Adding AI to your website: uses, costs, where to start

Adding AI to your website means connecting an AI model (such as ChatGPT or Claude) to one specific, repetitive job in your business: answering customers, drafting copy, pulling data out of forms and invoices, smarter search, or automating internal busywork. The "brain" itself is usually cheaper than people expect. For an assistant handling a few thousand conversations a month, usage fees land somewhere between a few dollars and a few tens of dollars. The real cost is building and maintaining the layer around it so the AI answers correctly, safely, and without leaking customer data. Start with one small, measurable job and expand from there.

What "adding AI" actually means

Think of AI as a very fast new hire who can read every document you hand them, but knows nothing about your company yet. They do good work only when given a clear job, the right documents, and someone checking their output. "Integration" is the work of putting that hire in the right seat in your process.

For a small or mid-sized business website, the five most common seats are:

Customer-support assistant. Answers questions about prices, availability and policies from your own documents, including outside office hours. I've written a separate guide on AI chatbots for your website.
Content drafting. Product descriptions, posts, reply templates. The AI drafts; a person reviews and edits.
Data extraction. Reads orders, invoices and sign-up forms, then fills in a spreadsheet or your business software. This is often the clearest saving in staff hours.
Smart search. A visitor types "sea-view apartment, dogs allowed" and finds the right listings, even when the page doesn't use those exact words.
Internal automation. Summarising customer feedback, sorting email, routing requests to the right person.

An honest note: if all you need is for your team to draft faster, a paid ChatGPT or Claude account is enough. Nothing needs to be built into your website. Integration earns its keep when the AI has to touch your data or your customers automatically.

What it costs: two very different bills

AI costs come in two separate parts, and a good proposal spells out both.

1. AI usage fees (paid monthly to the provider)

Providers such as OpenAI and Anthropic charge per token, a unit for how much text the AI reads and writes (roughly a small piece of a word). Current list prices per 1 million tokens:

Model	Input (AI reads)	Output (AI writes)
OpenAI GPT-5 mini	$0.25	$2
Anthropic Claude Haiku 5.5	$0.10	$0.50
Anthropic Claude Sonnet 5.5	$2	$10

From OpenAI's pricing page and Anthropic's pricing page, October 2026. The Haiku 5.5 price applies to prompts under 100,000 tokens.

Let's run the numbers. Say your assistant handles 3,000 conversations a month, and in each one the AI reads about 4,000 tokens (the question, the reference material, the chat history) and writes about 600:

With GPT-5 mini: roughly $6.60 a month.
With Claude Sonnet 5.5 (stronger, pricier): roughly $42 a month.

These are my estimates from stated assumptions, not figures measured on a real project, so budget double to be safe. The conclusion still holds: the "brain" is rarely the line item that hurts.

2. Build and maintenance

This is the main cost: preparing the documents the AI reads, writing the rules for what it may and may not say, connecting it to your forms, inbox or spreadsheets, handing over to a human when it isn't sure, keeping logs for review, and updating it when your prices or policies change. People in the field call this layer the harness, and it shapes the results more than the choice of model does. I go into it in The harness, not the model.

Off-the-shelf or custom?

An off-the-shelf tool (chatbot and helpdesk platforms) goes live fast, with no engineering team. In exchange you pay per seat or per outcome. Intercom's Fin AI agent, for example, lists from $0.99 per outcome, plus seats from $29 per seat per month (Intercom). At 1,000 outcomes a month, the outcome fees alone come close to $1,000, against a few tens of dollars in raw API fees.
A custom build costs more up front, but runs cheaply month to month and follows your actual process (for example, reading an invoice and writing it straight into the accounting tool you already use).

Rule of thumb: generic questions at low volume, try an off-the-shelf tool first. Work tied closely to your own data and process, at volume, and custom usually works out cheaper over time. My rates are on the pricing page.

The real risks

AI can make things up, confidently

AI sometimes produces information that sounds right but is wrong. This is called a hallucination. Even when asked only to summarise a document it was given, the best model on Vectara's leaderboard still adds false information in about 1.8% of summaries, and many popular models sit above 10%.

The consequences are real. In 2024 a Canadian civil tribunal ordered Air Canada to compensate a customer after its website chatbot gave wrong information about its bereavement-fare policy, and it rejected the argument that the chatbot was "responsible for its own actions" (American Bar Association). What the AI says on your website counts as you saying it.

How to reduce the risk: let the AI answer only from documents you've approved, make it say "I'm not sure, let me pass you to a colleague" when the information isn't there, and never let it promise prices or commitments on its own.

Customer data and privacy law

When the AI reads customer details (names, phone numbers, invoices), you are handing personal data to another company. Wherever you operate, three things are worth knowing:

You need a contract with the AI provider. Under the EU's GDPR, anyone processing personal data on your behalf must be bound by a written agreement setting out what they may do with it (GDPR Article 28). Ask your developer to confirm one is in place with the AI provider, not just with them.
Know where the data goes. Sending EU personal data to servers outside the EU is a regulated transfer, and fines for getting transfers wrong reach €20 million or 4% of worldwide annual turnover, whichever is higher (GDPR Article 83). OpenAI, for instance, offers data residency in a list of regions, including Europe, on request (OpenAI).
Know how it is used. OpenAI states that data sent through its API is not used to train its models unless you opt in, and is kept for up to 30 days for abuse monitoring by default (OpenAI). Ask for the equivalent statement from whichever provider your developer picks.

This is background so you know what to ask, not legal advice. If the AI will handle personal data, check your obligations with a lawyer in your own jurisdiction before going live. The safest rule of all: don't send the AI data it doesn't need.

How to evaluate an AI proposal

A trustworthy proposal usually covers five things: a specific job to solve, an estimate of monthly API fees at your volume, what the AI does when it doesn't know, where the data goes, and how success is measured. If two or more are missing, ask again.

Be wary of any proposal that promises "the AI never gets it wrong" or skips a trial run.

Questions to ask your web team

"Which model will you use? What are the estimated monthly API fees at my volume, whose name is the account in, and is there a spending cap?"
"Which documents will the AI answer from? When it has no answer, what happens? Does it hand over to a person?"
"Where does customer data go, which country are the servers in, is it used for training, and is a data processing agreement in place?"
"How will you test it before customers see it? Is there a set of sample questions and an accuracy rate?"
"If I want to switch to another model later, do we start from scratch?"
"When my prices or policies change, who updates the AI, and what does that cost?"

A serious developer answers with numbers. If all you hear is "don't worry, the AI is very smart", dig deeper.

Next step: where to start

In short: AI is worth it when it takes over a repetitive, measurable job. The "brain" is usually cheap. What deserves the investment is making it accurate, safe, and easy to maintain.

Three sensible first steps:

Pick one job that repeats and eats the most staff hours, usually answering familiar questions or keying in data from invoices and forms.
Run a small pilot for a few weeks, with a person reviewing the output.
Measure the hours saved or the share of correct answers, then decide whether to expand.

If you'd like help working out which job in your business is worth handing to AI, and what it would really cost to run, that's what my AI integration service is for. Get in touch.

Sources:

Adding AI to your website: what it can do, what it costs, where to start