How Much Does an AI Agent Cost to Build and Run?

What an AI agent costs to build

An AI agent costs $120,595 to build on average, according to Clutch's pricing data for AI developers. That average covers agents, chatbots, machine learning and language projects reviewed on Clutch. The most common band is $10,000 to $49,999, the average monthly spend is $11,553 (the highest of any Clutch development category) and the typical engagement runs about 10 months.

AI work also carries a rate premium. Axios, reporting Upwork data, found that freelancers doing AI work earn more than 40% more per hour than those doing non-AI work. That premium shows up in agency quotes too.

As with any average, a few large programs pull the number up. A focused agent that answers customer questions or books appointments usually sits in the lower bands. An agent that reads and writes to several business systems sits higher.

AI agent price bands from published agency packages

Several AI agencies publish their price bands, which makes the market easier to read. The ranges below are quoted from vendor pages. They are industry anchors, not Arcesso prices, and each row names the publisher.

The jump between rows tracks two things: how many systems the agent touches and how much it is trusted to do without a person checking. A bot that answers from a help center is cheap. An agent that changes orders in your ERP needs permissions, audit logs and testing.

What it costs to run an AI agent each month

Running an AI agent costs money every month, separate from the build. Buyers often miss this line because a demo runs on free credits. Production traffic does not.

Clutch's $11,553 average monthly spend is what clients pay the agency during an engagement. It is not the cost of running the agent. Your run cost is the bill from the platforms the agent uses, and it scales with volume.

• Model usage: OpenAI, Anthropic Claude and Google Gemini bill per token, at rates listed on their pricing pages. Larger models cost more per token.
• Voice minutes: phone agents pay per minute for telephony (for example Twilio) and for the voice platform (Vapi, Retell, ElevenLabs).
• Hosting and data: app hosting, a database and a vector store on platforms such as Vercel, Supabase or AWS.
• Monitoring and evals: logs, tracing and scheduled test runs that catch wrong answers before customers do.
• People: the staff time to review flagged conversations and approve actions.

How to estimate token and API costs before you build

Token cost is easy to estimate with four numbers: conversations per month, turns per conversation, tokens per turn and the model's published price. Multiply the first three to get monthly tokens, then apply the price from the provider's pricing page.

For example, 2,000 conversations a month with 6 turns each and 1,500 tokens per turn comes to 18 million tokens. Retrieval adds tokens because each answer carries pieces of your documents into the prompt. Tool calls add more, since the agent reads the tool result before replying.

Four levers keep the bill down. Route simple questions to a smaller, cheaper model and save the large model for hard ones. Cache repeated instructions. Limit how many document chunks go into each prompt. Set hard spending caps in each provider's dashboard so a loop or an abuse spike cannot run up a surprise invoice.

Build cost versus run cost for an AI agent

Build cost is paid once and run cost is paid every month, so compare them over a year, not at launch. A cheaper build that uses a large model for every message can cost more in year one than a careful build that routes traffic well.

What moves the price of an AI agent

Integrations and the accuracy bar move the price of an AI agent most. The model itself is rarely the expensive part of the build. The work is in connecting it safely to your data and proving it behaves.

• Number of tools: each system the agent can read or change (CRM, calendar, ERP, ticketing) adds integration and permission work.
• Data readiness: clean, current documents are cheap to search. Scattered PDFs and old wikis need cleanup first.
• Channel: web chat is simplest. WhatsApp and phone add channel rules, latency work and per-message or per-minute fees.
• Accuracy bar and evals: a support bot can tolerate a handoff. An agent that issues refunds needs a test suite and approvals.
• Human in the loop: review queues and approval steps add screens and logic.
• Compliance: personal or health data adds redaction, access controls and audit logs.
• Languages and volume: more languages mean more tests, and high volume justifies model routing work.

Why voice agents and AI receptionists cost more to run

Voice agents cost more to run than chat agents because every minute of a call is billed by several services at once. A typical call pays for telephony, speech-to-text, the language model and text-to-speech, often from different providers.

Arcesso builds voice agents on Vapi, Retell, LiveKit, ElevenLabs and Twilio, and already markets AI voice receptionists and appointment booking on arcesso.ai. The design choices that cut cost are the same each time: keep calls short with clear prompts, hand off to a person early when the caller needs one, and use a smaller model for routing.

Demand is growing fast. MarketsandMarkets projects the AI agents market to grow from $7.84 billion to $52.62 billion by 2030, a 46.3% annual growth rate. Fiverr reported searches for AI voice agent specialists up 49% between late 2025 and early 2026.

How to keep AI agent costs under control

The safest way to control AI agent cost is to prove one workflow before funding many. Pick the task with the most volume and the clearest success measure, such as booked appointments or tickets resolved without a person.

Build that one workflow with real data, an eval set of real questions and a spending cap. Measure accuracy, handoff rate and cost per conversation for a few weeks. Then decide whether to expand, change models or stop.

Arcesso runs discovery at a fixed price and credits it in full to the build. You own 100% of the code, prompts and eval sets from the first commit, in writing, so you can switch models or vendors later without losing the work.

When an AI agent is not worth building

An AI agent is not worth building when a simple rule or workflow does the job. If every case follows the same steps, an n8n or Make workflow without a model is cheaper to run and easier to trust.

Agents also struggle when volume is low, because the build cost never pays back, or when the source data is out of date, because the agent will repeat wrong answers with confidence. Fix the data first.

We will tell you if we are not the right fit. If your use case is a standard FAQ bot that an off-the-shelf help desk tool already offers, buy the tool. Custom agents earn their cost when they act inside your own systems.