What an AI customer support assistant really costs

Build cost, model usage, infrastructure and upkeep: what drives each, a formula for running costs and the questions to ask before signing.
4 min
09 Nov 2026
  • AI assistants
What an AI customer support assistant really costs

“How much does an AI support assistant cost?” has two answers. There is the one-time cost of building it, and the running cost of every conversation. Vendors often show only one of them. This article breaks both down so you can estimate your own case and compare quotes.

What you pay for

  1. Build: connecting your content, setting up search, writing the instructions, building the chat interface or connecting it to your help desk, testing.
  2. Model usage: the language model provider charges per amount of text processed (tokens).
  3. Infrastructure: hosting, the search index, logging. Usually small compared to the rest.
  4. Upkeep: reviewing answers, updating content, adjusting instructions, watching for model changes. This is the cost most often forgotten.

What drives the build cost

The language model itself is the easy part. Cost grows with:

  • Number and state of sources. One clean help center is quick. Ten sources across PDFs, wikis and a product database, with duplicates and outdated pages, take much longer.
  • Integrations. Answering general questions is simpler than “where is my order?”, which needs a secure connection to your order system and a way to identify the customer.
  • Channels. A website widget is one channel. Adding WhatsApp, Telegram, email or a help desk like Zendesk or Freshdesk adds work for each.
  • Languages. Models handle many languages well, but every language needs its own test questions and content checks.
  • Handover to people. Passing the conversation with its history to an agent in your help desk is a separate piece of work.
  • Quality requirements. A test set, evaluation runs and an approval process add time, and save much more later.

A focused first version, one audience, one or two channels, the main content sources and handover to a person, is typically a project of a few weeks. Ask for a fixed scope for that first version rather than an open-ended estimate.

How to estimate running costs

Model providers price per million tokens, separately for input (what the model reads) and output (what it writes). A token is roughly three quarters of an English word. In a RAG assistant most tokens are input, because every answer includes the retrieved passages.

A simple formula per conversation:

cost = turns × (input tokens per turn × input price + output tokens per turn × output price)

Example with explicit assumptions:

  • 3 turns per conversation;
  • about 4,000 input tokens per turn (instructions, retrieved passages, chat history);
  • about 300 output tokens per turn.

That is about 12,000 input and 900 output tokens per conversation. Multiply by your model’s current prices from the provider’s pricing page. Prices differ by more than ten times between small and large models and change often, so always calculate with today’s numbers. For many support use cases, a mid-sized model gives answers that are good enough at a fraction of the cost of the largest one. The test set tells you which model is sufficient.

Then compare with the cost of the same conversations handled by people: conversations per month × share the assistant resolves × average handling time × hourly cost. For most teams with real volume, model costs are a small part of this. The real savings and the real risk come from the share of conversations the assistant resolves correctly.

Ways to keep running costs down

  • Retrieve fewer, better passages instead of many long ones.
  • Use a smaller model for simple steps, like detecting the topic of a question, and a larger one only for the answer.
  • Use the provider’s prompt caching for the fixed part of the instructions where available.
  • Answer the most frequent questions from approved, ready-made answers.

Questions to ask before signing

  1. Is the price for a fixed scope? What exactly is included in the first version?
  2. Who pays for model usage, and is it billed at cost or with a markup?
  3. Who owns the code, the prompts and the search index? Can we move to another vendor or model?
  4. How is quality measured before launch, and what is the target?
  5. What happens when the assistant does not know the answer?
  6. What does support after launch include, and what does it cost per month?

The cheapest option is not always building

If your help desk already offers an AI add-on and your content lives entirely inside it, try that first. Custom development makes sense when you need to combine several sources, connect your own systems, control the model and the data, or when per-seat or per-resolution pricing becomes expensive at your volume.

For background on how such assistants work, read How to make an AI assistant answer from your own documents. We build support assistants with a fixed scope for the first version, see fixed-scope packages and our AI assistants service, or book a free 30-minute call to estimate your case.

    Book a free 30-min AI call