How much does it cost to run an AI agent per month (for real)

30-second summary

• An AI agent has four cost components: model API, infrastructure, integrations, and technical maintenance.
• For simple WhatsApp lead qualification: $100–200/month in total.
• For agents with RAG or multiple integrations: $200–1,100/month.
• The most underestimated cost isn't the API — it's the specialist's time that keeps everything running.
• The math works when saved hours exceed the sum of all components.

The most common question after understanding what an AI agent actually does is: how much does it cost per month, for real, running in production? The honest answer is "it depends" — but it depends on concrete factors you can calculate before you start.

What goes into an AI agent's bill?

1. Language model API

The AI processes text in tokens (one token ≈ ¾ of a word in English). Every call to the model — when the agent reads a message and formulates a reply — consumes input and output tokens, and you pay for each.

Benchmark price (Jun 2026): frontier models like Claude Sonnet or GPT-4o cost between $0.003 and $0.015 per 1,000 output tokens. A full lead qualification conversation on WhatsApp (8 to 12 message exchanges) consumes roughly 2,000 to 4,000 tokens in total.

In practice: 1,000 qualification conversations per month = $15–50 just on the API, depending on the model and average message size.

2. Execution infrastructure

The code that orchestrates the agent — receives the message, calls the API, logs in the CRM, sends the reply — needs somewhere to run. The most common options:

• Serverless (AWS Lambda, Google Cloud Run): pay per execution, scales automatically. For low to medium volumes: $6–25/month.
• VPS / dedicated server: makes sense at high volume or when multiple agents run together. $30–100/month.
• Self-hosted n8n: if your company already runs n8n on a VPS, the incremental cost of adding an agent is nearly zero.

3. Vector database (only for agents with RAG)

If the agent uses RAG — looking up internal documents before answering — you need a vector store:

• pgvector on existing PostgreSQL: zero additional cost if you already have the database.
• Pinecone: free up to 1 million vectors; paid plans from $25/month.
• Weaviate or Qdrant cloud: $16–60/month depending on document volume.

Agents without RAG — fixed-question qualification flows, for example — don't need a vector store.

4. External integrations

This is where the bill most often surprises people who don't calculate upfront:

• WhatsApp Business API (Meta): billed per conversation (24-hour window). Customer service conversations: $0.02–0.09 each depending on country. At 1,000 conversations/month, that's $20–90.
• CRM: most platforms don't charge for webhooks or API calls — the extra cost is zero beyond the plan you already pay.
• n8n cloud / Make / Zapier: $20–100/month depending on execution volume. Self-hosted n8n is free beyond the VPS.
• Auxiliary database (logs, conversation history): $4–16/month on a managed service.

5. Technical maintenance

The most underestimated component. An agent in production is not "set it and forget it":

• Prompts need to evolve as unexpected edge cases appear.
• Integration APIs change — WhatsApp, CRMs, and media platforms update contracts and endpoints.
• Volume grows and the infrastructure needs to scale with it.
• Monitoring is non-negotiable: if the agent starts responding incorrectly, who catches it?

Realistic estimate: 2 to 8 hours/month from a specialist to keep an agent running reliably. Depending on the system's complexity, that represents $60–640/month.

How much does it cost in practice? Three scenarios

Simple agent: WhatsApp lead qualification (no RAG)

Volume: 500 leads/month · basic CRM integration

• Model API (Claude Haiku or similar): $12–24
• Serverless infrastructure: $10–20
• WhatsApp Business API: $20–50
• Maintenance (2h/month): $60–120
• Total: $102–214/month

Agent with RAG: internal assistant or knowledge-based support

Volume: 1,000 queries/month · indexed documents · no WhatsApp

• Model API (Claude Sonnet or similar): $30–56
• Infrastructure + vector store: $20–40
• Maintenance (4h/month): $120–240
• Total: $170–336/month

Complex agent: multiple integrations and high volume

Volume: 5,000 interactions/month · WhatsApp + CRM + automated reports · RAG

• Model API: $100–180
• Infrastructure: $50–100
• WhatsApp Business API: $120–240
• Orchestration (n8n/Make): $40–100
• Maintenance (8h/month): $240–480
• Total: $550–1,100/month

What drives costs up unnecessarily?

Three mistakes that tend to double the bill:

Unoptimized prompts. A poorly written prompt sends unnecessary context on every call — and you pay for every token. Lean prompts cost less and perform better.

No prompt caching. When the same chunk of context (a manual, an FAQ, business rules) is resent on every call, prompt caching on models that support it reduces input costs by up to 90%. Ignoring it means paying for the same tokens over and over.

Wrong model for the task. Using Claude Opus or GPT-4o to classify "yes or no" costs 10x more than a smaller model with the same result. The rule: lightweight model for routing and triage, powerful model only where reasoning actually matters.

When does the agent pay for itself?

The math is straightforward: add up what you currently spend on the process the agent will replace — team hours, leads that go cold because of slow replies, manual CRM logging errors. If that number exceeds the agent's monthly cost, it pays for itself.

The break-even point typically appears in the second or third month, once the initial tuning stabilizes the agent's behavior. It's not Day 1 — but it's not a two-year bet either.

Example: a team spending 60 hours/month on manual lead qualification (at $15/hour) has a $900 labor cost for that task alone. An agent that does the same for $200/month frees up $700/month — before counting the extra conversions from leads that used to go cold.

area next handles this calculation before any project: estimated cost, expected return, payback timeline. Request a diagnosis.

Frequently asked questions