The cost of an AI agent on your website, explained

TL;DR

• The cost of an AI agent on your website has four layers: the model usage behind each answer, the vendor's price, your team's time and the extras. Most budgets count only the second.
• The model itself is cheap: at October 2026 API list prices, one grounded answer costs a fraction of a cent to a few cents. The subscription and the hours you spend on setup and upkeep usually decide the total.
• Judge the result by cost per outcome you care about, such as a qualified lead or a meeting booked, not by cost per message.

Layer 1: what one answer costs to produce

Every answer from an AI agent is a call to a large language model, billed in tokens, the chunks of text models read and write. An agent that answers from your website uses RAG: it finds a few relevant passages from your pages and sends them to the model with the visitor's question and instructions. A typical request might carry 3,000 input tokens and produce a 300-token answer.

At OpenAI's API list prices on October 21, 2026, that works out as follows:

Turning your site into a searchable knowledge base costs even less. OpenAI lists text-embedding-3-small at $0.02 per million tokens, so embedding 1,000 pages of about 1,000 tokens each costs around two cents.

These are illustrative calculations with one provider's list prices; other providers price in the same range, and cached input, batch processing and conversation length change the numbers. The point is the order of magnitude: the raw model cost of answering a visitor is small. What vendors charge on top pays for hosting, retrieval, the widget, analytics, integrations, support and margin.

Layer 2: what vendors charge

Vendors package that cost in different units. Some examples from their pricing pages, checked October 21, 2026:

The units are not interchangeable. An outcome or a resolved conversation may include several answers; a conversation pack counts conversations whether or not they end well. Our guide to AI agent pricing models explains how each unit shifts risk between you and the vendor and how to model them at different success rates.

For ASDK, all plan prices and limits are on the pricing page: Free includes 50 answers a month, Starter 1,000, Pro 5,000 and Business 20,000, with annual billing about 20% cheaper.

Layer 3: your team's time

This is the layer most budgets miss. An agent that answers from your content needs someone to:

• Set it up: connect the site, choose what it should and should not discuss, set up handoff and lead delivery.
• Test it before launch: a week of part-time work is typical for a proper test set; see how to evaluate an AI assistant's answers.
• Review conversations every week, especially unanswered questions and handoffs.
• Fix content when answers are wrong or missing. Fixes are usually small, but they need the person who owns the page.
• Handle handoffs: the conversations the agent passes to a person still take that person's time.

To put a number on it, the U.S. Bureau of Labor Statistics lists the median wage for customer service representatives at $21.53 an hour in May 2025 (BLS). Four hours a month of review and fixes at that rate is about $86 in wages, before overhead, which is comparable to many subscription fees. If the reviewer is a marketer or a founder, the real cost is higher.

Layer 4: the extras

• Branding removal. Often a higher plan or an add-on.
• More sites, seats or pages. Agencies and multi-brand companies hit these first.
• Voice minutes. Spoken conversations are priced per minute almost everywhere. ASDK lists extra voice minutes at $0.25 each for plans that include its Voice agent, which is coming soon.
• Integrations. A native CRM connector may need a higher plan; an automation tool in the middle has its own price. Zapier, for example, lists webhooks only on its paid plans (Zapier).
• Overage. Some products bill automatically above the allowance, others stop answering. Know which.

A worked example

An illustrative B2B website, not a benchmark: 3,000 visits a month to product and pricing pages. 3% of visitors start a conversation, which gives 90 conversations; each has three answers on average, which gives 270 answers a month.

Two observations. First, model usage is a rounding error at this volume; the plan and the time are what you pay for. Second, the sales-focused plan costs more than three times the basic one, so it has to produce results: if it adds four qualified meetings a month, that is about $32 per meeting on monthly billing. Whether that is good depends on what a meeting is worth to you, which you can estimate with the funnel math in AI sales agent for B2B websites.

The comparison people usually make

The common question is whether an agent is cheaper than people. A useful way to frame it: staffing one live-chat seat around the clock takes 168 hours a week, or about 728 hours a month. At the BLS median wage, that is roughly $15,700 a month in wages alone, before benefits, management and tools. An agent covering the same hours costs a small fraction of that.

But that comparison flatters the agent. It does not handle every conversation well, and the important ones, a large deal or an unhappy customer, should reach a person. The realistic picture is an agent that answers routine questions at any hour and hands the rest over with the context attached, so your team spends its time on fewer, better conversations. Good human handoff is part of the value, not a cost to minimize.

How to keep the cost down

Most of the levers are in your hands, not the vendor's:

• Better pages, shorter conversations. When the answer is clear on the page, the agent answers in one turn instead of three. Every content fix lowers usage and review time at once.
• Scope the agent. An agent that only covers your products, pricing and policies uses fewer answers than one that chats about anything. It is also safer.
• Hand over early when it matters. Ten back-and-forth answers that end in "let me connect you" cost more than a handoff after the second question, and annoy the visitor.
• Turn on what you use. Pay for the sales-focused plan when you are ready to act on qualified leads and meetings, not before.
• Watch the allowance. Set alerts before you reach the limit, so a busy month does not end in surprise overage or a silent widget.
• Review in batches. A fixed hour each week is cheaper than reacting to every odd transcript as it arrives.

Measure cost per outcome

Cost per message tells you little. Track the cost of what you want:

• Cost per qualified lead or per meeting booked for sales-focused agents.
• Cost per resolved question for support-focused agents.
• Cost per hour of human time saved, if you can measure handoffs avoided.

Divide the total of all four layers, not just the subscription, by the outcome count. Our guide to measuring an AI agent covers how to count those outcomes honestly.

Before you buy: five questions

If you are starting with no budget at all, free AI chatbot plans compared shows what each free tier covers.

An AI agent on your website is cheap to run and not free to own. The model costs pennies; the plan, the extras and the hours your team spends keeping answers right are the real budget. Count all four layers, divide by the outcomes you care about, and you will know whether the agent is worth it.