Short answer: a single AI-powered automation can cost a few hundred dollars to build, a document-answering assistant a few thousand, and a custom AI agent built into your own systems more than that. That's the one-time build. The running cost is separate. You pay the AI model provider per use, plus any automation platform or hosting you use. For many small business workloads, that's tens of dollars a month, not thousands. What moves the price most is how many systems the agent connects to and how much it's allowed to do on its own.
This guide breaks the cost into its parts, so you can sense-check any quote you get, including ours.
Key points
- Two separate bills: what you pay once to build the agent, and what you pay every month to run it.
- Scope drives build cost. Each extra app connection, each decision the AI makes and each approval step adds work.
- Running cost is mostly AI model usage, billed per token (small chunks of text). Prices differ by about 100x between the smallest and largest models.
- Pick the simplest thing that works. A fixed automation with one AI step costs far less to build and run than a fully autonomous agent.
- Budget for upkeep. Apps change their APIs, models get retired and your processes change.
What you're actually paying for
1. The build (one-time)
This is the design, building, testing and handover work. It depends on:
| Cost driver | Why it matters |
|---|---|
| Number of app connections | Every system the agent reads from or writes to (email, CRM, accounting, Slack, your database) needs authentication, data mapping and error handling. |
| How much the AI decides | One AI step inside a fixed workflow is simple. An agent that chooses its own next steps needs more guardrails and far more testing. |
| Your data | Answering from 20 clean PDFs is a different job from answering from 10 years of mixed documents with tables and scans. |
| Approvals and safety | Human sign-off before sending emails or moving money, audit logs and limits on what the agent can touch all take time to build. They're worth it. |
| Where it lives | A workflow in n8n or Make is quicker to build than a feature inside your own app, which also has to fit your codebase, tests and release process. |
| Testing (evaluation) | A reliable agent is tested against real example cases before launch. Skipping this is how pilots fail in production. |
2. Running costs (monthly)
AI model usage. Model providers bill per million tokens, with separate prices for input (what you send) and output (what the model writes). On the official price lists we checked in October 2026, the gap between models is large:
| Model (example) | Input, per 1M tokens | Output, per 1M tokens |
|---|---|---|
| Claude Haiku 5.5 (small, fast) | $0.10 | $0.50 |
| Claude Sonnet 5.5 (mid-size) | $2 | $10 |
| Claude Opus 5.5 (large) | $4 | $20 |
| OpenAI gpt-6-luna (small) | $0.10 | $0.50 |
| OpenAI gpt-6-sol (mid-size) | $2 | $10 |
Sources: Anthropic pricing and OpenAI API pricing, checked 13 October 2026. Prices change often, so check the source before you budget.
Two discounts are worth knowing. Prompt caching makes repeated input, such as the same instructions or documents sent on every request, much cheaper. Batch processing gives 50% off for work that doesn't need an instant answer, on Anthropic's Batch API for example.
Automation platform. If the agent runs on a workflow tool, that has its own plan. n8n's hosted plans start at €20 a month billed annually, and its Community Edition is free to self-host on your own server (n8n pricing). Make has a free plan and paid plans from $12 a month for 10,000 credits (Make pricing).
Hosting and storage. A custom agent or a document assistant needs somewhere to run and somewhere to keep its search index. For small workloads, this is often a modest cloud server.
Upkeep. Somebody has to fix things when a connected app changes its API, when the provider retires a model or when your process changes. Plan for some hours each month or a support arrangement.
A worked example: monthly model cost
This is a sample calculation, not a client case. It shows how the arithmetic works.
