A website AI chatbot costs $5k–$50k to build custom, or $50–$500 a month off the shelf, and deflects 30–60% of support conversations when it is grounded in your own content. The number that decides whether it works is not deflection: it is what happens on the conversations it cannot handle.
Gartner reported in 2026 that 91% of customer service and support leaders are under executive pressure to implement AI. If you are reading this, there is a reasonable chance you have a mandate rather than a hypothesis, which makes the useful question not "should we" but "what will this actually be like to own".
A quick scoping note, because people arrive at this decision from two directions. If you are shopping for an AI chatbot for website support specifically, you are buying a channel. If you are shopping for AI customer service software, you are usually buying a platform that includes the chatbot alongside ticketing, routing and reporting. The second is a bigger purchase and a harder migration, so it is worth being clear which one you are actually doing before you take demos. Everything below applies to the chatbot layer either way.
What deflection really looks like
Vendors quote 70–80%. In systems we run and inherit, grounded in the customer's own documentation, 30–60% is the honest band, and where you land inside it depends almost entirely on how good your documentation is.
That is worth restating because it is the whole game: a chatbot is a retrieval system with a conversational front end. If the answer exists in your help centre, it will find it. If your help centre is thin, out of date or contradicts itself, the bot will confidently reproduce all three problems at scale. Teams that spend two weeks fixing content before launch consistently outperform teams that spend two weeks tuning prompts.
Also be careful how deflection is counted. A conversation where the customer gave up and emailed you instead is not deflected, it is displaced. Measure resolution as confirmed by the customer, plus the volume arriving in your other channels.
Build or buy
| Off-the-shelf | Custom build | |
|---|---|---|
| Cost | $50–$500 / mo | $5k–$50k build |
| Live in | Days | 3–8 weeks |
| Grounding | Crawls your site and docs | Any source, including internal systems |
| Permissions | Usually none or coarse | Per-user, honouring existing rules |
| Actions | Limited | Order lookups, account changes, ticket creation |
| Data | Vendor-hosted | Your infrastructure if required |
Buy first if your use case is answering questions from public content. The products are good and the economics are hard to beat. Build when the chatbot needs to know things only your systems know (order status, entitlements, account state), when answers must respect who is asking, or when data residency rules out a vendor.
What breaks
- Confident wrong answers. The failure mode that damages trust. Mitigated by strict grounding, citations back to source, and a genuine "I do not know" path. A bot that admits uncertainty is more useful than one that never does.
- Stale content. You changed the refund policy; the bot did not notice. Whoever owns the policy has to own the source the bot reads, or you need automated freshness checks.
- The escalation cliff. Customer asks three times, gets nowhere, cannot find a human. This single design failure produces more complaints than wrong answers do.
- Prompt injection. If your bot can take actions, someone will try to talk it into taking the wrong one. Covered in LLM application security.
- Silent quality drift. Answer quality degrades gradually after a content change or model update, and without monitoring nobody notices for weeks.
The maintenance nobody quotes
Budget roughly half a day a month of someone's attention for a steady-state chatbot, plus tooling. That covers reading a sample of conversations, checking the ones marked unresolved, refreshing content the bot got wrong, and re-running evals after a model change. Two model migrations a year is a reasonable planning assumption.
It is not a large number. It is not zero either, and treating it as zero is why chatbots that launched well are embarrassing eighteen months later.
Designing the escalation
The most important twenty minutes of the project. Decide:
- What triggers escalation: explicit request, repeated failure, detected frustration, or a topic allowlist.
- Where it goes, and what happens outside business hours.
- What the human sees. Full transcript and what the bot already tried, so the customer never repeats themselves.
- Whether the customer can skip straight to a human. They should be able to.
Get this right and a mediocre bot is still a good experience. Get it wrong and an excellent bot generates complaints.
What it is worth
For a team handling 2,000 support conversations a month at roughly $4 of loaded agent time each, 40% deflection is about $38k a year. Against a $25k build and $8k a year to run, it pays back inside the first year, and the second year is close to pure return.
The less-quantified benefit is coverage: instant answers at 2am, in the customer's language, at any concurrency. For companies selling across time zones that is often worth more than the deflection saving.
Our AI chatbot development page covers how we build these, and chatbot vs AI agent is the right read if you are not sure which of the two you need.
Common questions
How much does an AI chatbot for a website cost?
Off-the-shelf website chatbots cost $50–$500 a month. A custom build grounded in your own content and systems costs $5,000–$50,000, with most landing at $15,000–$30,000, plus $200–$1,500 a month to run. Build rather than buy when the bot needs data only your internal systems hold, or when answers must respect per-user permissions.
What percentage of support tickets can an AI chatbot deflect?
Expect 30–60% when the chatbot is grounded in your own documentation, not the 70–80% commonly quoted. The main variable is documentation quality, not model quality. Also check how deflection is counted: a customer who gave up and emailed instead has been displaced, not deflected.
How do I stop an AI chatbot giving wrong answers?
Ground it strictly in your own content, show citations back to the source document, and give it a real path to say it does not know rather than guessing. Then monitor: sample conversations weekly, review everything marked unresolved, and re-run a fixed evaluation set after any content or model change.
How much maintenance does an AI chatbot need?
Budget about half a day a month of a person’s time plus tooling costs. That covers sampling conversations, reviewing unresolved ones, refreshing content the bot answered badly, and re-running evaluations. Plan for one or two model migrations a year, since providers retire models on their own schedule.