AI chatbot for your website: build, buy, or embed?
Who owns the answers, data and risk when you add an LLM assistant to a business site. Cost shapes, vendor questions and the risks that stay with you.
"Add a chatbot to the website" now means something specific: an LLM-backed assistant that answers in natural language, not a scripted widget with three buttons. The build-or-buy decision for that assistant is less about technology fashion and more about who owns the answers, the data and the risk. This guide lays out the decision for a business site, checked against current vendor documentation and the OWASP GenAI Security Project on 19 November 2026.
What you are actually choosing
Buying means embedding a hosted chatbot product: a script tag or plugin, a monthly subscription, and the vendor's model, hosting and dashboard. Building means your own assistant: a model API (billed by usage), your own retrieval over your own content, your own UI, and your own guardrails. A middle path - a platform's embeddable assistant pointed at your documentation - behaves like buying for cost and like building for configuration effort.
The cost structures are different shapes
A bought assistant is a predictable subscription line item. A built assistant is usage-based: model API billing scales with traffic and answer length, and the engineering cost is front-loaded - retrieval, evaluation, safety handling and the UI all exist before the first visitor asks anything. API pricing is public and changes; check the model provider's current pricing page when you model costs rather than trusting any article, including this one. Unbounded usage is a real failure mode either way: our guide to unbounded consumption and cost abuse covers the caps to put in place.
The risks do not transfer with the invoice
Whether you build or buy, three risks stay with the site owner:
- Wrong answers presented with confidence. An assistant that invents your refund policy creates a support problem, not a novelty. See misinformation and overreliance for the pattern and the mitigations.
- Instruction attacks through content. If the assistant reads your pages, tickets or uploaded documents, hostile text in that content can steer it. Our prompt injection explainer covers what reduces the damage.
- Data exposure. An assistant with access to private data can disclose it; one with broad tool access can act on it. The boundaries are described in sensitive information disclosure and excessive agency. The OWASP GenAI Security Project catalogs the same risk families at the industry level.
Buying moves the infrastructure work to the vendor; it does not move these answers-and-data questions. Ask any vendor how their product handles each one, in writing.
When buying is the honest choice
- The job is deflecting repetitive questions whose answers already exist in public documentation.
- Volume is modest and the cost of a wrong answer is a mildly annoyed visitor, not a contractual commitment.
- Nobody on the team will own evaluation, prompt maintenance and incident response for an assistant.
When building earns its cost
- The assistant must answer from your own data - orders, accounts, internal policy - which means retrieval over private sources and per-user access control. Our RAG explainer for founders describes that architecture.
- The assistant takes actions (booking, refunds, changes) where you need explicit, tested authorization boundaries.
- Traffic is high enough that per-conversation API billing beats subscription tiers, or compliance requires controlling where conversations are processed.
Questions that separate real vendors from demos
- What content does the assistant answer from, and how do I see which source it used for a given answer?
- What happens when it does not know - silence, a guess, or a handoff to a human?
- How are conversations stored, and can they be excluded from model training?
- What usage caps and alerts exist, and what does a traffic spike cost?
- How do we export our configuration if we leave?
If you decide a custom assistant is the right shape, our generative AI development page describes how such engagements are scoped, and the project scoping guide helps write the requirement down first.
Sources
- OWASP GenAI Security Project (LLM/GenAI risk catalog): https://genai.owasp.org/
- OpenAI API pricing (check current rates when modeling costs): https://openai.com/api/pricing/
Drafted with AI assistance from these sources and Omnistu's own referenced guides. Pricing and product capabilities change; verify against current vendor documentation before committing.