AI in Hiring · 9 min read

What Does an AI Agent Cost? One-Off Costs, Running Costs and a Worked Example

What does an AI agent cost? One-off and running costs explained, with a worked example you can fill in, cost levers, hidden costs and build vs buy vs SaaS.

Door Ingmar van Maurik · Founder & CEO, Making Moves


Short answer

The cost of an AI agent has two parts: one-off costs to build it (analysis, design, integrations, an evaluation set, a security review and documentation) and running costs to operate it (model usage, hosting, monitoring, maintenance and human review). Model usage is billed per token. It grows with the number of conversations, the length of your prompt and the number of steps per task. A simple formula lets you estimate it in advance. In many setups, integrations and human review end up weighing more than the model costs themselves.

First: what do we mean by an AI agent?

An AI agent is a system that carries out a task in several steps: it reads input, chooses an action, calls a tool (your ATS, CRM or calendar, for example), looks at the result and continues until the task is done. Those steps drive a large share of the cost. A chatbot answering one question makes one model call. An agent that screens a candidate, schedules an interview and writes a summary easily makes six or more. If you want the basics first, read what AI agents are.

One-off costs: what you pay to build it

ComponentWhat it involvesWhat drives the cost

|---|---|---|

AnalysisMapping the process, baseline, setting a goalHow well the process is already documented DesignSteps, tools, permissions, human approvalNumber of decision points and exceptions IntegrationsConnections to ATS, CRM, email, calendar or databaseAPI quality, authentication, test environments Evaluation setReal, labelled examples to measure qualityHow many situations you want to cover and who labels Security reviewPrompt injection, data leaks, least privilegeSensitivity of the data and number of tools Documentation and handoverExplanations, runbooks, training your teamWhether you want to run the system yourself

Integrations are usually the biggest uncertainty. A clean, documented API connects quickly. A legacy system without a test environment takes far more time than the model.

An indication, and we stress indication: a tightly scoped pilot with one or two integrations is often in the range of a few thousand euros, while an agent with several integrations, strict privacy requirements and a thorough evaluation set can easily run into tens of thousands of euros or more. The real price depends entirely on scope. So always ask for the breakdown, not just the total.

Running costs: what it costs per month

  • Model usage. You pay per token, usually at a different price for input (what you send to the model) and output (what it returns). Costs scale with volume, prompt length and steps per task.
  • Hosting. Servers, databases, storage for documents and logs.
  • Monitoring. Tracing of every step, dashboards for quality and cost, alerts.
  • Maintenance and model updates. Models get replaced or change behaviour. Every switch needs a new evaluation round. The APIs of your own systems change too.
  • Human review. Staff time spent on borderline cases, spot checks or approving actions.
  • Worked example: calculate it yourself

    The formula for monthly model costs:

    Model cost = conversations × steps × (input tokens × input price + output tokens × output price) / 1,000,000

    Prices are per million tokens. All numbers below are hypothetical. They are not real vendor prices. Replace them with your own volumes and the current prices of the model you choose.

    Assumption (example)Value

    |---|---|

    Conversations per month5,000 Steps per conversation6 Input tokens per step3,000 Output tokens per step300 Example input price€2 per million tokens Example output price€8 per million tokens

    Worked out:

  • Input: 5,000 × 6 × 3,000 = 90 million tokens × €2 = €180
  • Output: 5,000 × 6 × 300 = 9 million tokens × €8 = €72
  • Total: €252 per month, roughly €0.05 per conversation
  • Now assume 70% of the input is a fixed instruction you reuse with prompt caching, and that cached tokens cost 10% of the normal price in this example. Input then costs 63 million × €0.20 + 27 million × €2 = €66.60. The total drops to €138.60.

    Two things stand out. First: if retries push your agent to an average of 9 steps instead of 6, model costs rise by half. Second: if a person reviews 5% of conversations at 4 minutes each, at an assumed hourly cost of €40, that is 250 × 4 minutes = just over 16 hours, about €667 per month. In this example, human review costs more than the model.

    How to keep costs down

  • Pick the right model per step. A small, cheap model for routing and classification, a larger one only for steps that need real reasoning.
  • Use prompt caching for fixed instructions, tool descriptions and standing documents.
  • Keep context short. Send only relevant passages, not the whole file.
  • Cap steps and retries. A maximum number of steps per task and of retries stops a stuck agent from running up costs.
  • Batch what can wait. Overnight processing of CVs, for example, can often be done more cheaply through batch processing.
  • Set cost ceilings and alerts. Per day, per user and per task, so a bug does not produce a surprise bill.
  • Hidden costs

  • Failures. An agent that books the wrong meeting or wrongly rejects a candidate costs time and trust.
  • Rework. Staff correcting output, handling exceptions or resolving complaints. That time rarely appears in a quote.
  • Compliance. GDPR, data processing agreements, logging and explainability. AI used for recruitment and selection is classed as high risk under the EU AI Act, with requirements for documentation, human oversight and monitoring.
  • Build, buy or SaaS?

    OptionAdvantagesWatch out for

    |---|---|---|

    Build in houseFull control, ownership of code and dataYou need people who can both build and run it Have a partner build itFast start, experience with pitfallsAgree on ownership, handover and maintenance SaaS productLive quickly, predictable subscriptionLess customisation, per user or volume pricing, lock-in risk

    Compare not just the starting price but the total cost over two to three years, including operations and switching costs. How that works for recruitment software is covered in the true cost of ATS software. For recruitment, we build an AI hiring system that fits your existing processes.

    Is your consultant's estimate realistic?

    A realistic estimate has a few features:

  • One-off costs are broken down by component, with integrations listed separately.
  • It includes a calculation of running costs, with tokens per task, volume and the assumptions behind them.
  • There is budget for an evaluation set and a security review.
  • Maintenance and model updates are included, not just the build.
  • The consultant names the biggest uncertainty and how to reduce it, for instance with a pilot.
  • If any of that is missing, ask more questions. You will find more in how to evaluate AI consultants.

    Frequently asked questions

    What does an AI agent cost per month?

    It depends on volume, prompt length, number of steps and the chosen model, plus hosting, monitoring and human review. Work it out with the formula above and your own assumptions.

    What does it cost to have an AI agent built?

    Mainly the number of integrations, the sensitivity of the data and how thorough the evaluation needs to be. Ask for a breakdown per component rather than a single total.

    Are model costs the biggest item?

    Not always. For many agents, integrations, maintenance and human review weigh more. The worked example shows how that can play out.

    How do I avoid a surprise bill?

    Set cost ceilings and alerts, cap steps and retries, and monitor cost per task from day one.

    Summary

  • An AI agent has one-off and running costs. Look at both.
  • Model costs scale with volume, prompt length and steps. Calculate them with your own assumptions.
  • Keep costs down with model choice per step, caching, short context, limits and alerts.
  • Do not forget failures, rework and compliance.
  • A good quote is broken down and states its assumptions.
  • Want to know whether the people building or assessing your AI agents really have these skills? Take a look at the AI Skills Assessment or get in touch.


    Book intake call · View our AI Hiring System