What is an AI Worker?

An AI Worker is an autonomous AI agent that works like a real teammate: it has a personality, specialized skills and memory, runs tasks end to end and integrates with your tools (Slack, Teams, Gmail, Notion…).

How many AI Workers are available?

Specialized AI Workers are available for every team, covering marketing, development, design, data, human resources, finance and customer support.

Can I try it for free?

Yes, a free trial is included with no commitment. You can add or remove AI Workers at any time and cancel whenever you want.

Yes. Your data is encrypted end to end, hosted in isolated environments, and never used to train models. Workers Delos is GDPR-compliant and SOC 2-certified.

What Does an AI Agent Run Really Cost? Five Budget Controls Before You Scale

By Delos — 2026-11-11

Cost is one of the three reasons Gartner gives for agentic projects being cancelled. Here is how to measure and cap the cost of one agent run.

A demo agent costs almost nothing. The same agent running 5,000 times a day on real data is a line in the CFO's budget review. Gartner forecasts that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing rising costs, unclear business value and insufficient risk controls (Anushree Verma, Senior Director Analyst at Gartner, quoted by IT Brief, June 2025). Cost is therefore a design question, not an invoice you read afterwards. This guide gives you a way to measure and cap it.

Why agent costs surprise teams

A chatbot answers once. An agent loops: it plans, calls a tool, reads the result, decides again. Each pass sends the growing conversation back to the model. Two runs of the same task can differ widely in cost because the path is not fixed. That variance is the point of an agent, and also its budget risk.

What a run really costs

Think of one run as four multiplied or added drivers (see chart): the number of steps, the tokens per step, the price of the model, and the cost of tools and human review.

Steps per run

Loops and retries are the biggest multiplier. An agent that retries a failing tool call five times pays for five calls and produces nothing.

Tokens per step

Context grows at every step: instructions, history, tool results. Long tool outputs pasted back in full are a classic waste. Return only the fields the agent needs. Our guide to agent memory architecture explains how to structure what an agent keeps and retrieves.

Model price

Not every step needs the largest model. Classifying an email or extracting a date can go to a smaller, cheaper model. Reserve the large one for planning and hard judgment.

Tools and review

External API calls are billed separately, and a human checking the output has a real cost in minutes. Include both in the number, or the business case is flattering.

!Four drivers of AI agent cost per run: steps, tokens per step, model price, and tool and human review costs

Five controls to put in place before you scale

!Five budget controls for AI agents: measure real runs, cap steps and spend, route cheaper models, assign a daily budget owner, and compare cost with task value

1. Measure cost per run on real data. Run 50 to 100 real cases, including messy ones, and record the median and the worst run. The worst run is your budget risk.

2. Set a step limit and a spend cap per run. When the cap is hit, the agent stops and hands over to a person. No silent overruns.

3. Route simple steps to a cheaper model. This is usually the quickest saving and needs no change to the business logic.

4. Alert on a daily budget, with a named owner. Someone must receive the alert and be allowed to pause the agent.

5. Compare cost per run with the value of the task. If a run costs more than the task is worth, fix the design before adding volume.

A worked example (illustrative numbers, our assumption)

Suppose an agent that triages supplier invoices uses 8 steps and the team measures a median of 0.12 EUR per run and a worst case of 1.10 EUR. At 2,000 runs a day, the median means about 240 EUR a day; if one run in twenty hits the worst case, the bill rises noticeably. The cap in control 2 turns that unknown into a ceiling. These figures are invented to show the arithmetic, replace them with your measurements.

Start with the simplest design

Anthropic's guidance on building agents recommends finding the simplest solution first and adding agentic complexity only when it demonstrably pays off. A fixed workflow is cheaper and more predictable than an open loop; use an agent where the path truly cannot be known in advance. See our comparison of single-agent and multi-agent architectures before adding another layer of orchestration.

How Delos Workers help

Delos Workers run with defined scope, permissions and a human checkpoint, so a cap and an owner are part of the setup rather than an afterthought. Start with one worker, one task, measured for two weeks. Pair the budget controls with least-privilege access for agents to keep the scope of each run explicit.

Conclusion: measure the worst run, cap it, route cheap steps to cheap models, name an owner, and compare to value. Do these five before you scale.