The question has two answers, separated by two orders of magnitude. You can put a chatbot on your site this afternoon for a few hundred dollars a month. A custom agent, built against your systems, runs $15,000 to $50,000. Both are chatbots in the sense that a person types and something types back. They do different jobs, and most of the confusion in this market comes from calling them the same thing.
This piece lays out what our AI chatbots and agents work costs and what the difference in price buys.
The published ranges
These come from the public estimator on our /connect page. Every quote we send is specific to the project; these are the brackets projects land in.
| Service | Typical range |
|---|---|
| AI chatbot or agent | $15,000 to $50,000 |
| Custom AI application | $25,000 to $100,000 |
| Workflow automation | $8,000 to $30,000 |
| AI consulting | $5,000 to $15,000 |
A support agent answering from a documentation site sits near the bottom of the chatbot line. An agent that reads a customer's order history, checks inventory, then reschedules a delivery sits near the top, or crosses into the application line if it needs an interface of its own. Consulting is the smallest line because it produces a plan rather than a system; some clients buy it first to find out which of the other lines they actually need.
What the cheap option does
Embed widgets start at a few hundred dollars a month. They crawl your public site, index it, then answer questions from that index. Setup is an afternoon.
For FAQ deflection that is a fair trade. If a quarter of your inbound email is shipping times and return windows, a widget absorbs a good share of it, and a few hundred dollars a month against a support inbox is easy math. We tell clients to try one first when their questions really are that shallow.
The ceiling arrives quickly. The widget knows what is on your website. It does not know what is in your ticket history, your policy documents, your product database. It can describe your return policy; it cannot start a return. When a customer asks something the site does not cover, it guesses or apologizes, and you find out from a review rather than a dashboard.
What the money buys
A custom agent is four pieces of work. A widget is a thin version of one of them.
Grounding. The agent answers from your own documents, tickets, policies, retrieved at question time. This is what people mean by RAG. The practical effect is that answers trace back to a source, so a wrong answer is a content problem you can fix at the source instead of a mystery.
Integration. The agent is wired to the systems it needs to touch: bookings, orders, CRM. This is where "chatbot" stops being the right word. Reading a booking is one level of work. Changing one is another, with permissions, confirmations, an audit trail, and a plan for what happens when the API is down.
Evaluation before launch. A test set of real questions with known-good answers, run against every change. You want a measured accuracy figure on your own questions before customers start asking them, and you want to know that today's fix did not break last month's behavior.
Monitoring after launch. Conversations get logged and analyzed. Resolution rates get tracked. The answers the agent missed become next month's material, so the thing improves on a schedule. An unmonitored agent decays, because your products change and its knowledge sits still.
Four kinds of agent
Support. Answers customer questions, resolves what it can, hands the rest to a person with the conversation attached. Priced by how many systems it has to read from.
Intake. Qualifies inbound leads or requests, collects the facts a human needs, routes to the right queue. Often the fastest payback, because the alternative is a senior person reading unqualified inquiries.
Knowledge. Internal. Answers staff questions from company documents so new hires stop interrupting the people who know things. No customer risk, which makes it a reasonable first build.
Action. Does things. Books, cancels, reschedules, updates a record, opens a ticket. The highest value and the highest scrutiny; every write needs a confirmation path and a way to undo.
What moves your number
How many systems it touches. Each integration is scope of its own: authentication, error handling, someone else's API that changes without telling you. Adding the second integration usually costs more than the first did, because the two have to agree about the same customer.
How costly a wrong answer is. An agent quoting store hours has a low cost of error. An agent quoting contract terms to an enterprise customer has a high one, and evaluation work scales with that number.
Whether your content exists. Grounding needs something to ground in. A company with written policies starts faster than one where the policy lives in three people's memory, and writing that content down is often the first invoice.
We build Claude-first, so model selection is one fewer variable in a scope conversation.
If the budget is tight, narrow the job instead of thinning the build. One agent that handles your top twenty questions correctly is worth more than one that handles two hundred at seventy percent. That is how most of our AI chatbots and agents projects start, with a first version scoped small enough to measure. Our broader AI services pages cover the other shapes this work takes when a conversation is the wrong interface.
The fastest way to get a number is to name the systems the agent has to touch and the questions it has to answer. Send us that list and start a project; we will come back with a price and a scope you can cut.