How Much Does It Cost to Build an AI Chatbot for Customer Support?

A couple of years ago an AI support assistant was a nice demo. Now plenty of companies have one answering real customers. So the question we get these days isn't whether to add one. It's whether to buy an off-the-shelf tool or build something custom, and what each one really costs once it's handling thousands of conversations a month.

Here are both options with real numbers, including how to work out the monthly AI bill yourself.

Buying vs building

Buying means turning on the AI agent inside a help desk you might already use, like Intercom, Zendesk or Freshdesk. It reads your help centre and answers customers. You can be live within days.

Building means a custom assistant on top of models from Anthropic, OpenAI or Google, connected to your own docs, product data and systems. It takes longer to set up, but you control how it behaves, where the data lives and what it costs.

Neither is always right. It mostly comes down to three things: how many conversations you handle, whether the assistant only needs to answer questions or also needs to do things, and how much control you need over your data.

What buying costs

Most off-the-shelf AI agents charge per resolved conversation, per seat, or both. Intercom's Fin, one of the best known, charges $0.99 per resolution at the time of writing. That's on top of the help desk seats you're already paying for.

At low volume that's great value. Resolve 300 conversations a month and you pay about $300, with almost no engineering work.

The maths changes as you grow, because the bill goes up exactly as the bot gets more useful:

You're also limited to what the platform supports. If you want the assistant to check an order in your own database, change a booking or follow your own refund rules, you'll hit that wall fairly quickly.

What building costs

A custom build costs more up front and much less to run. Typical ranges for an agency build this year:

Those numbers assume an India-based team. US and UK agencies usually quote two to three times more.

Working out the monthly AI bill

This is the part most teams are unsure about, so let's go through an example.

Say a typical support chat has 6 back-and-forth messages. Each time, the model receives about 3,000 tokens (your instructions, the relevant help articles and the conversation so far) and writes back about 300.

Across the whole chat that's roughly 18,000 tokens in and 1,800 tokens out.

Now use a fast, cheap model. Claude Haiku 4.5 costs $1 per million input tokens and $5 per million output tokens at the time of writing:

• Input: 18,000 × $1 per million = $0.018
• Output: 1,800 × $5 per million = $0.009
• Total: about 3 cents per conversation

At 5,000 conversations a month, that comes to roughly $135 in model costs. Prompt caching, which reuses the parts of each request that don't change at a much lower price, can bring it down further.

The model bill isn't everything, of course. You'll also pay for hosting, a vector database for search, monitoring, and a few hours of maintenance each month. For most teams that adds up to a few hundred dollars a month at this volume. Compare that with roughly $3,000 to $5,000 a month on per-resolution pricing.

So which should you choose?

Our rough rule:

• Buy if you get fewer than about 1,000 conversations a month that a bot could resolve, your answers already live in a help centre, and you want something running this week.
• Build if you're heading past a few thousand conversations a month, the assistant needs to take actions in your own systems, or there's data you can't send to a third-party help desk.

Plenty of companies do both, one after the other. They start with an off-the-shelf tool to learn what customers actually ask, then build their own once the volume makes it worth it.

What makes a support bot good

The model is honestly the easy part. Quality comes from everything around it.

It needs to answer from your own content. That means pulling the right help articles or data for each question (this is what people mean by RAG) and showing where each answer came from.

You need to test it properly before launch. Collect a few hundred real customer questions with the right answers and measure the bot against them. Run that test again every time you change the prompt, the model or the content.

It has to know when to hand over. If a customer is frustrated, the topic is sensitive, or the bot simply doesn't know, it should pass the chat to a person along with the full conversation.

If it can take actions, set limits. Refunds above a certain amount should need a human to approve them, and every action should be logged.

And keep the content fresh. A great bot reading out-of-date docs will give confident, wrong answers. Connect it to wherever your docs actually live and re-index automatically.

Data and privacy

If customers share personal or account details in chat, check three things.

First, training. The enterprise API plans from the big model providers don't train on your data by default, but get that confirmed in writing for whichever provider you use.

Second, where the data lives. With a custom build, your documents, search index and chat logs can stay in your own cloud account and region.

Third, access. The bot should only be able to see information that the customer in front of it is allowed to see. Quick builds often miss this.

How long it takes

A focused proof of concept can be live in 2 to 4 weeks. A production assistant with integrations, proper testing and human handoff usually takes 6 to 10 weeks. Assistants that take actions inside core systems take longer, mostly because of testing and approval flows.

If you're weighing up buying vs building for your support volume, our AI team can run the numbers with you on a free call.