Adding AI to an existing product costs $5,000–20,000 for one feature on a screen you already have, $15,000–60,000 for a chatbot or search over your own content, $20,000–80,000 for an assistant inside the product, and $40,000–150,000 for an agent that takes actions. After launch, you also pay a monthly AI bill, usually $50 to $10,000.
AI quotes vary wildly because "add AI" can mean a single summary button or a system that acts on your customers' data. The price depends on how much the AI can see, how much it can do, and how right it has to be. Below: prices by type, the monthly running cost with a worked example, when an off-the-shelf tool is enough, two real AI products, what drives the cost, and the red flags in an AI quote. If you're still deciding whether AI belongs in your product at all, start with when AI belongs in a product.
AI feature cost by type
Freelance prices in US dollars, for a product that already exists. The monthly cost is what you pay the AI model provider.
| Type | Build | Timeline | Monthly AI cost |
|---|---|---|---|
| One AI feature on an existing screen | $5,000–20,000 | 2–4 weeks | $50–500 |
| Chatbot or search over your content | $15,000–60,000 | 4–10 weeks | $200–2,000 |
| Assistant inside your product | $20,000–80,000 | 6–12 weeks | Depends on use |
| AI agent that takes actions | $40,000–150,000 | 8–16 weeks | $1,000–10,000 |
| Type | Typical range | Real projects |
|---|---|---|
| One AI feature on a screen | $5,000–20,000 | — |
| Chatbot or search over your content | $15,000–60,000 | — |
| Assistant inside your product | $20,000–80,000 | — |
| AI agent that takes actions | $40,000–150,000 | — |
| An AI SaaS I built | — | The Chat Company: $20,000 |
An example of the first row: a button that summarizes a long record, drafts a reply, or sorts incoming requests. The fourth row is different in kind: the AI changes records, sends messages, or runs workflows, so it needs limits, approvals, and much more testing.
What an AI feature costs every month
AI models are billed by usage, measured in tokens, roughly pieces of words. Every question and every answer uses some. So the monthly bill grows with how many people use the feature, how often, and how much text goes in with each request, like the documents a chatbot reads before answering.
It can be kept predictable. The usual controls are spending limits on the provider account, a cap on how long each conversation can run, using smaller, cheaper models for simple tasks, and tracking use per customer. The Chat Company uses two of them: token use is tracked per business, and conversation length is capped, so AI costs grow with customers rather than by surprise. Ask for an estimate of a normal month and a busy month before the build starts, based on your real traffic.
What an AI chatbot costs per month: a worked example
Here's the arithmetic for a chatbot that answers from your website, using the providers' published prices in September 2026. The assumptions are round numbers for illustration: 10,000 conversations a month, four questions per conversation, and each question sending about 3,000 tokens (the relevant pages plus the chat so far) and getting about 300 back. That's 120 million tokens in and 12 million out each month.
| Model | Price per million tokens, in / out | Monthly AI bill |
|---|---|---|
| OpenAI gpt-5-mini | $0.25 / $2 | About $54 |
| Anthropic Claude Haiku 4.5 | $1 / $5 | About $180 |
| Anthropic Claude Sonnet 5 | $2 / $10 | About $360 |
Two things stand out. The model you pick changes the bill several times over, so it's worth testing a cheaper model on your real questions before defaulting to the biggest one. And most of the cost is what goes in, not what comes out: the pages the bot reads before answering. Sending fewer, better-chosen pages with each question is usually the biggest saving. On top of the model bill, budget for hosting and the search index that stores your content. Prices change often, so recheck them before you commit.
Off-the-shelf AI tool or a custom AI feature?
For a support chatbot, you don't always need to build anything. Tools like Intercom's Fin plug into your help desk and charge per answered conversation: $0.99 per outcome on Intercom's pricing page. At 1,000 resolved conversations a month, that's about $990 a month, with no build cost and a working product on day one.
Buy when the job is standard, like answering support questions from a help center, and the volume is modest. Build when the AI is part of what customers pay you for, when it needs your product's own data and actions, or when volume is high enough that paying per conversation costs more than running your own. The Chat Company is the build case: the chatbot is the product itself, sold to other businesses, so renting someone else's would have left no product to sell.
Two real AI products and what they cost
The Chat Company: an AI chatbot SaaS for $20,000
The Chat Companylets any business add a chatbot to its website that answers from the business's own pages. It began as an AI-generated demo. I kept about 20%, rewrote the rest, and built the full product in nine weeks for $20,000.
The core is a technique called RAG (retrieval-augmented generation): before answering, the bot looks up the most relevant parts of that business's content and gives them to the model. That keeps answers grounded in the website instead of guesswork. Two safeguards mattered as much as the AI itself: every lookup is filtered by business, so one customer's content can never appear in another customer's chatbot, and if the lookup is slow or fails, the chat still answers instead of breaking. It now serves 30+ businesses and handles 100K+ messages a day.
A B2B sales platform: AI inside a larger product
A confidential B2B sales and gifting platform is the second example. Its name and budget are confidential. There the AI isn't a chat box. It works inside the product's own workflows: helping set up campaigns, personalizing landing pages and emails from contact data, and taking actions through the platform's own tools. I built those systems as part of a team of 14 over 30 weeks, and the platform has run about 10,000 campaigns for 20+ companies. It's what the assistant and agent rows in the table look like in practice: most of the work is the product around the AI, the data, the permissions, and the actions, not the model itself.
What makes an AI feature cost more
Messy data
A chatbot is only as good as what it reads. Duplicate, outdated, or scattered documents need sorting before the build, and that work is often underestimated.
How right it has to be
A draft someone edits can be wrong sometimes. An answer a customer relies on can't. Higher stakes mean more testing, source citations, and fallbacks.
What the AI is allowed to do
Reading is cheaper than acting. Every action, like updating a record or sending an email, needs permissions, limits, and often a human approval step.
Access to private data
An assistant that sees each user's own data must never show it to anyone else, which adds careful permission checks throughout.
How to keep an AI feature affordable
- Start with one task. The one users do most, where AI clearly saves time.
- Use existing models.Don't train your own for a business feature.
- Give the AI your content at answer time. Cheaper and easier to update than retraining anything.
- Keep a person in the loop. Let the AI draft, and a human approve, before letting it act alone.
- Measure use from day one. So you know the monthly bill before it grows.
The same thinking applies to building a whole product around AI, as covered in what a web product costs. Of the five, keeping a person in the loop matters most early on. It catches wrong answers before customers see them, and the corrections show you exactly where the AI needs better data or clearer instructions, which is cheaper than guessing.
Red flags in an AI development quote
A good AI quote reads like any other good quote: a named task, a clear scope, a price, a timeline, and an honest estimate of running costs.
A brief to get an accurate AI quote
See how I add AI to products on my AI features page, or send me your brief, and I'll tell you what it would cost to build and to run each month.
