How much does an AI agent cost?
Quotes for what sounds like the same agent can differ by a factor of ten. The gap usually comes down to what the agent is connected to and how much it is allowed to do. Here is how the costs break down.
The short answer
Plan for three costs: setup, running costs, and the time your team spends supervising. A ready-made platform can start around a few hundred dollars a month. A custom agent connected to your systems usually starts at several thousand dollars and rises with its integrations and permissions.
For a typical customer conversation, AI model usage is often only cents. People and integrations cost more than the AI itself.
Typical price ranges
| Option | Published price | Fits when |
|---|---|---|
| Business AI assistant | About USD 20 to 25 per seat a month | Staff need help, and a person checks every result |
| Ready-made agent platform | One UAE vendor lists AED 547 to 1,098 a month, before Meta fees, hosting and setup | Standard support or sales chat with few integrations |
| Custom agent | One development company’s guide runs from USD 20,000 for a simple agent to over USD 200,000 | It must act in your own systems or follow your own rules |
| keyt | A first AI agent falls under Build, from USD 6,000, with a fixed written quote | You want a scoped first agent you own |
The market figures come from vendors with something to sell, so treat them as rough benchmarks and compare written quotes for the same scope.
What moves the price
- Integrations: each system the agent reads from or writes to, especially older ones without an API
- Actions: replying is cheap; changing records, sending quotes or moving money needs approvals and logs
- Knowledge: how much content there is, and how messy it is
- Languages: Arabic dialects and mixed Arabic and English need their own testing
- Volume and uptime: what happens at peak, and who responds when something fails
- Rules: data location, audit trails and regulated sectors
A worked example of running costs
AI models charge per token, roughly a piece of a word. At a list price of USD 2 per million input tokens and USD 10 per million output tokens, take a ten-message conversation where each turn sends about 3,000 tokens of instructions and history and returns about 150 tokens.
| Item | Calculation | Cost |
|---|---|---|
| Input | 30,000 tokens × USD 2 per million | USD 0.06 |
| Output | 1,500 tokens × USD 10 per million | USD 0.015 |
| One conversation | About USD 0.08 | |
| 2,000 conversations a month | About USD 150 |
This is an illustration: real prompts, models and prompt caching change the result, and agents that make many tool calls use more. Measure your own usage in the first weeks.
On WhatsApp, Meta charges per delivered message by category and country. Since 1 October 2026, service replies are charged too, after the first 1,000 a month per business number. Our cost calculator estimates this with editable rates.
How to judge the return
Monthly saving = hours saved × loaded hourly cost, plus extra sales from faster replies, minus running costs and supervision time. Use your own numbers from the current process, and be careful with the sales figure: count only what you can trace.
An agent is rarely worth it at low volume. If a person handles a few dozen enquiries a week without trouble, improve the process first.
Free tools for this
Sources
- Anthropic: model pricing
- SleekFlow: AI agent cost in the UAE (a platform vendor’s figures)
- Cleveroad: AI agent development cost (a development company’s figures)
- Meta: pricing on the WhatsApp Business Platform
Rules change. Check the official text, or ask your accountant or lawyer, before you rely on a figure.
Get a fixed quote for your first agent
Tell us the task and the systems involved. We will scope it, quote it in writing, and tell you if a cheaper tool would do the job.