Every AI chat vendor looks affordable on the pricing page, because the pricing page shows month one. What decides whether you are happy is month three, when the assistant is working and the volume has moved.
Four shapes are in common use, and they behave completely differently as that volume grows. The sticker price is the least interesting thing about them.
The same assistant, four ways to be charged
Move the volume and watch which numbers move with it.
The point of the exercise is not which column is cheapest at 12,000. It is that each column answers to something different: headcount, traffic, text volume, or a boundary somebody published.
What each shape is really charging for
Per seat is the shape worth reading twice, because its incentive runs backwards against what you are buying. The assistant's job is to answer without a person. In a per-seat contract that shows up as nothing at all until somebody leaves the team.
The five questions that decide the bill
The shape is the frame. These are what actually move the number.
- What counts as one unit? A conversation, a message, an answer, or a session? A twelve-message thread is one conversation and twelve messages, and the difference is an order of magnitude.
- Does a failed answer count? If the assistant says it does not know, some vendors still bill it. Ask.
- What happens at the limit? Three behaviours exist: it stops answering, it charges overage, or it quietly degrades. Only one of those is acceptable without a warning.
- Are seats charged for people who only read? Analytics viewers and managers add up quickly on per-seat plans.
- What does annual buy? Usually two months. Occasionally it buys a lock on a price that is about to change, which is worth more.
A plan that says unlimited has a limit somewhere: in the terms, in a rate limiter, or in a phone call once you are big enough to notice. Prefer a published number you can plan against over a word you cannot.
Where we sit, plainly
We use the fourth shape, and here are the numbers so you can hold them against anybody else's. Starter is $75 a month with 3,000 answers, Pro is $199 with 8,000, Business is $499 with 20,000. Yearly is ten months for twelve. Every other limit, from documents to storage, is on the pricing page rather than in a sales call.
There is also a fifth shape as an add-on, because some teams genuinely want the meter. For $99 a month on any plan you connect your own Anthropic and OpenAI keys, those providers bill you directly for the model calls, and we stop counting answers against your plan. Everything else, indexing, retrieval, widgets and analytics, stays where it was.
The honest weakness of our shape is the same as everybody else's: there is a step at the boundary. What makes a step tolerable is seeing it coming, which is why the number is on the page.
welcomeai.devEvery limit each plan enforces, written downHow to estimate your own volume before you sign
You do not need a spreadsheet. Take the support conversations you handled last month, add the questions that arrive by email and get answered in one line, and assume between a fifth and a third of the visitors who would have contacted you will ask the assistant instead.
Then double it, because a question that used to be too much trouble to ask gets asked when the box is right there. That is the number to price against, and it is also the number that makes per-conversation pricing uncomfortable.
Questions people ask
Work out your monthly answers first, then read pricing pages with that one number in hand. Every vendor's page is designed to be read without it.
welcomeai.devThe cost of an AI answer, broken down
