AI Cost CalculatorWhat actually drives the bill
Two things set a chatbot's cost: how many messages it handles, and how much text each one carries. Messages come from your traffic - visitors, times the share who open the chat, times how many messages each chat runs to. And every message the bot reads carries its standing instructions, the visitor's text, a bit of the earlier chat, and anything it looked up. That earlier-chat part is the one that grows, which is why one long twelve-message chat costs more than two short six-message ones. A bot that answers quickly is both nicer to use and cheaper to run - those two goals rarely fight.
The other two dials are search and model choice. Looking things up in your help docs makes answers better, but it adds text to every message, so only pull in what you need. And a small, fast model handles most support and FAQ questions fine - save the premium models for tricky reasoning or areas where a wrong answer is costly. You can compare models on your own traffic in the chatbot cost calculator. Whatever the bill comes to, it usually looks tiny next to the support staff time a good bot saves - the support ROI calculator puts numbers on that side.
Start with the traffic, not the bot
Website visitors aren't chatbot users. Start with your monthly visitors, then estimate how many notice the chat, how many open it, how many actually send a message, and how many keep going. Then split the quick one-line questions from the full back-and-forth chats. A bot tucked away in the help centre gets opened far less than one that pops up on every pricing page.
Use a range until you have real numbers. A quiet-month and a busy-month version is far more honest than assuming every visitor starts a four-message chat.
Chats grow one message at a time
On the first message, the bot reads its instructions and a single question. By the sixth, it's also re-reading most of the earlier chat. That's why cost can climb faster than the message count.
You don't have to re-send everything forever. Keep the recent messages as-is, summarise the older ones, drop the small talk, and store steady facts like the customer's plan separately. The aim is to keep what actually changes the answer - not every "hi" and "thanks".
Handing off costs something, but so does refusing to
Passing a chat to a person costs staff time. But refusing to hand off can be worse - the bot keeps giving unhelpful answers while the customer gets more annoyed. Set clear triggers for when to hand over: it's failed a few times, it's unsure, the request needs account access, or the customer just asks for a human.
When you do hand off, pass along the chat and a short summary. Making the customer start their story over wastes both the bot's work and the agent's time.
A simple support example
Say 4,000 people open the bot in a month. Half ask just one question; the rest average five messages. That's 12,000 messages, not 20,000. If 35% of chats need a person, count the other 65% as "handled" only when the bot truly sorted the issue out - not just when it replied.
Then try different models on the same chats. A small model may handle routine policy questions well, while billing or troubleshooting needs a stronger one or a quick handoff. Mixing models keeps quality up without paying top price for every "hello".
Before you launch
- Measure cost on real support questions, not a polished demo.
- Set a target reply length and a limit on chat length.
- Test what happens when the help docs are missing, out of date or contradict each other.
- Make the "talk to a person" option easy to find, and carry the chat into the handoff.
- Track how often it resolves the issue, how often people come back, satisfaction, and cost per resolved chat.
The chatbot calculator gives you the running cost. The quality numbers tell you whether the saving is real.