The three lanes, defined without hype
Buy: a support-bot product (Intercom Fin-class, Chatbase-class) pointed at your help docs, live in days, priced monthly per resolution or per seat, limited to the vendor's model, flows and data handling. Build: a custom retrieval pipeline over your knowledge base with your escalation rules and tone, on API models you choose, more control, real engineering, the chatbot development service is this lane. Fine-tune: training a model on your data to change its behavior, the expensive option, and the one most often bought when it is not needed.
The decision, honestly
Buy when the job is standard support over standard documentation: your help center, your policies, your FAQ, typical questions. The bought bots are genuinely good at this now, and the honest consultant says "buy" before quoting anything, the failure mode is real but bounded, and you learn what customers actually ask. Build (retrieval/custom) when the knowledge is proprietary, product catalogs with rules, internal runbooks, data that must stay in your stack, or the escalation flows, tone, and integrations are the differentiator. Retrieval-augmented generation (RAG) over your documents is the standard architecture; the RAG development page covers it technically. Fine-tune when, and this is rare, the behavior you need is a style or format no amount of retrieved context produces, you have thousands of cleaned examples, and you have the appetite to maintain a trained model. Fine-tuning to "teach the bot your facts" is the classic misuse: retrieval does that better, cheaper, and updateable this afternoon.
The costs, shaped honestly (verify current pricing before deciding)
Buy: per-resolution pricing that scales with success, cheap at low volume, real money at scale, plus the subscription. Build: engineering up front, then API costs that scale with usage (the LLM cost ceilings guide covers the token math), plus maintenance of the knowledge base the bot reads. Fine-tune: training cost + hosting + the same API costs + a re-training cycle whenever the facts change. The ceilings conversation happens before the build, the AI-assisted build process and its eval gates apply to every lane.
The honest middle path most businesses should take
Buy first, on a short leash: deploy a bought bot over clean docs, watch what customers actually ask for a quarter, measure containment and complaints. That data is the spec for a custom build if the limits bite, and if they never bite, you just saved a build. The sequence buys knowledge with subscriptions instead of engineering. When the limits do bite, proprietary knowledge, data control, escalation depth, the agent and RAG pages cover what building adds. Bring the docs and the complaint logs; the recommendation follows the data.