What does the build cost?
Mine starts at From $799. That is the same starting point as any small web app, because the work is mostly not the model: it is your files, the permissions, the interface and the tests. Small jobs start at $49: a fix, a small automation, a change to a page. Each kind of work has a starting price below; most work is quoted after a free first call as one fixed price, agreed in writing, because the final number depends on the work.
Three things move the build price more than any other decision. How many document types have to be read. Whether some files are private to some staff. Whether answers must be logged so you can audit what was said to whom.
A single shared inbox of PDFs for five people is a small build. Twelve departments, three sources and a rule about who may read payroll is a different project, and anyone who quotes the two the same price has not looked at your files.
What are the four parts of the bill?
This is the shape of it. The first is one-off; the other three arrive every month whether anyone asks a question or not, so price the year rather than the project.
| Part | When you pay it | What moves the number |
|---|---|---|
| Build: retrieval, interface, permissions, tests | Once, as a fixed price | Number of document sources and who may see each |
| Hosting and the search store | Monthly | Size of the document set and whether you self-host or rent |
| Model calls: embeddings plus answers | Monthly, by usage | Questions asked per month and how much text each answer reads |
| Upkeep: model versions, platform changes, new files | Monthly, or per visit | How often your documents change and how fast you need them reflected |
Why do two quotes for the same chatbot differ so much?
Because they are not the same product. A demo that answers well when you ask it something you already know the answer to costs very little. A system that names its source, refuses when the files do not say, keeps private documents out of the wrong hands and still works after your policy changed in March is where the hours go.
The visible part is a chat box. Underneath it, somebody has to chunk the documents, embed them, keep the index current, decide what happens when two files disagree, and write the tests that catch a bad answer before a customer does. The gap between quotes is usually which of those are included, and most buyers only discover the list after the first wrong answer.
What does it cost to run each month?
It follows usage, and usage is predictable if you ask two questions first: how many people will use it, and how many questions each will ask in a day. A staff assistant for a team of eight is a rounding error. A customer-facing bot on a shop with real traffic is a line item worth planning for.
The costs people did not expect, in the order they usually appear:
- Embedding the whole document set again every time something changes, instead of only what moved.
- Answering each question by stuffing in far more retrieved text than the answer needed.
- Paying for a bigger model on questions a smaller one gets right, which is why I keep the simple cases simple.
- Letting logs and old indexes pile up on a store that bills by size.
How do you pay less without getting less?
Two decisions carry almost all of the saving, and both are made before any code exists.
- Narrow the document set. Ten well-kept files beat two hundred archived ones on both cost and accuracy, because there is less to read and less contradiction to lose the answer in.
- Bound the scope. Answer the twenty questions you actually get, and hand everything else to a person. An assistant that admits it does not know is cheaper than one that always tries.
- Fix the documents first. Duplicates, scanned images and out-of-date versions are the usual reason a project runs long, and they are free to clean before the build starts.
- Own the accounts. When the assistant runs on your own hosting and your own model key, the monthly bill is visible and switchable instead of bundled into a platform price that goes up at renewal.
What to remember
- The build starts at From $799 as one fixed price, and the honest question to ask any developer is which of the four parts of the bill that includes.
- Only the build is one-off. Hosting, model calls and upkeep arrive all year, so compare quotes on a twelve month total.
- Usage follows the number of questions and how much text each answer reads, which is why a narrow document set and a bounded scope are the two biggest savings available.
- A rented platform can look cheaper in month one and cost more by month twelve, because the result does not leave with you.
Sources
Primary documentation, for anyone who wants to check the details.