool calls, sandbox runtime, and potential downstream costs rise at the same time.

OpenAI reports more than US$600 of daily API consumption for its median researcher and more than US$7,000 at the 90th percentile. Those figures cannot be transferred to a mid-sized company. The direction still matters: a fixed per-user licence does not fully represent the economic and operational consumption of an autonomous system doing work.

## From licence budget to work budget

Google frames its current capabilities explicitly as FinOps for AI: flexible billing and cost controls for agents are intended to make expenditure manageable. This suggests a practical operating model. Do not only grant access to an agent; give each use case a bounded work budget.

That budget should connect four measures:

- **Purpose:** What verifiable outcome may the agent produce?
- **Capacity:** How many concurrent runs, tool calls, and runtime are permitted?
- **Cost:** What monthly and escalation limit applies, including downstream cloud services?
- **Approval:** Who may increase the budget, model tier, or permissions?

## What this means for DACH organisations

Do not start with a general “agent flat rate.” Choose one defined process, such as test-case creation or document classification, and measure throughput, quality corrections, human review time, and total cost together. A cost limit without a quality metric merely creates pressure to save; a productivity target without a limit hides consumption.

The object of control is therefore not the chat, but the work performed. Organisations that make capacity, spending, and decision rights visible for each agent workflow can scale without the bill rising faster than demonstrable value.

## Sources

- [OpenAI — Research acceleration: The view inside OpenAI](https://openai.com/index/research-acceleration-view-inside-openai/)
- [Google Cloud — What Google Cloud announced in AI this month](https://cloud.google.com/blog/products/ai-machine-learning/what-google-cloud-announced-in-ai-this-month)