"How much will it cost?" is always the first question, and it should be. I'd rather explain where the money goes than give a number that means nothing without context. Here's how I break it down for clients.
The three parts of the cost
1. Building it (one time)
This is design and development: connecting your systems, writing the agent's logic and rules, building the approval flow, testing with your real data, and launch. It depends on:
- How many systems it connects to. Gmail only, or Gmail plus your order system plus WhatsApp?
- How much it is allowed to do on its own. Drafting replies is simpler than issuing refunds.
- Whether it needs a user interface, such as a dashboard to review and approve.
- How messy the data is. Clean, written answers are quick to use. Knowledge that lives only in someone's head has to be written down first.
2. AI model usage (monthly)
You pay providers like Anthropic (Claude) or OpenAI per token, which roughly means per piece of text read and written. For typical business tasks such as classifying emails, drafting replies or answering chatbot questions, this is usually the smallest part of the bill.
Ways I keep it low:
- A small, cheap model for simple steps like classification and routing.
- A stronger model only where quality matters, like replies customers read.
- Caching answers to repeated questions.
- Monthly spending limits set at the provider.
3. Hosting and maintenance (monthly)
A small server or serverless functions, a database, monitoring, and someone keeping an eye on it: updating content, adjusting rules, and handling changes when a provider updates its API.
Estimate the return first
Before building anything, I ask clients to estimate one number: hours per month spent on this task today.
Then a simple sum: hours saved × what an hour costs you, compared with the build cost plus the monthly running cost. Add the harder-to-measure wins, such as faster replies, fewer missed leads, staff doing better work, and the decision usually becomes obvious, one way or the other.
If the task takes two hours a month, automation is probably not worth it. If it takes two hours a day, it almost certainly is.
Ways to start small
- Draft-only mode first: the AI drafts, a person sends. Lower risk, faster to build.
- One workflow first, the most repetitive one, then expand.
- Use what you already have. Your existing inbox, sheets and CRM, not a new platform.
What I won't do
I won't quote a price before understanding the workflow, and I won't recommend AI where a simple rule or a better form would do the job. Sometimes the cheapest automation is no AI at all.
If you want a real estimate, describe the task you'd like to automate: what happens today, how often, and which tools you use. Read more about my AI automation service.
