# How to control AI agent costs: limits, economy mode and cost per task

What drives AI agent spend and how to cap it: monthly hard limits, economy mode, batching and cost per task, with a worked example for a small business.

Published 2026-11-05. Updated 2026-11-05. Source: https://beereau.com/blog/control-ai-agent-costs

> **Short answer:** To control AI agent costs, set a hard monthly limit, track cost per task, and decide in advance what happens near the limit. The main drivers of spend are how much text each task reads and writes, how many steps and retries it takes, and which model it uses. Batching non-urgent work and an automatic economy mode keep urgent tasks running while the rest gets cheaper.

AI agents do not charge like software seats. They charge like a contractor billing for minutes, except the minutes are text processed. That makes them cheap on a quiet month and surprising on a busy one. This guide explains what drives the bill and the four controls that keep it predictable.

## What drives AI agent costs?

Most AI services bill by **tokens**, which are small chunks of text, roughly parts of words. Every time an agent reads an email, your policies or its own earlier notes, those tokens count as input. Every reply, summary or plan it writes counts as output.

So the cost of a task comes from a few things:

| Cost driver | Why it adds up | How to reduce it |
|---|---|---|
| Context size | The agent rereads long threads, policies and history on each step | Summarise long threads; share only the docs a task needs |
| Number of steps | Planning, tool calls and reviews each cost a round trip | Clear instructions; fewer hand-offs for simple tickets |
| Retries and loops | An agent stuck on a failing step can repeat it many times | Step limits and timeouts per task |
| Model choice | Stronger models cost more per token | Use smaller models for sorting and extraction |
| Urgency | Real-time requests cost full price | Batch anything that can wait |
| Reviews | A second agent checking a draft doubles part of the work | Review only risky categories, not every reply |

Example: a refund question that arrives with a 12-message thread costs more than a one-line "where is my order?", even if both get a two-sentence reply, because the agent has to read the whole thread first.

## How does a monthly hard limit work?

A **hard limit** is a ceiling on spend that the system enforces, not just reports. A **soft limit** sends you an alert while spending continues. For a small business, only the hard kind protects you from the bad month: a spam wave, a product recall or an agent stuck in a loop over a weekend.

A good hard limit has three stages:

1. **Normal:** everything runs at full speed.
2. **Warning zone:** spend reaches a threshold (for example 80%), and the system starts saving money on its own.
3. **Limit reached:** you choose what happens next, instead of the system choosing for you.

In beereau, every office has a hard monthly limit. At 80% it switches to economy mode: urgent work continues and the rest is batched. At 100% you choose to pause or allow a capped overage. Nothing goes over silently.

## What is economy mode, and what should it change?

**Economy mode** is a set of cheaper behaviours that kick in near the limit. It should change *how* work is done, not drop work. Sensible economy-mode rules:

- Urgent tickets (angry customers, payment problems, anything with a deadline) keep full priority.
- Non-urgent tickets wait and are processed together at set times.
- Weekly reports and long summaries get shorter.
- Optional second-pass reviews are skipped for low-risk categories.

The point is that your customers should not notice economy mode on the things that matter. A customer whose order is lost still gets an answer today. A customer asking about next season's colours may get it this afternoon instead of this minute.

## Why does batching cut AI costs?

**Batching** means collecting tasks that do not need an instant answer and processing them together. Providers reward this because it lets them use idle capacity.

The discounts are real and documented. OpenAI's Batch API gives a "50% cost discount compared to synchronous APIs" with a 24-hour completion window ([OpenAI docs](https://developers.openai.com/api/docs/guides/batch)). Anthropic's Message Batches API also cuts costs by 50%, with most batches finishing in less than an hour ([Anthropic docs](https://platform.claude.com/docs/en/build-with-claude/batch-processing)).

A related lever is **caching**: when an agent reuses the same long instructions or documents on every task, some providers charge much less for the repeated part. Anthropic, for instance, prices cache reads at 0.1 times the normal input price for most of its models ([Anthropic docs](https://platform.claude.com/docs/en/build-with-claude/prompt-caching)). You will not set this up yourself in a finished product, but it is worth asking any vendor whether they use it, because it is part of why some AI teams cost less than others.

