The short answer
To build one: our AI Agent engagement starts from $8,000. A written proposal fixes the price of each milestone before work starts, and nothing is billed before you accept it. The work starts with a deposit of 30% of the price, credited against the milestones.
To run it: the AI model's usage and, on WhatsApp, Meta's messaging fees. Both are billed by those providers to accounts in your name, so you see every charge and none of it passes through us.
If the job isn't clear yet, a Product Sprint from $5,000 turns it into a written spec, a prototype and a priced plan before anyone commits to the build.
What the build pays for
An agent that talks to your customers is more than a prompt. The price covers five pieces of work:
- The agent, with tools wired to your systems: your catalog, your calendar, your CRM or your orders.
- Retrieval over your documents, so every answer has a source you can check.
- An eval suite: graded test cases from your real conversations, with a pass threshold, run on every change.
- A second AI provider that takes over when the first one fails, so an outage doesn't stop your customers.
- A console where your team reviews conversations and takes over the ones the agent hands off.
The evals are what tell you the agent is ready. On this site, the AI that sorts incoming briefs has 31 graded cases and stays off until it passes 90% of them. Your agent gets the same rule, with a threshold we agree with you.
What moves the price
Two agents on the same channel can cost very different amounts. These questions decide it, roughly in the order they matter:
| Question | It costs less when | It costs more when |
|---|---|---|
| How many jobs? | It does one job: answering, booking or qualifying leads. | It does several, each with its own tools and evals. |
| What does it touch? | It reads from one clean source, like a product catalog. | It writes into your systems: orders, bookings, refunds. |
| Where does it talk? | On one channel, WhatsApp or your website. | On several that must give the same answers. |
| What does it know? | Your answers already live in a catalog or a help center. | They're spread over PDFs, old chats and people's heads. |
| Which languages? | One. | Several, each with its own test cases. |
| What can go wrong? | A wrong answer costs little. | Money, health or legal topics, which need stricter rules and more cases. |
Each of these adds work we can name in the proposal, so the price of every milestone says what it pays for.
What it costs each month
Once it's live, an agent has three running costs. None of them goes through us:
- Model usage. AI providers charge for the text the model reads and writes, so the cost grows with the number of conversations and their length. Anthropic's pricing has the current rates for Claude.
- Channel fees. On WhatsApp, Meta bills messages by type and by country, and a reply to a customer who wrote to you is billed differently from a message you start. Its pricing page has the rates.
- Hosting. The agent's server and database, in your own cloud account.
We measure cost from the pilot on. Every model call is logged with what it cost, the way this site's lead triage stores the cost of each run, so you know the cost per conversation before the agent is open to everyone.
If you want us to keep improving it after launch (new cases, model updates re-run against the evals, monitoring), that's a Retainer from $3,500 a month. It's optional: the code, the accounts and the docs are yours either way.
Questions to ask before you hire anyone
Whoever builds it, these answers tell you what a quote includes:
- 01How will we know it's ready? Ask for graded test cases and a pass threshold, not a demo.
- 02What does it never do, and when does it hand off to a person? Both should be written down before the build.
- 03What happens when the AI provider is down?
- 04Who owns the code, the prompts and the accounts at the end?
- 05Which costs are billed through the vendor, and which go to your own accounts?
- 06Is the price fixed per milestone, or billed by the hour?
When an AI agent isn't worth it
Sometimes the honest answer is a smaller tool. An agent rarely pays off when:
- Few people write to you, and your team answers them without strain.
- The answers change every day and nobody keeps a source for them.
- A clear page, a form or a booking link would solve it.
We'd rather say so on the first call than build something you don't need. If it does fit, send us a brief and a person replies with questions or a first plan.