Every Visitor Sentence Turns the Meter
Put a chat box on a page and each thing a visitor types becomes a request to a language model living somewhere else. Each request is priced. Most people know that much going in. What comes apart later is the assumption that the total tracks the number of questions.
The model remembers nothing between requests. A conversation only feels continuous because everything said so far is shipped along again every time. So the tenth question, typed at the same length as the first, registers far larger on the meter. Three things ride in every request: the visitor’s sentence, your own block of instructions telling the model what role to play and what never to say (the system prompt), and the answer that comes back.
Which puts the room to cut cost in conversation design rather than in the feature list. How many turns one conversation is allowed to run, when older messages get trimmed, whether the instructions that go out unchanged every time are cached. Decisions left out of the scope document turn up on the bill instead.
What Rides Along With “A Bot That Knows Our Company”
Requests usually arrive in roughly those words. The model does not know your company, so a second piece of machinery gets built: something that searches your own material for the relevant passage and hands it over alongside the question. Gathering the material, cutting it into chunks, indexing it so it can be searched, and picking the right chunk per question all sit inside that machinery. What moves the quote is the state of the material. Pulling from tidy web pages and documents is not the same job as pulling from scanned PDFs and spreadsheets that every department formats differently.
That machinery is never finished either. Change a price or a policy and the index has to be rebuilt, and if it is not rebuilt the bot states last month’s terms with total confidence. Without an agreement on who updates the source and how many days it takes to land, half the wrong answers come from here.
A Customer Who Gets a Strange Answer Says Nothing
A chatbot answers in sentences even when it has nothing. Rather than leave a blank it fills the gap with something plausible. So what has to be settled before building is not an accuracy target but a handling procedure. Whether conversations are logged, and if they are, who reads them and when.
Without logs you cannot know it broke. Customers who get a strange answer rarely say so. They close the window.
Draw the boundary at the same time. Whether a refund applies, when the delivery lands, what something costs: promises that turn into disputes when broken stay out of the chat window and go to a person. Build the handoff button and the conditions that make it appear at the start, and there are fewer apologies to write later.
The Disclosure Line Is a Checklist Item, Not a Build Task
If any of your users sit in the EU, the AI Act’s transparency obligations apply from 2 August 2026: people have to be told when they are dealing with an AI system, and synthetic output has to be marked as machine generated. The European Commission publishes the implementation timeline those dates come from. Your own market may have added a rule of its own, so that is worth checking before launch rather than after. As a line of copy it costs almost nothing to build. Deciding which screen it belongs on and how it is worded is better done while the screens are being drawn.
On the Day the Model Changes, the Code Is Fine
When an external service changes its contract, the integration breaks visibly. A language model gives you no such signal. The call stays identical while the character of the answers moves. Switch to a newer version and the same question comes back phrased differently, and hold on to an older one and it eventually gets an end-of-life date. Nothing in the code changed, and the answers still have to be checked again. So write down twenty representative questions with the answers you want, and run those twenty again every time the model is touched. Without that list the only verdict available is “it feels off.” The version of this where the contract itself breaks is in An Integration Starts Being Work After the First Call Succeeds.
We Start From What It Is Allowed to Answer
You can come to us still undecided about whether to add one at all. Weple’s AI feature work draws the map first, what the bot may be asked and where it hands over to a person, then splits the build cost and the monthly cost across that map. If you already keep a list of common questions or a help document, that is the starting material, and if not, we go through the support inbox together. Units of work are listed on the pricing page. In a first conversation what we count is not how many questions arrive, but how many turns one of them runs.
In a similar spot
Send us where things stand and we reply with the scope and the price within 24 hours.