What can be batched in a support inbox:

- Product questions with no deadline
- Follow-ups to customers who have gone quiet
- Tagging, summaries and the weekly report
- Drafts that wait for your approval anyway

What cannot be batched: anything where delay creates a new problem, like a failed payment, a shipping error or a complaint that is already escalating.

## How do you measure cost per task?

**Cost per task** is total AI spend for a period divided by the number of tasks completed, ideally split by task type. It is the number that tells you whether an AI team is worth it.

To get it, you need a record of every task with three fields: who handled it, what it cost and who approved it. beereau shows exactly that on every task, so you can see which kinds of work are expensive and why.

Example: an office spends $90 of AI usage in a month and closes 1,200 tickets. Average cost per ticket is $0.075. Split by type, tracking questions might come in far below that average and refund cases with long threads far above it. That split tells you where instructions or thread summaries will save the most.

Compare it with people. In May 2025 the median US customer service representative earned $21.53 an hour ([BLS](https://www.bls.gov/ooh/office-and-administrative-support/customer-service-representatives.htm)). Even at several tickets an hour, the wage alone is well above the AI cost in the example, before taxes, benefits or management time. The full comparison, including outsourcing, is in [AI customer support cost](/blog/ai-customer-support-cost).

## What spending mistakes do small businesses make with AI agents?

- **No ceiling.** Usage-based billing with only email alerts is how a quiet month turns into a large invoice.
- **One model for everything.** Sorting an email does not need the most expensive model.
- **Unbounded loops.** Agents without step limits can keep retrying a broken task.
- **Sharing everything.** Attaching your whole knowledge base to every task inflates every bill.
- **Reviewing everything twice.** Reviews are worth it on refunds and complaints, not on "thanks, got it" replies.
- **Ignoring the weekly view.** Cost per task drifts as your products, policies and thread lengths change.

## How should you set your first AI budget?

1. Count last month's tickets or tasks by type.
2. Start with a vendor plan that has a hard limit built in, not open-ended usage billing. For example, beereau's first team is $100 a month with the office included, and the Support team is sized for about 1,500 tickets a month.
3. Run two weeks with approvals on, so you see real costs per task type.
4. Adjust: move non-urgent categories to batch, tighten instructions for expensive ones.
5. Review cost per task weekly alongside your [customer support metrics](/blog/customer-support-metrics).

If you are planning beyond support, [running a company with AI agents](/blog/run-a-company-with-ai-agents) covers how budgets work across several teams, and [human in the loop AI](/blog/human-in-the-loop-ai) explains why approvals also act as a cost control: an agent that has to ask before a refund cannot commit your money on its own.

## Next step

Look at your last month of support email and mark which tickets truly needed an answer within the hour. Everything else is a candidate for batching. Then pick a monthly number you would be comfortable losing on a bad month, and make that your hard limit before you turn on any AI agent.

## Frequently asked questions

### Why are AI agent costs so hard to predict?

Because agents charge by the work they actually do, not by seat. A long email thread, a retry loop or an agent that rereads the same documents on every step can cost many times more than a simple reply. A hard monthly limit and a per-task cost view turn that uncertainty into a number you control.

### What is a hard spending limit for AI agents?

It is a monthly ceiling the system will not cross without your say. When spend reaches the limit, work pauses or continues only within a capped overage you approved. It is different from a soft alert, which only warns you while spending keeps going.

### What is economy mode for AI agents?

Economy mode is a lower-cost way of working that kicks in as you approach your budget. Urgent work continues as usual, while non-urgent tasks are grouped and processed in batches. In beereau it switches on automatically at 80% of the monthly limit.

### Does batching really make AI cheaper?

Often, yes. Several AI providers sell batch processing at a discount in exchange for slower turnaround. OpenAI and Anthropic both document a 50% discount for batch requests compared with standard requests.

### How do I know what one AI task costs?

Use a tool that records cost on every task, next to who handled it and who approved it. Then divide monthly spend by completed tasks of each type to get cost per task, and compare it with what the same task costs when a person does it.

### Is it cheaper to use the most powerful AI model for everything?

Usually not. The strongest models cost more per word processed, and many tasks, such as sorting emails or extracting an order number, do not need them. Matching the task to the right level of model is one of the biggest cost levers.

### What should happen when an AI agent hits its budget?

You should decide. A good setup gives you the choice to pause non-urgent work until next month or allow a capped overage, rather than silently stopping support or silently overspending.