Say an internal assistant answers 1,000 questions a month. Each question sends about 3,000 tokens to the model (instructions, the question and some matching passages from your documents). Each answer is about 500 tokens. That's 3 million input tokens and 0.5 million output tokens a month.
| Model | Input cost | Output cost | Monthly total |
|---|---|---|---|
| Claude Haiku 5.5 | 3 × $0.10 = $0.30 | 0.5 × $0.50 = $0.25 | about $0.55 |
| Claude Sonnet 5.5 | 3 × $2 = $6 | 0.5 × $10 = $5 | about $11 |
| Claude Opus 5.5 | 3 × $4 = $12 | 0.5 × $20 = $10 | about $22 |
Three things push this number up in real life:
- Agents take several steps. An agent that plans, calls a tool, reads the result and then answers sends the growing conversation back to the model at every step, so tokens multiply with the number of steps.
- Long context. Pasting whole documents into every request costs far more than retrieving only the relevant passages, which is what RAG does.
- Volume. A customer-facing assistant on a busy website can see many times the volume of an internal one.
The point: for a typical small business assistant, the model bill is usually not the big number. The build, and keeping it working, are.
Typical price ranges by type of project
Here are our own published packages, as a reference point. They're in USD and exclude your tool and model accounts, which are billed to you directly by the provider (full details):
| What you need | Our package | Price |
|---|---|---|
| Automate one repetitive task: 1 workflow, up to 3 apps, 1 AI step | Quick Win | $490 |
| Automate a whole process: up to 3 workflows, up to 5 apps, error alerts | Workflow Pro | $1,200 |
| An assistant that answers from your documents, or a simple agent using 2 tools, on your website or in Slack | AI Assistant | from $2,500 |
| Add 1 AI feature to your existing app in a 2-week sprint | AI Integration Sprint | from $3,000 |
A fully custom AI agent that works across many systems, makes decisions on its own and needs approval steps and audit logs is bigger than these packages. That's quoted after a scoping call, because the cost depends on the drivers in the first table.
Quotes from other providers will vary widely. Compare what's included rather than the headline number.
How to keep the cost down
- Start with one process. Pick one task that is repetitive, rule-heavy and annoying. Prove it works, then expand.
- Use the simplest design that works. Anthropic's own guidance on building agents recommends starting simple and adding agent behavior only when it's needed, because agentic systems "often trade latency and cost for better task performance." Not sure which you need? See AI agent vs chatbot vs automation.
- Match the model to the task. Small models handle sorting, extracting and short answers well. Save the large models for steps that need real reasoning.
- Own your accounts. Have model and platform accounts in your company's name, so you see the real running cost and keep control if you change providers.
- Agree the scope in writing. A fixed scope with a clear list of what's included protects both sides from surprise bills.
Questions to ask before you sign
- What exactly is included, and what would count as extra work?
- Who owns the code, the workflows and the prompts when the project ends?
- Whose accounts do the model and platform bills go to?
- How will you test it before launch, and what happens if it gets something wrong?
- What does support cost after handover?
FAQ
How much does it cost to run an AI agent per month?
It depends on volume, the model and how many steps the agent takes. In our sample calculation, 1,000 questions a month cost between about $0.55 and $22 in model fees, depending on the model. Add any automation platform plan, hosting and upkeep on top.
Is it cheaper to use ChatGPT Business or Claude seats instead of building an agent?
For staff who need a general assistant, seat plans are often the cheapest option. For example, OpenAI lists ChatGPT Business from $20 per user a month billed annually (OpenAI business pricing). A built agent makes sense when the work needs to run automatically, connect to your systems or serve your customers.
Why do some AI projects say "from" instead of a fixed price?
Because the cost depends on things that vary between businesses: how many documents you have and what state they're in, which tools need connecting and what your existing software looks like. A fair provider confirms a fixed price in writing after seeing those.
Can I pay for an AI project in AED?
For our own work, yes. Our website shows prices in USD, and clients in the UAE get their quote and invoice in AED.
Want a number for your project?
Tell us what you'd like to automate and which tools you use today. In a free 30-minute call, we'll suggest the simplest way to do it and follow up with a written scope and price. You can also see what we build on our AI agents page.