Pricing a chatbot is not really pricing software. It is pricing an answer to the question "how much of what our customers ask is already written down somewhere". A company with three years of well-kept help articles and a company with a PDF from 2019 need the same code and completely different amounts of work, which is why a single number quoted off a one-line brief is close to meaningless.
Build cost by project size
We size chatbot projects the same way we size everything else, by how much distinct functionality is involved rather than how many pages get indexed. Indexing forty pages and indexing four hundred is nearly the same job. Adding a live order lookup is not.
| Tier | What it typically includes | Timeline | Typical cost |
|---|---|---|---|
| Focused | One knowledge source (your site or your help centre), chat widget, handoff to email | ~2 weeks | From $3,000 |
| Standard | Several sources including docs and PDFs, handoff into a helpdesk or CRM, gap dashboard | ~3 weeks | From $5,000 |
| Substantial | Signed-in user context, live actions (order status, bookings, account lookups), more than one channel | ~5 weeks | From $9,000 |
Fixed-scope, fixed-price, agreed in writing before work starts. They assume a US or Canadian client, content that exists in some form even if it needs tidying, and a scope that has been through one real conversation rather than guessed from a form submission.
The running cost, honestly
This is where expectations are usually wrong in both directions. People either assume the model bill will be enormous, or forget it exists.
Model usage scales with conversation volume, not with how many people on your team have a login. A support bot handling a few thousand conversations a month typically costs tens of dollars in model spend, not hundreds, because a grounded answer is a short prompt and a short reply. On top of that sits a small retrieval service and its database, which is another modest monthly line on infrastructure you already pay for or can add cheaply.
Both bills land in your own accounts. We do not resell model access at a markup, partly because it is not a business we want and mostly because it makes the running cost impossible for you to audit.
Custom build against a SaaS subscription
The honest answer is that SaaS wins year one on cash and loses later on control. A subscription is cheaper to start, requires no engineering from you, and gets steadily more expensive as your volume grows, because the pricing is per seat or, increasingly, per resolved conversation. That second model in particular means success costs you more.
A custom build inverts it. Higher up front, then usage priced at what the provider charges, with the code in your repository and no vendor able to reprice you. For most businesses the lines cross somewhere in the second year. If you expect to still be running this in three years, or if your content is a genuine competitive asset you would rather not upload wholesale to a third-party index, the build is usually the better deal. If you want something live by Friday and you are testing whether customers will even use it, buy the subscription. We will say that on the call.
What gets left out of most quotes
Four line items disappear from chatbot proposals with suspicious regularity.
- The content work. If nine of your top twenty questions are not answered anywhere in writing, no amount of retrieval will find them. Someone has to write those answers. Budget for it, or accept a lower ceiling.
- The integration. A bot that cannot pass a conversation into the tool your team actually watches will get ignored by your team, and then by your customers.
- Shadow mode. A week of the bot drafting answers your staff approve before anything reaches a customer. It is the cheapest week in the project and the one most likely to get cut.
- Upkeep. HubSpot's State of Service data shows teams whose help centre was updated in the last 30 days reporting around 45% deflection, against 18% for teams that had not audited theirs in six months. Same software, two and a half times the outcome. Whoever owns that content owns the result.
How to sanity-check any quote you get
Three questions worth asking whoever you end up talking to, us included.
- Are you quoting deflection or resolution, and which definition are you using? A vendor reporting 80% deflection can be performing worse than one reporting 55% resolution, because a deflected conversation includes the customer who gave up.
- What happens when the bot does not know? If there is no designed refusal path, you are buying a liability. In Moffatt v. Air Canada (February 2024) the British Columbia Civil Resolution Tribunal held the airline responsible for a wrong answer its chatbot gave a customer, rejecting the argument that the bot was a separate entity.
- Who owns it at the end, and what happens if we stop paying you? If the answer involves a platform, price the exit before you price the entry.
Want a fixed number for your site rather than a range? Send us the URL and we will tell you what is answerable today.
Book Free Consultation