Owners who are drowning in the same five customer questions ask me about AI chatbots more than almost anything else right now. The build can be cheap. The question that decides whether it is a good idea has nothing to do with cost: what happens when it answers wrong?
Key Takeaways
- Off-the-shelf AI support bots are priced per conversation or per resolution, and are the right start for most owners.
- A custom assistant makes sense when it needs your own data, your own rules, or to take actions like booking.
- The running cost of the AI model is often small. The cost of wrong answers can be large.
- Every business bot needs a clear handoff to a human.
Three Ways To Get One
1. A bot built into a tool you already use. Many help desks, website builders, and booking tools now include an AI assistant. Cheapest and fastest to try.
2. A dedicated AI support product. Priced by usage. For example, Intercom lists its Fin AI agent at $0.99 per outcome. Easy to set up, trained on your help articles.
3. A custom assistant. Built for your business, connected to your own systems, able to look up an order or book a slot, with rules you control.
What a Custom Assistant Costs
Using the same bands as our free honest read:
| Version | What it does | Typical build cost |
|---|---|---|
| FAQ assistant on your website | Answers from your own documents, hands off to email | $3,000 to $12,000, one developer |
| Assistant that takes actions | Looks up bookings or orders, creates requests, with a human handoff | $15,000 to $45,000, small team |
Then there is the running cost of the AI model itself, which is usually priced per million "tokens" (roughly, pieces of words). As a reference point, Anthropic lists its Claude Haiku 4.5 model at $1 per million input tokens and $5 per million output tokens. For a small business answering a few hundred questions a month, model costs on a small model like that tend to be modest. Hosting, monitoring, and maintenance usually cost more than the model.
The Question That Matters Most
Our honest read asks one question for every AI assistant idea: "What happens when it answers wrong? Who catches it, and what does a wrong answer cost you?"
The answer decides the design:
- A wrong answer is harmless (store hours, parking): let it answer freely.
- A wrong answer is expensive (prices, availability, refunds): it must look the answer up in your real data, never guess.
- A wrong answer is dangerous (medical, legal, safety): it should hand off to a person, not answer.
The same read lists the typical hard parts: handing off to a human before a wrong answer costs a customer, connecting it to real data instead of guessing, and keeping it honest about what it does not know.
What To Prepare Before Any Build
- The 20 questions customers ask most, with your real answers.
- The questions the bot must never answer.
- Where the true data lives (booking system, price list, order system).
- Who gets the handoff, and how fast they reply.
That list is most of the specification. It also tells you whether option 1 or 2 is enough.
A Concrete Version
A veterinary clinic gets about 40 calls a day, and most are the same: opening hours, whether a vaccine is in stock, and appointment requests.
They start with option 2, an AI support product trained on their FAQ. It handles hours and general questions well. It cannot see vaccine stock or the appointment book, so it guesses, which is exactly the expensive kind of wrong.
They then commission a small custom assistant: it checks stock from their inventory export, offers real open appointment slots, and creates a booking request that staff confirm. Anything about a sick animal goes straight to a person, by design. With the booking system connection, that lands in the small-team band.
The Honest Counterpoint
Many businesses do not need a chatbot. A clear FAQ page, a good booking form, and an auto-reply often cut the same calls for almost nothing. Some customers also dislike talking to bots and will call anyway. Try the cheap fixes first, and measure: if calls do not drop, a bot may not help either. Our post on what to automate first in a small business covers how to pick.
Frequently Asked Questions
Can I build a chatbot with an AI builder?
A simple FAQ bot, often yes. Connecting it safely to live data is where you want a developer. See can AI build my app without a developer.
Will the AI make things up?
It can. Grounding answers in your own data and limiting what it may answer reduces that a lot, but a human handoff is still needed.
Does a chatbot need to follow privacy rules?
If it collects personal information, yes, and more so for health or financial data. This is general information, not legal advice.
How do I know it is working?
Track three numbers: questions it resolved, questions it handed off, and complaints about wrong answers.
The Bottom Line
Start with an off-the-shelf bot, build custom when it needs your data or has to act, and design around wrong answers from day one. For a verdict on your own assistant idea, use the honest read. Our software factory builds custom assistants at a fixed price. Related: hiring engineers for an AI startup.
Roberto Espinoza is CEO of Ruzora, which builds custom software for business owners at a fixed price and places pre-vetted senior LATAM engineers with US teams. Get a free honest read on your idea.
